José Cordeiro · Slimane Hammoudi

Leszek Maciaszek · Olivier Camp

Joaquim Filipe (Eds.)

Enterprise
LNBIP 227

Information Systems
16th International Conference, ICEIS 2014
Lisbon, Portugal, April 27–30, 2014
Revised Selected Papers

123
Lecture Notes
in Business Information Processing 227

Series Editors
Wil van der Aalst
Eindhoven Technical University, Eindhoven, The Netherlands
John Mylopoulos
University of Trento, Povo, Italy
Michael Rosemann
Queensland University of Technology, Brisbane, QLD, Australia
Michael J. Shaw
University of Illinois, Urbana-Champaign, IL, USA
Clemens Szyperski
Microsoft Research, Redmond, WA, USA
More information about this series at http://www.springer.com/series/7911
José Cordeiro Slimane Hammoudi
•

Leszek Maciaszek Olivier Camp

•

Joaquim Filipe (Eds.)

Enterprise
Information Systems
16th International Conference, ICEIS 2014
Lisbon, Portugal, April 27–30, 2014
Revised Selected Papers

123
Editors
José Cordeiro Olivier Camp
INSTICC MODESTE/ESEO
Polytechnic Institute of Setúbal Angers
Setúbal France
Portugal
Joaquim Filipe
Slimane Hammoudi INSTICC
MODESTE/ESEO Polytechnic Institute of Setúbal
Angers Setúbal
France Portugal
Leszek Maciaszek
Wroclaw University of Economics
Wroclaw
Poland

ISSN 1865-1348 ISSN 1865-1356 (electronic)

Lecture Notes in Business Information Processing
ISBN 978-3-319-22347-6 ISBN 978-3-319-22348-3 (eBook)
DOI 10.1007/978-3-319-22348-3

Library of Congress Control Number: 2015944743

Springer Cham Heidelberg New York Dordrecht London

© Springer International Publishing Switzerland 2015
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the
material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now
known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book are
believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors
give a warranty, express or implied, with respect to the material contained herein or for any errors or
omissions that may have been made.

Printed on acid-free paper

Springer International Publishing AG Switzerland is part of Springer Science+Business Media

(www.springer.com)
 Preface

The present book includes extended and revised versions of a set of selected papers from
the 16th International Conference on Enterprise Information Systems (ICEIS 2014),
held in Lisbon, Portugal, during April 27–30, 2014. The conference was sponsored by
the Institute for Systems and Technologies of Information, Control and Communication
(INSTICC), held in cooperation with the Association for the Advancement of Artificial
Intelligence (AAAI), IEICE Special Interest Group on Software Interprise Modelling
(SWIM), ACM SIGART - ACM Special Interest Group on Artificial Intelligence,
ACM SIGMIS - ACM Special Interest Group on Management Information Systems,
ACM SIGCHI - ACM Special Interest Group on Computer–Human Interaction, and in
collaboration with the Informatics Research Center (IRC).
The conference was organized in six simultaneous tracks: “Databases and Infor-
mation Systems Integration, Artificial Intelligence and Decision Support Systems,
Information Systems Analysis and Specification, Software Agents and Internet Com-
puting, Human–Computer Interaction, and Enterprise Architecture”. The book is based
on the same structure.
ICEIS 2014 received 313 paper submissions from 50 countries in all continents.
From these, after a blind review process, only 47 were accepted as full papers, of which
24 were selected for inclusion in this book, based on the classifications provided by the
Program Committee. The selected papers reflect state-of-art research work that is often
oriented toward real-world applications and highlight the benefits of information sys-
tems and technology for industry and services, thus forming a bridge between the
worlds of academia and enterprise. These high-quality standards were maintained and
reinforced at ICEIS 2015, which was held in Barcelona, Spain, and will continue to be
reinforced in future editions of this conference.
Furthermore, ICEIS 2014 included five plenary keynote lectures given by Kecheng
Liu (University of Reading, UK), Jan Dietz (Delft University of Technology, The
Netherlands), Antoni Olivé (Universitat Politècnica de Catalunya, Spain), José Tribolet
(INESC-ID/Instituto Superior Técnico, Portugal), and Hans-J. Lenz (Freie Universität
Berlin, Germany). We would like to express our appreciation to all of them and in
particular to those who took the time to contribute with a paper to this book.
On behalf of the conference Organizing Committee, we would like to thank all
participants. First of all the authors, whose quality work is the essence of the confer-
ence, and the members of the Program Committee, who helped us with their expertise
and diligence in reviewing the papers. As we all know, producing a conference requires
VI Preface

the effort of many individuals. We wish to thank also all the members of our Orga-
nizing Committee, whose work and commitment were invaluable.

February 2014 José Cordeiro

Slimane Hammoudi
Leszek Maciaszek
Olivier Camp
Joaquim Filipe
 Organization

Conference Co-chairs
Joaquim Filipe Polytechnic Institute of Setúbal/INSTICC, Portugal
Olivier Camp ESEO, MODESTE, France

Program Co-chairs
Slimane Hammoudi ESEO, MODESTE, France
Leszek Maciaszek Wroclaw University of Economics, Poland and
Macquarie University, Sydney, Australia
José Cordeiro Polytechnic Institute of Setúbal/INSTICC, Portugal

Organizing Committee
Marina Carvalho INSTICC, Portugal
Helder Coelhas INSTICC, Portugal
Bruno Encarnação INSTICC, Portugal
Ana Guerreiro INSTICC, Portugal
André Lista INSTICC, Portugal
Filipe Mariano INSTICC, Portugal
Andreia Moita INSTICC, Portugal
Raquel Pedrosa INSTICC, Portugal
Vitor Pedrosa INSTICC, Portugal
Cláudia Pinto INSTICC, Portugal
Cátia Pires INSTICC, Portugal
Susana Ribeiro INSTICC, Portugal
Rui Rodrigues INSTICC, Portugal
Sara Santiago INSTICC, Portugal
André Santos INSTICC, Portugal
Fábio Santos INSTICC, Portugal
Mara Silva INSTICC, Portugal
José Varela INSTICC, Portugal
Pedro Varela INSTICC, Portugal

Senior Program Committee

Balbir Barn Middlesex University, UK
Senén Barro University of Santiago de Compostela, Spain
Albert Cheng University of Houston, USA
Jan Dietz Delft University of Technology, The Netherlands
VIII Organization

Schahram Dustdar Vienna University of Technology, Austria

António Figueiredo University of Coimbra, Portugal
Nuno Guimarães Lasige/ISCTE-University Institute of Lisbon, Portugal
Jan Jürjens TU Dortmund and Fraunhofer ISST, Germany
Kecheng Liu University of Reading, UK
Pericles Loucopoulos Harokopio University of Athens, Greece
Andrea de Lucia Università degli Studi di Salerno, Italy
Yannis Manolopoulos Aristotle University, Greece
José Legatheaux Martins FCT/UNL, Portugal
Masao Johannes Matsumoto Solution Research Lab, Japan
Alain Pirotte Université catholique de Louvain, Belgium
Matthias Rauterberg Eindhoven University of Technology, The Netherlands
Colette Rolland Université Paris 1 Panthéon-Sorbonne, France
Narcyz Roztocki State University of New York at New Paltz, USA
Abdel-Badeeh Mohamed Ain Shams University, Egypt
Salem
Bernadette Sharp Staffordshire University, UK
Alexander Smirnov SPIIRAS, Russian Academy of Sciences, Russian
Federation
Ronald Stamper Measur Ltd., UK
François Vernadat European Court of Auditors, France

Program Committee
Lena Aggestam Region Västra Götaland, The Hospital in Alingsås,
Sweden
Miguel Angel Martinez University of Murcia, Spain
Aguilar
Adeel Ahmad Laboratoire d’Informatique Signal et Image de la Côte
d’Opale, France
Antonia Albani University of St. Gallen, Switzerland
Patrick Albers ESEO - Ecole Superieure D’Electronique de L’Ouest,
France
Abdullah Alnajim Qassim University, Saudi Arabia
Mohammad Al-Shamri Ibb University, Yemen
Rainer Alt University of Leipzig, Germany
Andreas S. Andreou Cyprus University of Technology, Cyprus
Wudhichai King Mongkut’s University of Technology Thonburi,
Assawinchaichote Thailand
Tamara Babaian Bentley University, USA
Cecilia Baranauskas State University of Campinas - Unicamp, Brazil
Rémi Bastide ISIS - CUFR Jean-François Champollion, France
Bernhard Bauer University of Augsburg, Germany
Orlando Belo University of Minho, Portugal
Jorge Bernardino Polytechnic Institute of Coimbra - ISEC, Portugal
Frederique Biennier INSA Lyon, France
 Organization IX

Sandro Bimonte Irstea, France

Jean-Louis Boulanger CERTIFER, France
Coral Calero University of Castilla – La Mancha, Spain
Daniel Antonio Callegari PUC-RS Pontificia Universidade Catolica do Rio
Grande do Sul, Brazil
Luis M. Camarinha-Matos New University of Lisbon, Portugal
Manuel Isidoro University of Granada, Spain
Capel-Tuñón
Glauco Carneiro Salvador University (UNIFACS), Brazil
Angélica Caro University of Bio-Bio, Chile
Nunzio Casalino Università degli Studi Guglielmo Marconi, Italy
Marco Antonio Casanova PUC-Rio, Brazil
Luca Cernuzzi Universidad Católica “Nuestra Señora de la Asunción”,
Paraguay
David Chen Laboratory IMS, France
Ming-Puu Chen National Taiwan Normal University, Taiwan
Shiping Chen CSIRO ICT Centre Australia, Australia
Shu-Ching Chen Florida International University, USA
Max Chevalier Institut de Recherche en Informatique de
Toulouse UMR 5505, France
Nan-Hsing Chiu Chien Hsin University of Science and Technology,
Taiwan
Witold Chmielarz Warsaw University, Poland
William Cheng-Chung Chu Tunghai University, Taiwan
Daniela Barreiro Claro Universidade Federal da Bahia (UFBA), Brazil
Pedro Gouvêa Coelho State University of Rio de Janeiro, Brazil
Francesco Colace Università Degli Studi di Salerno, Italy
Cesar Collazos Universidad del Cauca, Colombia
Antonio Corral University of Almeria, Spain
Mariela Cortés State University of Ceará, Brazil
Karl Cox University of Brighton, UK
Sharon Cox Birmingham City University, UK
Broderick Crawford Pontificia Universidad Catolica de Valparaiso, Chile
Maria Damiani University of Milan, Italy
Vincenzo Deufemia Università di Salerno, Italy
Dulce Domingos University of Lisbon, Portugal
César Domínguez Universidad de La Rioja, Spain
António Dourado University of Coimbra, Portugal
Juan C. Dueñas Universidad Politécnica de Madrid, Spain
Alan Eardley Staffordshire University, UK
Sophie Ebersold Université Toulouse II-LeMirail, France
Hans-Dieter Ehrich Technische Universität Braunschweig, Germany
Fabrício Enembreck Pontifical Catholic University of Paraná, Brazil
Sean Eom Southeast Missouri State University, USA
Hossam Faheem Ain Shams University, Egypt
João Faria FEUP, University of Porto, Portugal
X Organization

Jamel Feki University of Sfax, Tunisia

Edilson Ferneda Catholic University of Brasília, Brazil
Maria João Silva Costa Universidade Portucalense, Portugal
Ferreira
Paulo Ferreira INESC-ID/IST, Portugal
George Feuerlicht University of Technology, Sydney (UTS), Australia
Barry Floyd California Polytechnic State University, USA
Rita Francese Università degli Studi di Salerno, Italy
Ariel Frank Bar-Ilan University, Israel
Ana Fred Instituto de Telecomunicações/IST, Portugal
Lixin Fu University of North Carolina, Greensboro, USA
Mariagrazia Fugini Politecnico di Milano, Italy
Maria Ganzha SRI PAS and University of Gdansk, Poland
Mouzhi Ge Technical University of Munich, Germany
Johannes Gettinger University of Hohenheim, Germany
Daniela Giordano University of Catania, Italy
Raúl Giráldez Pablo de Olavide University of Seville, Spain
Pascual Gonzalez Universidad de Castilla-La Mancha, Spain
Robert Goodwin Flinders University of South Australia, Australia
Raj Gopalan Curtin University, Australia
Feliz Gouveia University Fernando Pessoa/Cerem, Portugal
Virginie Govaere INRS, France
Janis Grabis Riga Technical University, Latvia
Maria Carmen Penadés Universitat Politècnica de València, Spain
Gramaje
Gerd Groener University of Koblenz-Landau, Germany
Sven Groppe University of Lübeck, Germany
Tom Gross Bauhaus University Weimar, Germany
Wieslawa Gryncewicz Wroclaw University of Economics, Poland
Slimane Hammoudi ESEO, MODESTE, France
Karin Harbusch Universität Koblenz-Landau, Germany
Markus Helfert Dublin City University, Ireland
Wladyslaw Homenda Warsaw University of Technology, Poland
Wei-Chiang Hong Oriental Institute of Technology, Taiwan
Miguel J. Hornos University of Granada, Spain
Hesuan Hu Nanyang Technological University, Singapore
Kai-I Huang Tunghai University, Taiwan
Miroslav Hudec University of Economics in Bratislava, Slovak
Republic
Arturo Jaime Universidad de La Rioja, Spain
Wassim Jaziri College of Computer Science and Engineering,
Saudi Arabia
Sabina Jeschke RWTH Aachen University, Germany
Edson Oliveira Jr. State University of Maringá, Brazil
Nikitas Karanikolas Technological Educational Institute of Athens (TEI-A),
Greece
 Organization XI

Andrea Kienle University of Applied Sciences, Dortmund, Germany

Marite Kirikova Riga Technical University, Latvia
Alexander Knapp Universität Augsburg, Germany
Natallia Kokash Leiden University, The Netherlands
John Krogstie NTNU, Norway
Rob Kusters Eindhoven University of Technology and Open
University of the Netherlands, The Netherlands
Wim Laurier Université Saint-Louis, Belgium
Ramon Lawrence University of British Columbia Okanagan, Canada
Jintae Lee Leeds School of Business at University of Colorado,
Boulder, USA
Alain Leger France Telecom Orange Labs, France
Daniel Lemire TELUQ, Canada
Joerg Leukel University of Hohenheim, Germany
Lei Li Hefei University of Technology, China
Da-Yin Liao Applied Wireless Identifications, USA
Therese Libourel University of Montpellier II (IRD,UR, UAG), France
Luis Jiménez Linares University of de Castilla-La Mancha, Spain
Panos Linos Butler University, USA
Stephane Loiseau LERIA, University of Angers, France
João Correia Lopes Universidade do Porto/INESC Porto, Portugal
Maria Filomena Cerqueira Universidade Portucalense Infante D. Henrique,
de Castro Lopes Portugal
Miguel R. Luaces Universidade da Coruña, Spain
Wendy Lucas Bentley University, USA
André Ludwig University of Leipzig, Germany
Mark Lycett Brunel University, UK
Jose Antonio Macedo Federal University of Ceara, Brazil
Leszek Maciaszek Wroclaw University of Economics, Poland and
Macquarie University, Sydney, Australia
Cristiano Maciel Universidade Federal de Mato Grosso, Brazil
Rita Suzana Pitangueira Federal University of Bahia, Brazil
Maciel
S. Kami Makki Lamar University, USA
Pierre Maret Université de Saint Etienne, France
Herve Martin Grenoble University, France Katsuhisa Maruyama,
Ritsumeikan University, Japan
Viviana Mascardi University of Genoa, Italy
David Martins de Matos L2F/INESC-ID Lisboa/Instituto Superior Técnico,
Portugal
Wolfgang Mayer University of South Australia, Australia
Zaamoune Mehdi Irstea, France
Andreas Meier University of Fribourg, Switzerland
Jerzy Michnik University of Economics in Katowice, Poland
Marek Milosz Lublin University of Technology, Poland
Michele Missikoff IASI-CNR, Italy
XII Organization

Vladimír Modrák Technical University of Košice, Slovak Republic

Ghodrat Moghadampour Vaasa University of Applied Sciences, Finland
Pascal Molli LINA, University of Nantes, France
Lars Mönch Fern Universität in Hagen, Germany
Valérie Monfort Université de Paris1 Panthéon Sorbonne, France
Francisco Montero University of Castilla-la Mancha, Spain
Carlos León de Mora University of Seville, Spain
João Luís Cardoso Federal University of São Carlos, Brazil
de Moraes
Fernando Moreira Universidade Portucalense, Portugal
Nathalie Moreno University of Malaga, Spain
Haralambos Mouratidis University of East London, UK
Pietro Murano University of Salford, UK
Tomoharu Nakashima Osaka Prefecture University, Japan
Ovidiu Noran Griffith University, Australia
Jose Angel Olivas Universidad de Castilla-La Mancha, Spain
Andrés Muñoz Ortega Catholic University of Murcia (UCAM), Spain
Samia Oussena University of West London, UK
Sietse Overbeek University of Duisburg-Essen, Germany
Mieczyslaw Owoc Wroclaw University of Economics, Poland
Claus Pahl Dublin City University, Ireland
Tadeusz Pankowski Poznan University of Technology, Poland
Eric Pardede La Trobe University, Australia
Rodrigo Paredes Universidad de Talca, Chile
Massimiliano Di Penta University of Sannio, Italy
Dana Petcu West University of Timisoara, Romania
Yannis A. Phillis Technical University of Crete, Greece
Josef Pieprzyk Macquarie University, Australia
Luis Ferreira Pires University of Twente, The Netherlands
Ángeles Saavedra Places University of A Coruña, Spain
Malgorzata Lublin University of Technology, Poland
Plechawska-Wojcik
Geert Poels Ghent University, Belgium
Michal Polasik Nicolaus Copernicus University, Poland
Luigi Pontieri National Research Council (CNR), Italy
Jolita Ralyte University of Geneva, Switzerland
T. Ramayah Universiti Sains Malaysia, Malaysia
Pedro Ramos Instituto Superior das Ciências do Trabalho e da
Empresa, Portugal
Francisco Regateiro Instituto Superior Técnico, Portugal
Ulrich Reimer University of Applied Sciences St. Gallen, Switzerland
Nuno de Magalhães Ribeiro Universidade Fernando Pessoa, Portugal
Michele Risi University of Salerno, Italy
Alfonso Rodriguez University of Bio-Bio, Chile
Daniel Rodriguez University of Alcalá, Spain
Oscar Mario Institute of Technology of Hermosillo, Mexico
Rodriguez-Elias
 Organization XIII

Erik Rolland University of California at Merced, USA

Luciana Alvim Santos Embrapa Agriculture Informatics, Brazil
Romani
Jose Raul Romero University of Cordoba, Spain
David G. Rosado University of Castilla-la Mancha, Spain
Gustavo Rossi Lifia, Argentina
Artur Rot Wroclaw University of Economics, Poland
Francisco Ruiz Universidad de Castilla-La Mancha, Spain
Belen Vela Sanchez Rey Juan Carlos University, Spain
Luis Enrique Sánchez Sicaman Nuevas Tecnologias S.L., Spain
Manuel Filipe Santos University of Minho, Portugal
Jurek Sasiadek Carleton University, Canada
Isabel Seruca Universidade Portucalense, Portugal
Ahm Shamsuzzoha University of Vaasa, Finland
Jianhua Shao Cardiff University, UK
Mei-Ling Shyu University of Miami, USA
Markus Siepermann TU Dortmund, Germany
Alberto Rodrigues Silva Instituto Superior Técnico, Portugal
Sean Siqueira Federal University of the State of Rio de Janeiro, Brazil
Hala Skaf-molli Nantes University, France
Michel Soares Federal University of Sergipe, Brazil
Ricardo Soto Pontificia Universidad Catolica de Valparaiso, Chile
Chantal Soule-Dupuy University of Toulouse 1, France
Chris Stary Johannes Kepler University of Linz, Austria
Stefan Strecker University of Hagen, Germany
Vijayan Sugumaran Oakland University, USA
Hiroki Suguri Miyagi University, Japan
Lily Sun University of Reading, UK
Jerzy Surma Warsaw School of Economics, Poland
Miroslav Sveda Brno University of Technology, Czech Republic
Ryszard Tadeusiewicz AGH University of Science and Technology, Poland
Tania Tait Maringá State University, Brazil
Mohan Tanniru Oakland University, USA
Sotirios Terzis University of Strathclyde, UK
Claudine Toffolon Université du Maine, France
Ying-Mei Tu Chung Hua University, Taiwan
Theodoros Tzouramanis University of the Aegean, Greece
José Ângelo Braga de Universidade Atlântica, Portugal
Vasconcelos
Michael Vassilakopoulos University of Thessaly, Greece
Christine Verdier LIG - Joseph Fourier University of Grenoble, France
Bing Wang University of Hull, UK
Dariusz Wawrzyniak Wroclaw University of Economics, Poland
Hans Weghorn BW Cooperative State University of Stuttgart,
Germany
Hans Weigand Tilburg University, The Netherlands
Viacheslav Wolfengagen Institute JurInfoR, Russian Federation
XIV Organization

Ouri Wolfson University of Illinois at Chicago, USA

Stanislaw Wrycza University of Gdansk, Poland
Ing-Long Wu National Chung Cheng University, Taiwan
Mudasser Wyne National University, USA
Hongji Yang De Montfort University, UK
Eugenio Zimeo University of Sannio, Italy

Additional Reviewers
María Luisa Rodríguez University of Granada, Spain
Almendros
Almudena Sierra Alonso Universidad Rey Juan Carlos, Spain
Pierluigi Assogna Libera Università Internazionale Degli Studi Sociali
(luiss), Italy
Alexander Bock University of Duisburg-Essen, Germany
Emmanuel Desmontils Faculté des Sciences et des Techniques, Université de
Nantes, France
Adolfo Almeida Duran Federal University of Bahia, Brazil
Fausto Fasano University of Molise, Italy
Tiago Ferreto PUCRS, Brazil
Francesco Folino ICAR-CNR, Italy
Francisco Javier Pablo de Olavide University, Spain
Gil-Cumbreras
Massimo Guarascio ICAR-CNR, Italy
Agnieszka Jastrzebska Warsaw University of Technology, Poland
Paloma Cáceres García Rey Juan Carlos University, Spain
de Marina
Asmat Monaghan Brunel University, UK
Luã Marcelo Muriana UFF, Brazil
Aurora Ramirez University of Córdoba, Spain
Monica Sebillo Università degli Studi di Salerno, Italy
Panagiotis Symeonidis Aristotle University of Thessaloniki, Greece
Pawel Szmeja Polish Academy of Sciences, Poland
Francesco Taglino CNR-IASI, Italy
Mahsa Teimourikia Politecnico di Milano, Italy

Invited Speakers
Kecheng Liu University of Reading, UK
Jan Dietz Delft University of Technology, The Netherlands
Antoni Olivé Universitat Politècnica de Catalunya, Spain
José Tribolet INESC-ID/Instituto Superior Técnico, Portugal
Hans-J. Lenz Freie Universität Berlin, Germany
 Contents

Invited Papers

Semiotics in Digital Visualisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

Kecheng Liu and Chekfoung Tan

Data Fraud Detection: A First General Perspective . . . . . . . . . . . . . . . . . . . 14

Hans-J. Lenz

Databases and Information Systems Integration

ERP System Integration: An Inter-organizational Challenge in the Dynamic

Business Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
Tommi Kähkönen, Andrey Maglyas, and Kari Smolander

Using Complex Correspondences for Integrating Relational Data Sources . . . 57

Valéria Pequeno, Helena Galhardas, and Vânia M. Ponte Vidal

Towards an Efficient and Distributed DBSCAN Algorithm

Using MapReduce. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
Ticiana L. Coelho da Silva, Antonio C. Araújo Neto,
Regis Pires Magalhães, Victor A.E. de Farias, José A.F. de Macêdo,
and Javam C. Machado

Discovering Frequent Patterns on Agrometeorological Data with TrieMotif . . . 91

Daniel Y.T. Chino, Renata R.V. Goncalves, Luciana A.S. Romani,
Caetano Traina Jr., and Agma J.M. Traina

An Approach to the Discovery of Accurate and Expressive Fix-Time

Prediction Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
Francesco Folino, Massimo Guarascio, and Luigi Pontieri

Artificial Intelligence and Decision Support Systems

Solving the Unrelated Parallel Machine Scheduling Problem with Setup

Times by Efficient Algorithms Based on Iterated Local Search . . . . . . . . . . . 131
Matheus N. Haddad, Luciano P. Cota, Marcone J.F. Souza,
and Nelson Maculan

Using Data Mining for Prediction of Hospital Length of Stay:

An Application of the CRISP-DM Methodology . . . . . . . . . . . . . . . . . . . . . 149
Nuno Caetano, Paulo Cortez, and Raul M.S. Laureano
XVI Contents

Applying Artificial Immune Systems for Deploying Router Nodes

in Wireless Networks in the Presence of Obstacles . . . . . . . . . . . . . . . . . . . 167
Pedro Henrique Gouvêa Coelho, Jorge Luís Machado do Amaral,
José Franco Machado do Amaral, Luciane Fernanda de Arruda Barreira,
and Adriano Valladão de Barros

Information Systems Analysis and Specification

Performance Evaluation of Aspect-Oriented Programming Weavers . . . . . . . . 187

Michel S. Soares, Marcelo A. Maia, and Rodrigo F.G. Silva

ABOR: An Automatic Framework for Buffer Overflow Removal

in C/C++Programs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
Sun Ding, Hee Beng Kuan Tan, and Hongyu Zhang

ValueApping: An Analysis Method to Identify Value-Adding

Mobile Enterprise Apps in Business Processes . . . . . . . . . . . . . . . . . . . . . . 222
Eva Hoos, Christoph Gröger, Stefan Kramer, and Bernhard Mitschang

An Overview on Aspect-Oriented Requirements Engineering Area . . . . . . . . 244

Paulo Afonso Parreira Júnior and Rosângela Dellosso Penteado

SpecQua: Towards a Framework for Requirements Specifications

with Increased Quality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
Alberto Rodrigues da Silva

Software Agents and Internet Computing

Analyzing Virtual Organizations’ Formation Risk in P2P

SON Environments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285
Rafael Giordano Vieira, Omir Correia Alves Junior,
and Adriano Fiorese

CRAWLER-LD: A Multilevel Metadata Focused Crawler Framework

for Linked Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302
Raphael do Vale A. Gomes, Marco A. Casanova, Giseli Rabello Lopes,
and Luiz André P. Paes Leme

A Framework Reactive and Proactive for Pervasive Homecare

Environments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 320
Alencar Machado, Daniel Lichtnow, Ana Marilza Pernas,
Amel Bouzeghoub, Iara Augustin, Leandro Krug Wives,
and José Palazzo Moreira de Oliveira
 Contents XVII

Human-Computer Interaction

Designing an IDTV Application in a Situated Scenario: A Participatory

Approach Based on Patterns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 341
Samuel B. Buchdid, Roberto Pereira, and M. Cecília C. Baranauskas

Video Stream Transmodality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 361

Pierre-Olivier Rocher, Christophe Gravier, Julien Subercaze,
and Marius Preda

STAR: Speech Therapy with Augmented Reality for Children with Autism
Spectrum Disorders . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 379
Camilla Almeida da Silva, António Ramires Fernandes,
and Ana Paula Grohmann

Interpretation of Construction Patterns for Biodiversity Spreadsheets . . . . . . . 397

Ivelize Rocha Bernardo, Michela Borges,
Maria Cecília Calani Baranauskas, and André Santanchè

Enterprise Architecture

Integrating Business Information Streams in a Core Banking Architecture:

A Practical Experience. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 417
Beatriz San Miguel, Jose M. del Alamo, and Juan C. Yelmo

PRIMROSe: A Graph-Based Approach for Enterprise Architecture

Analysis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 434
David Naranjo, Mario Sánchez, and Jorge Villalobos

A Review and Evaluation of Business Model Ontologies:

A Viability Perspective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 453
A. D’Souza, N.R.T.P. van Beest, G.B. Huitema, J.C. Wortmann,
and H. Velthuijsen

Enterprise-Specific Ontology-Driven Process Modelling . . . . . . . . . . . . . . . . 472

Nadejda Alkhaldi, Sven Casteleyn, and Frederik Gailly

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 489

Invited Papers
 Semiotics in Digital Visualisation

Kecheng Liu ✉ and Chekfoung Tan

( )

Informatics Research Centre, University of Reading,

Whiteknights campus, Reading, RG6 6UD, UK
[email protected], [email protected]

Abstract. Digital visualisation is a way of representing data and information

with the aid of digital means, engages human interpretation on information in
order to gain insights in a particular context. Digital visualisation is always
purposeful that will illustrate relationships; discover patterns and interdependen‐
cies; or generate some hypothesis or theory. However, it is still bound to the
semantic (interpretation of visual displays and the meaning in the context) and
pragmatic (effect and intention to be achieved) issues. These two issues can be
addressed by semiotics, a formal doctrine of signs introduced by Peirce back in
the 1930s; where digital visualisation is seen as a process of abduction. Abduction
is a key process of scientific inquiry, which involves norms. Norms are patterns,
regulations, rules and laws which are the reflection of knowledge in a cultural
group or an organisation; which has an effect on the human interpretation on
information. This paper pioneers a new perspective of digital visualisation by
positioning digital visualisation as a process of abduction and proposes the key
principles in digital visualisation.

Keywords: Semiotics · Digital visualisation · Norms · Abduction process ·

Hypotheses formation

1 Introduction

Digital visualisation is a way of representing data and information with the aid of digital
means. It ranges from a simple form such as a graph or chart to a complex form like
animated visualisations that allows user to interact with the underlying data through
direct manipulation [1]. The notion of digital visualisation engages human interpretation
on information in order to gain insights in a particular context [2–4]. Hence, it is a
complex process involving multiple disciplines, including the socio-technical element.
The social element relates to human perception in interpreting information. The tech‐
nical element on the other hand, refers to the technology used to enable visualisation,
for example the SAS suite [5] that offers visual analytics to support interactive dashboard
and reporting.
A typical process of digital visualisation starts from data collection, followed by data
transformation and filtering, and finally the visual display [adapted from 4]. The primary
goal of digital visualisation is enhance the efficiency of understanding of certain
phenomenon through visualised data and information; but it also aims to address latent

© Springer International Publishing Switzerland 2015

J. Cordeiro et al. (Eds.): ICEIS 2014, LNBIP 227, pp. 3–13, 2015.
DOI: 10.1007/978-3-319-22348-3_1
4 K. Liu and C. Tan

aspects, such as semantic and pragmatic issues [6, 7]. The semantic issues are related to
the interpretation of visual displays and the meaning in the context. The pragmatic issues
are concerned with the effect and intention to be achieved on the users through digital
visualisation.
Digital visualisation is always purposeful that will illustrate relationships; discover
patterns and interdependencies; or generate some hypothesis or theory. The user’s
hypothesis would very much influence on what data would be interested in the analysis
and be included in the visualisation. Therefore, there will be always a set of questions
highly relevant in any visualisation, such as data availability, access, format (data itself
and display format), meaning (i.e. interpretation), purpose of data presented, and effect
of visualised data on the recipients. Such questions can be best answered by drawing
input from semiotics.
This paper pioneers a new perspective on digital visualisation by positioning digital
visualisation as a process of abduction. Section 2 illustrates the components, process
and the challenges of digital visualisation. Section 3 describes the semiotics perspective
to digital visualisation. Section 4 suggests a set of principles for digital visualisation
aftermath of the notion of semiotics. This paper is concluded with the future research
insights by intertwining the notion of semiotics, abduction process especially with digital
visualisation.

2 Digital Visualisation

2.1 Components and Process

Digital visualisation deals with three components: data, information and knowledge
[adapted from 8], in which can be explained by social and technical perspective (see
Fig. 1). In the social facet, human users perceive data in the forms of e.g. figures,
numbers and charts. Such a presentation of the data enables the human users to
understand the data better and to obtain information efficiently through the
processing of the data, which bring in the meaning and insight of the data to the
user. In the technical facet, digital visualisation assists the users further to gain
knowledge through interaction, interpretation and abstraction of data and informa‐
tion to produce generic types or patterns based on classifications of contexts and
application domains. Table 1 summarises the description of the components of
digital visualisation from the social and technical perspective.
The process of digital visualisation captures the intentions from users, exposes the
effect of data and addresses the purposes of data analysis through graphical images.
This is reflected in the four main stages in the process of digital visualisation
[adapted from 4, 9] (see Fig. 2): (1) data collection, (2) data transformation and
filtering, (3) visual mapping and, (4) display and interaction. In the data collection
stage, data is gathered through various sources. The collected data can be structured
or unstructured data. According to Baars and Kemper [10], the structured data is data
that has predefined format and can be processed by computing equipment. Unstruc‐
tured data is data that usually does not have a specific format such as social media
data, images, and audio files [11]. The data transformation and filtering stage
 Semiotics in Digital Visualisation 5

Fig. 1. Social and technical facet of digital visualization.

Table 1. Description of the digital visualisation components.

Component Social (Human Perception) Technical (Enabling Visualisation)

Data Figures, numbers, text Representing abstracted data sche‐
matic form, including attributes
or variables for the units of data
(direct display, no deep
processing)
Information Processed data that answers Showing/emphasising on relation‐
what, who, where, when, ship between data items; with
why and how questions interpretation; show semantics
Knowledge Application of data that Representing the effect on organi‐
explains the type or patterns sation through visual analytics
of a situation or context on social, on economy, e.g. drop
on sales, decision-making by
looking at alternatives

explores the collected data through various techniques such as extract, transform and
load (ETL) and data mining. Data are extracted from the source databases are trans‐
formed into a desired format for analysis purpose, and then loaded into the destina‐
tion databases [12]. Some of the common data mining techniques in dealing with
structured data are clustering and categorisation for visualisation purpose. The visual
mapping stage contains a graphic engine that manipulate the process of transforming
the focus data with their geometric primitives (e.g. points, lines) and attributes (e.g.
colour, position, size) to image data. The display and interaction stage provides the
user interface control that enables users to interact with data from multiple perspec‐
tives (e.g. drill down, expand), where visual analytics take place [13]. This is the
stage where meaning is interpreted by uses when interacting with data.
6 K. Liu and C. Tan

2.2 Challenges
Digital visualisation enables users to interpret and interact with data for a certain
purpose, such as aiding the decision making process. However, digital visualisation is
still prone to the semantic and pragmatic issue [adapted from 6]. Digital visualisation is
more than a simply graphical representation [14]. The semantic issue relates to whether
the visualisation conveys the right level of information to users. Users have a tendency
to establish a relationship based on what they see. Therefore, if the graphical represen‐
tation is not pointedly used, it will exacerbate the fallacies of the perceived meaning by
the users. On the other hand, the intention or purpose of digital visualisation is not
reflected, where digital visualisation is seen as an analytical reasoning process [15].
There is always a purpose how digital visualisation should be designed. There are usually
many data sources are required to achieve the purpose. Hence, this leads to challenges
such as data reliability, selection of the right information and the selection of the right
graphics to visualise information. Thus digital visualisation is only effective when there
is a clear purpose (guided by the pragmatic instance) on how data should be visualised
so it conveys a clear meaning to users (incorporated with the semantic aspect) [13].

3 Semiotics Perspective to Digital Visualisation

3.1 Signs and Norms

Semiotics, a formal doctrine of signs introduced by Peirce back in the 1930s [16], shows
great relevance to digital visualisation. A special branch is organisational semiotics
which has been developed by Stamper and his colleagues [17] to study the effective use
of information in business context. Data, under the study through visualisation, are signs.
The human interpretation on information is closely related to the five types of norms
(see Table 2 for the description of norms): perceptual, cognitive, evaluative, denotative
and behavioural [18]. Norms as a generic term are patterns, regulations, rules and laws
which are the reflection of knowledge in a cultural group or an organisation [19]. Norms
and signs (data) are intertwined [20]. In this research context, norms are employed to
interpret the signs; where a sign is a data object that conveys information and produces
an effect towards users in directing or prescribing users’ action, such as an action of
making a decision.

3.2 Abduction in Piercean Logical System

Abduction is a reasoning approach introduced in Piercean logical system [16], together
with induction and deduction. Abduction, according to Peirce [16], is a form of semiotics
interpretation [21], that guides the process of forming a hypothesis, which leads to
introducing new ideas. Deduction is the process of explaining the hypothesis; and
induction, is the process of evaluating the hypothesis. Abduction aims to find the best
explanation of the observed phenomenon [22]. In Piercean logical system, abduction
contributes to the human understanding of a phenomenon, by making sense of the
observed signs [adapted from 23].
 Semiotics in Digital Visualisation 7

Fig. 2. The process of digital visualization.

Kovács and Spens’s [24] and Thagard’s [25] propose the abduction process based
on Peirce’s [16] definition on abduction. The abduction process starts by a puzzlement
(an observed phenomenon) that leads to a search for explanation, followed by using
prior knowledge to establish a hypothesis and validating the hypothesis by conducting
more observations.

3.3 Digital Visualisation Is a Process of Abduction

From a semiotic perspective, digital visualisation is a process of abduction which is a
key process of scientific inquiry, i.e. a process of generating new knowledge. Peirce’s
abduction is a reasoning approach to fashion a hypothesis of the observed phenomenon,
hence the more clues that are collected, it is more likely the hypothesis leads to a truthful
conclusion [21]. Digital visualisation is therefore seen as a process of abduction from a
semiotics perspective; in which abduction is a key process of scientific inquiry for
generating new knowledge. On the basis, the process of abduction involves the norms
extensively. When encountering with a new phenomenon, prior knowledge will enable
us to produce some initial, but often, plausible explanations. Abduction thus allows us
to generate hypotheses (which should be plausible) and further to determine which
hypothesis or proposition to be tested [23, 26]. This aligns with Peirce’s definition on
abduction as “the process of forming explanatory hypotheses”, and the “only kind of
argument which starts a new idea” (Peirce, 1935).
Figure 3 illustrates the four key stages in digital visualisation as a process of
abduction: (1) search for explanation, (2) generation of hypotheses, (3) evaluation
and acceptance of hypotheses and, (4) effect from the hypotheses. In the search for
explanation stage, users establish an initial hypothesis based on what they see through
digital visualisation (e.g. graphical display of abstract information). This stage
commences the process of abduction. In the generation of hypotheses stage, users
8 K. Liu and C. Tan

Fig. 3. Digital visualisation is a process of abduction.

derive some patterns of knowledge based on their prior knowledge. At the same time,
users are anticipating for further insights grounded on what they see. Visual analytics
take place here to enable users interact with data by providing navigations such as
drill down for more information. In the evaluation and acceptance of hypotheses
stage, users verify or refute the perceived visual objects based on their prior knowl‐
edge. This stage adopts Popper’s [27] refutationism approach, where the formed
hypothesis can be refuted by the observed facts. And lastly, in the stage of effect from
the hypotheses, users reaffirm or refine the established hypothesis. The five types of
norms (perceptual, cognitive, evaluative, denotative and behavioural) correspond with
each stage in the process of abduction in digital visualisation, which will be illus‐
trated in Sect. 4.2.

4 The Principles of Digital Visualisation

Research leads us to believe that the three principles are important in digital visualisation
as illustrated in the following sub sections.

4.1 Digital Visualisation as Shared Semiosis

Digital visualisation is a sign-process that involves the production and consumption of

signs. Any digital visualisation will begin with the awareness of context. The design of
a digital visualisation is usually determined by certain purposes and effects that are
intended to be achieved. On the production side of this process, the producer may wish
to achieve a certain effect as he or she has certain intention and purposes through the
visualisation. The producer of the visualisation may situate the visualisation in a context
and relate it to some purposes. On the consumption side, a user, who could be any one
who sees the visual display in a different context and for a different purpose, may be led
to an effect that is same or totally different from what is intended by the producer. To
ensure the success of the communication through the use of visual display, it is important
 Semiotics in Digital Visualisation 9

Fig. 4. Shared semiosis process engaging producer and user.

to understand the semiotics in the production and utilisation of visualisation. Figure 4

illustrates the shared semiotic processes (i.e. semiosis) engaging the producer and user.
It is therefore vital to see digital visualisation as a shared semiosis; where there is an
agreement made between the producer and user on how and what data should be visual‐
ised, and producer should understand what purpose (or context) user tries to achieve
through digital visualisation.

4.2 Digital Visualisation as Norm-Centric Activities

Effective digital visualisation heavily relies on knowledge expressed as norms. Digital

visualisation is seen as an interactive artefact, which is usually designed based on
producer’s reasoning [28]. As mentioned in Sect. 4.1, the design of digital visualisation
should engage producer and user. Norms serve as the foundation and drive for the design
of a digital visualisation for the producer. On the other hand, the user incorporates norms
into the digital visualisation for discovering new knowledge. Both the producer and user
will invoke their norms in production (design of the graphical display) and utilisation
(use of the graphical display to gain new knowledge and perform certain actions).
Table 2 illustrates the five types of norms and their effect on the producer (or the
designer) and user by taking the digital visualisation design for the monthly sales report
for a department store as an example. The graphical display should clearly indicate how
the values relate to one another, represent the figure correctly and make it easy to navi‐
gate between figures. As digital visualisation is a process of abduction, the central role
of the subject who is engaged in the process must be recognised. If the goal of the digital
visualisation is to discover new knowledge, interactivity of digital visualisation to allow
the user’s manipulation of data in visualisation is essential.

4.3 Digital Visualisation as Artefacts as Well as Process

Digital visualisation is seen as an artefact, where it is a set of graphical displays (e.g.

charts and graphs) that conveys information. The graphical displays produce effect
towards designers (who determine the design of the graphical display) and users (who
use the displayed data to perform certain actions). The graphical display should clearly
translate the abstract information, that can be easily, efficiently, accurately and mean‐
ingfully perceived [29]. At the same time, there is an important feature (or it can be
called affordance) of digital visualisation is that it is not just static display but allows
(or affords) interactive inquiry. Digital visualisation hence is also seen as a process of
10 K. Liu and C. Tan

Table 2. Norms and their effect in Digital Visualisation (with an example of a monthly sales
report for a department store).
Types of norms Descriptions Effect on producer Effect on user
Perceptual Relate to how human Relate to the perception on Apply the prior knowledge
perceive signs in an sales figures. For to perceive the sales
environment. For example, sales figures figures. For example,
example, the distance are described by month, sales figures displayed
between two bollards is categories of products on the chart are
wide enough to drive and customer gender presented in month and
the car through categories
Cognitive Relate to how human inter‐ Affiliate with the interpre‐ Employ the prior knowl‐
pret the observed signs. tation of the sales edge to interpret the
For example, an illumi‐ figures. For example, sales figures. For
nated red light above an the low sales figure of example, the lowest bar
orange and green product A means the that represents product
means stop product is not selling A in the bar chart means
well in the particular the poor sales of
month. This leads to a product A in the month
graphic representation
to highlight the poor
sales
Evaluative Explain why human have Associate with the assess‐ Assess the performance.
beliefs, values and ment of the sales figure. For example, the low
objectives. For For example, lower sales figures identify
example, frankness in sales figures should poor performance
debating between attract more attention in through the aid of
employee and boss; the display by certain digital visualisation
openness in voicing distinctive shape or which otherwise may
personal views colour not be noticed
Denotative Direct the choices of signs Relate to the selection of Deploy the shared knowl‐
for signifying based on the graphical display edge of shaping and
the cultural back‐ that best represents the colour coding between
ground. For example, sales figure. For the producer and the
stop signals are red and example, red means user. For example,
octagonal under performance and certain shape or colour
green means acceptable represent certain mean‐
performance that meets ings
the expectation
Behavioural Govern human behaviour Relate graphic information Adopt the business knowl‐
within the regular with actions. For edge to perform
patterns. For example, example, the poor actions. For example,
if a library book is performance will be poor sales lead to
overdue than the linked with the produc‐ certain corrective
borrower must pay a tion of some warning actions such as
fine messages for attention strengthening
marketing effort or
reviewing the price

abduction. Affordance, originated by Gibson [30], is extended to the study of the real
world for understanding patterns of human behaviour. Digital visualisation can assist
human to interact with data to enhance understanding of the meaning of data, e.g. to
acquire more information. Furthermore, it will also enable one to discover pattern, regu‐
larities and norms from the visualised data, e.g. to obtain knowledge related to types of
problems and situations. With such interactivity, a user is able to explore data visually
to generate, refine and testify his or her hypothesis as an abductive reasoning process.
 Semiotics in Digital Visualisation 11

5 Conclusions and Future Research Insights

This paper initiates a new perspective on digital visualisation by positioning digital

visualisation as a process of abduction; through the generation, verification, refutation
and refinement of hypotheses. Some insights have been provided in the components and
process of digital visualisation and the semantic and pragmatic challenges embedded
within digital visualisation. Challenges in human perception and intention in digital
visualisation have been addressed by incorporating norms into the semiotics of visual‐
isation and visual analytics. This paper posits three key principles, i.e. digital visuali‐
sation is seen as: (1) a shared semiosis, (2) norm-centric activities and, (3) artefacts as
well as process. These three key principles contribute to how digital visualisation should
be designed (from the producer’s perspective) and how digital visualisation helps in
gaining new knowledge when new patterns and relationships are discovered (from the
user’s perspective).
This paper has established a vivid implication of semiotics to digital visualisation.
There are three main future research insights proposed on this basis. First of all,
researchers can adopt the three key principles suggested, especially the utilisation of
norms to increase the interactivity of digital visualisation. The interactive feature should
be able to draw and keep the attention of users, and enable users to drill down into data
and customise reporting [6]. In addition, users should be able to interact with each other
through information sharing.
The second potential area that researchers can look into is to understand the perspec‐
tive anticipated by users in manipulating the data. In relation to the interactivity of digital
visualisation, it is expected that the users will become more used to interpreting data
and creating new knowledge. Researchers should look into the human factors, the norms
in perceiving and reacting to visualisation result [14] (see Table 2).
The third area researchers can explore is the application of norms captured through
the abduction process into the technical implementation. Apart from the technicality of
the graphical display, researchers can focus on data collection, data processing and
analysis, in conjunction with the trend of big data. Digital visualisation should enable
users to do real time data analysis; and has the hold data in-memory in which it is
accessible to multiple users.

Acknowledgements. We would like to thank Dr Huiying Gao, visiting scholar to the Informatics
Research Centre of Henley Business School, University of Reading, from School of Management
and Economics of Beijing Institute of Technology, for her valuable and insightful comments to
this paper.

References

1. Chen, C.-H., Hrdle, W., Unwin, A.: Handbook of data visualization. Springer-Verlag,
Heidelberg (2008)
2. Czernicki, B.: Introduction to data visualizations. In: Czernicki, B. (ed.) Next-Generation
Business Intelligence Software with Silverlight 3, pp. 127–164. A press, New York (2010)
12 K. Liu and C. Tan

3. Robert, S.: Information Visualization - Design for Interaction. Pearson Educated Limited,
Essex (2007)
4. Ware, C.: Information visualization: perception for design. Elsevier, Somerville (2013)
5. SAS.: SAS Visual Analytics, 28 Feb 2014. http://www.sas.com/en_us/software/business-
intelligence/visual-analytics.html
6. Azzam, T., et al.: Data Visualization and Evaluation. New Dir. Eval. 2013(139), 7–32 (2013)
7. Tufte, E.R.: Beautiful evidence, vol. 1. Graphics Press, Cheshire (2006)
8. Chen, M., et al.: Data, information, and knowledge in visualization. IEEE Comput. Graph.
Appl. 29(1), 12–19 (2009)
9. Liu, S., Cui, W., Wu, Y., Liu, M.: A survey on information visualization: recent advances
and challenges. Vis. Comput. 1–21 (2014)
10. Baars, H., Kemper, H.-G.: Management support with structured and unstructured data—an
integrated business intelligence framework. Inf. Syst. Manage. 25(2), 132–148 (2008)
11. Zikopoulos, P., et al.: Harness the Power of Big Data the IBM Big Data Platform. McGraw
Hill Professional, New York (2012)
12. Singhal, A.: An Overview of Data Warehouse, OLAP and Data Mining Technology, In: Data
Warehousing and Data Mining Techniques for Cyber Security, pp. 1–23. Springer, New
york(2007)
13. Davenport, T., Merino, M.: State of the Art Practice with Visual Analytics. Harvard Business
Review, New York (2013)
14. Alexandre, D.S., Tavares, J.M.R.: Introduction of human perception in visualization. Int. J.
Imaging Robot. 4(10), 45–69 (2010)
15. Thomas, J.J., Cook, K.A. (eds.): Illuminating the path: The research and development agenda
for visual analytics. IEEE Computer Society Press, Los Alamitos (2005)
16. Peirce, C.S.: Collected Papers of Charles Sanders Peirce: Pragmaticisms and Pragnoaticism,
Scientific Metaphysics, vol. 5-6. Belknap Press, Cambridge (1935)
17. Stamper, R.: Organisational semiotics: Informatics without the computer. In: Liu, K., et al.
(eds.) Information, Organisation and Technology: Studies in Organisational Semiotics, pp.
115–171. Kluwer Academic Publishers, Boston (2001)
18. Stamper, R., et al.: Understanding the roles of signs and norms in organizations-a semiotic
approach to information systems design. Behav. Inf. Technol. 19(1), 15–27 (2000)
19. Liu, K.: Semiotics in information systems engineering. Cambridge University Press,
Cambridge (2000)
20. Stamper, R.: Information systems as a social science. In: Bouissac, P. (ed.) Information
System Concepts: An Integrated Discipline Emerging. IFIP — The International Federation
for Information Processing, pp. 1–51. Springer, New York (2000)
21. Moriarty, S.E.: Abduction: a theory of visual interpretation. Commun. Theory 6(2), 167–187
(1996)
22. Wirth, U.: What is abductive inference? In: Bouissac, P. (ed.) Encyclopedia of Semiotics, pp.
1–3. Oxford University Press: Oxford, New York (1998)
23. Yu, C.H.: Abduction? deduction? induction? is there a logic of exploratory data analysis?.
In: Annual Meeting of the American Educational Research Association, ERIC Clearinghouse,
New Orleans, LA (1994)
24. Kovács, G., Spens, K.M.: Abductive reasoning in logistics research. Int. J. Phys. Distrib.
Logistics Manage. 35(2), 132–144 (2005)
25. Thagard, P.: Abductive inference: from philosophical analysis to neural mechanisms. In:
Feeney, A., Heit, E. (eds.) Inductive Reasoning: Experimental, developmental, and
Computational Approaches, pp. 226–247. Cambridge University Press, Cambridge (2007)
 Semiotics in Digital Visualisation 13

26. Gregory, R. Muntermann, J.: Theorizing in design science research: inductive versus
deductive approaches. In: Thirty Second International Conference on Information Systems
(ICIS 2011), Shanghai (2011)
27. Popper, K.R.: Conjectures and Refutations, vol. 192. Routledge and Kegan Paul, London
(1963)
28. de Souza, C.S.: Semiotic engineering: bringing designers and users together at interaction
time. Interact. Comput. 17(3), 317–341 (2005)
29. Few, S.: Data visualization for human perception. In: Soegaard, M., Dam, R.F. (eds.) The
Encyclopedia of Human-Computer Interaction, 2nd edn, pp. 93–104. The Interaction Design
Foundation, Aarhus (2013)
30. Gibson, J.J.: The Ecological Approach to Visual Perception. Lawrence Erlbaum Associates
Inc, Hillsdale (1968)
 Data Fraud Detection: A First General
Perspective

Hans-J. Lenz(B)

Institute of Statistics and Econometrics, Freie Universität Berlin,

Boltzmannstr. 20 K30, 14195 Berlin, Germany
[email protected]

Abstract. We try to present a first broad overview on data fraud, and

give hints to data fraud detection (DFD). Especially, we show examples of
data fraud that happened at anytime of human mankind, all around the
world, and affects all kind of human activities. For instance, betrayers are
entities of the society, industry, banks, services, health-care, non-profit
organizations, art, science, media or even a government or the Vatican.
We consider four main areas of data fraud: spy out, plagiarism, manip-
ulation and fabrication of data.
Of course, there is not only interest on data fraud itself but on its
detection, too. Although improvements of data fraud detection is evi-
dent, it seems that the intellectual creativity and capacity of the betray-
ers is unlimited. Especially, the Internet with its various services and the
mobile communication opened the Pandora box for criminal acts. Fur-
thermore, one may state the hypothesis that while the ethics behavior of
people decreases over time the data fraud rate is continuously increasing.
There does not exist an omnibus data fraud detector, and the author
supposes there will be never one upcoming due to the heterogeneity of
the domain. For instance, compare the domains“spy out” in industry
and“data fabrication” of observational or experimental studies in sci-
ence. It is a matter of fact that the interest and need of science, business
and governmental authorities is increasing over time for improving tests
of data fraud detection. This paper can be viewed as a modest attempt
for stimulating research into this direction.

1 Introduction
One can best unfold the complexity of data fraud by classifying fraud into four
classes. All of them are driven by three time-invariant features of the societies,
business and human beings: Power, glory and money. As saying goes“Knowledge
is Power” power is the driving force of Spy out in the military and secret service
area, while knowledge and profit(money) are the main factors in business. The
activities of the secret services of all countries like Central Intelligence Service
(CIA) or NSA in USA, Military Intelligence (MI6) in UK, the Russian secret
service (CBP) etc. go back far into history. No doubt, they have ever had influ-
enced failure and success of military actions at any time. Less power and glory
but more profit (money) has been the forcing power of industrial or business
c Springer International Publishing Switzerland 2015
J. Cordeiro et al. (Eds.): ICEIS 2014, LNBIP 227, pp. 14–35, 2015.
DOI: 10.1007/978-3-319-22348-3 2
 Data Fraud Detection: A First General Perspective 15

spy out. There is no sharp boundary to the extensive“silent” storing and analytic
analyses of customer’s data, mostly without explicit permission. Consider only
the massive accumulation of such data by Amazon, Ebay or Google. Together
with nearly unlimited memory and Big Data Analytics it demonstrates the
increasing risk of leaving Recommendation Systems, simple pull/push systems,
and entering into a Total Information World with loss of any privacy of anybody,
anywhere and anytime as an extension of ubiquitous computing.
The second domain of data fraud is related to Plagiarism. It concerns mainly
the illegal usage of data of somebody else, sometimes existing as self plagiarism.
Since the start of the digital era plagiarism M.Sc. or Ph.D. studies have gained
more and more attention, although plagiarism is not limited to master or doc-
toral studies, cf. plagiarism happening in composing, drawing and painting. As
we shall see later the main interest in plag detection in the domain PhD disserta-
tions is caused by spectacular cases and the development of improved methods
(citation based plag detection besides of substring matching), [4], and“swarm
intelligence”. The last approach makes use of the efficiency of many, more or less
independently acting plag hunters pooling information about the same entity.
Our main interest, however, is devoted to Data Manipulation. Its main char-
acteristic is given by the dishonest manipulation of existing own or foreign data.
The objective is gaining prestige or making money. Evidently, data manipula-
tion happened at any time, anywhere and every domain of life is influenced. For
instance, the annual fraud rate of the new U.S. health care system is estimated
to be about 10%, [29]. Consider the scandalous manipulation of clinical trails at
the Medical University of Innsbruck (MUI), Austria, by Dr. H. Strasser, [19], or
the Libor 3(5)-months interest rate manipulation jointly performed by a cartel
of Deutsche Bank, Royal Bank of Scotland (RBS), Union Bank of Switzerland
(UBS), and Barclays Bank, UK, [20]. Greece fudged its annual depths-GDP rates
in the years before the country applied for entering the Euro zone, [22]. In recent
times the dishonest shuffling of ranks for TV shows by non-profit organizations
like ADAC, [23], or ZDF, Germany, [24], or even black money transfers and money
laundry by the Vatican, [25], caused much attention in the media.
A question of historical interest remains to be answered later: Who is known
to be the first betrayer in science? It is a tragedy of science that that profession
is more and more inclined to manipulate data gained from empirical (observa-
tional or experimental) studies. The causes are simple the pressure caused by
the dominating principle“Perish or Publish”, the increasing know-how needed
for applying sophisticated methods of data science, and the need of fund raising
for ongoing research. Yet, there exists a further variant of data fraud which is
far beyond the“simple manipulation” we talked about.
Data Fabrication. seems to be the most awful or“mostly criminal” form of
data fraud. Instead of manipulating data the betrayer selects an easier way of
collecting data: He simply generates the artificial data he needs. Evidently, this
safes cost and time, and, consequently, increases the chances of fund raising or
early publication because application forms or research papers can be submit-
ted with some time lead. Furthermore, the data will perfectly fit the betrayer’s
16 H.-J. Lenz

Ontology
of Data Fraud

Data Fraud types

Spy out Plagiarism Manipulation Fabrication

People, dishonest usage criminal change criminal production

Industry,…
Governments of data of others of real data of fictive data
E x.: Ex.: Ex.: Ex.:
NSA-Merkel C. Ptolemy(~ 100 ) Organic donations, Fraud in clinical
USA ( 2013 ) Munich ( 2012 ) trials, UK ( 2011 )

Data Security String matching, Benford's Law, Data- Inlier and outlier tests,
Data Privacy citation based model-conforming Model-conforming tests,
De/Encoding analysis, swarm tests, Inlier / outlier Checks of experiments,
intelligence tests, Research trust Research trust

Fig. 1. Data fraud types, examples and detectors.

hypothesis or objective if’professionally” created. A shocking case of such aca-

demic data fabrication is related to the Dutch psychologist D. Stapel, Univ. of
Tilburg, [21] whose forgery finally was stopped in 2013.
In the following we shall give a short history of spectacular cases of data
fraud. We pass by spying out, and turn to plagiarism with a special emphasis on
doctoral dissertations.We deeply look at data manipulation, fabrication and the
corresponding detection techniques. From the methodological point of view the
domains manipulation and fabrication strongly overlap. Consequently, the same
methods of detection can be used, i.e. Statistics, Data Mining, and Machine
Learning. Today, that bundle is labeled Data Analytics.
An overview on the four types of data fraud is given in Fig. 1 where we limit
ourselves to one example per each fraud type, and list some fraud detectors at
the bottom line.

2 A Short History of Data Fraud

To best of the author’s knowledge the first spectacular case of data fraud (data pla-
giarism) happened in ancient Egypt. Claudius Ptolemy ( 85 – 165 p. C.) was
a leading astronomer and mathematician at Alexandria, the center of the antic
world of science. He was the father of the Almagest, the famous Arabic star cal-
endar. He used the astronomical data from Hipparchos of Nicaea ( 190 – 120
a. C.), but he did not refer to him, [6]. Historians may argue here that the current
point of view of referencing might not be adequate for that science period. Today
such kind of fraud is called Data Plagiarism. The lesson for detecting such a
fraud is to carefully Check the Provenance of observational data referred to in
a publication.
 Data Fraud Detection: A First General Perspective 17

Almagest

World system of Plotemy(1661)

Ptolemy
Hipparchos

Fig. 2. Ptolemy, his centric world view, the Almagest, and Hipparchos, images: [10–13].

Galileo Telescope

Document about the

Heliocentric System

Inclined Plane

Fig. 3. Galileo, his heliocentric system, telescope and plane model, images: [14–17].

There is no doubt that Galileo Galilei (1564 – 1641) was one of the giants in
science. For example, he made ingenious contributions to astronomy and physics.
Especially, his astronomical research on the heliocentric system, and his exper-
iments with bodies sliding down inclined planes are famous. At that time, the
devices for measuring run time of moving objects produced“large” measurement
errors. Galileo believed in the validness of his cinematic hypotheses, and for
supporting them he decided to reduce the measurement errors. Today, there is
no doubt that Galileo’s experiments could not have been run in its way. As [7,
176] stated“The Genius was motivated by the objective of supporting the final
break-through of his ideas.” Refer to [8,27] for further details. Simply speak-
ing, motivated by prestige Galileo used data manipulation thus violating the
Principle of Reproduceability of well-performed experiments.
18 H.-J. Lenz

Fig. 4. Newton“Cuius genius humanibus superavit”, image: [18].

A further hero of science was Sir Isaac Newton (1643 - 1726) whose scien-
tific contributions to astronomy, mechanics (acceleration, gravitation and forces)
and mathematics (Calculus) opened a new era of physics, especially kinemat-
ics. In his book Principia Newton convinced due to the reported high precision
of his observations far from being legitimate, [7,9]. Or as [7, 176] or [9, 1118]
put it“Nobody was so brilliant and effective in cheating than the master math-
ematician.” Newton suppressed the real imprecision of his measurements being
anxious about a non perfect fit and casting doubts on his theory in the sci-
entific community. In order to support his hypotheses he faked the output of
his experiments and the observations by downsizing the errors. Accordingly, this
manipulation contradicts the Principle of Reproduceability. His motive certainly
was not profit, but fear of loss of prestige.
Data manipulation has many variants as we shall see later. One form is the
trick used by experimenters to select and publish only a proper subset of results
or runs. A“representative” case is given by the physicist Robert Millikan, US
Nobel Price Winner, 1868 – 1953. The claim is his strikingly precise measure-
ments of the charge of electrons in 1913 being quite better than those of his rival
Felix Ehrenfeld who experienced large deviations, [7,8]. Milikan’s lab protocols
showed later that he published only the ‘best 58 out of 140’ experiments having
smallest measurement variance, [8,34] and [7, 176–177]. Today such a misbe-
havior of running experiments is called Experimental Selection Bias and
contradicts, of course, the Principle of Reproduceability. Here again Prestige
combined with Nobel price winning expectations was presumable the driving
force of data fraud.
We continue presenting historical cases of data fraud and turn to the Deutsche
Bank as an example from the business sector. Alternatively, we could have
selected from the U.S. economy the Enron case with falsification of balance
sheets, [47]. Josef Ackermann, born 1948, was the chairman of the board of
directors 2002–2012, and was responsible for all claims. Under his leadership the
Deutsche Bank was involved in many scandals like manipulating the prices of
 Data Fraud Detection: A First General Perspective 19

Fig. 5. Libor manipulation by the cartel, image: [20].

food products, unlawful CO2 -emission permits and financial derivatives trading,
and, last but not least, manipulating the 3(6) months Libor interest rate within
a bank cartel consisting of Barclays, Deutsche Bank, RBS and UBS, [20]. The
legal task of the group was a“fair price fixing” of both interest rates, but the
cartel used its economic power for down or upraising of the Libor for making
moderate but continuous profits, cf. Fig. 5.
Another domain where forgery typically happens is health care, cf. [29] for
fraud detection in the new U.S. medicare system. Especially medical universities
are affected institutions because of the pressure to economically manage clinics,
fund raising for research or corruption. The last motive was underlying extensive
sport doping and manipulation of facts and figures at the Univ. of Freiburg,
Germany, cf. [27]. Above we already mentioned the case of illegal experiments
at the Medical University of Innsbruck (MUI), Austria, as reported by Nature,
[19]. In Germany tricksters manipulated the lists of organic transplantations
at several of the 49 German transplantation centers like Göttingen, München,
Münster, Regensburg etc. [26]. In order to detect fraud in such cases independent
experts are employed who checked with diligence all mails, lab protocols, lists,
reports, time-schedules and revenues and expenditures vouchers with respect
20 H.-J. Lenz

to facts and prior knowledge in a qualitative and quantitative way. We call

these (qualitative) manual activities assisted by analytical techniques Manual
Inquest of Data Fraud. They are based on Abductive Reasoning, and are
similar to trouble shooting from an investigation point view.
Next, we turn to economics. In this domain data fraud happens for gain-
ing more economic influence or reward. As an example take Greece in the late
nineties. The crucial period for Greece to enter the Euro zone was 1997 − 1999.
Greece manipulated its official macro-economic statistics, and made the EC
member countries, the EC authorities in Bruxelles and Luxembourg and the
public believe that the deficit had fallen under the Maastricht limit of 3%. How-
ever, later the (new) Greece Finance Minister, Mr. Aligoskoufis, confessed: “It
has been proven that the deficit had not fallen below 3% in every year since
1999.” [22]. Indeed, the figures for 1997-1999 were 6.44, 4.13 and 3.38%.
Let us close with politics and consider governments who do not completely
obey democratic rules. It is a matter of fact that the elections in Russia 2011
were fudged, [28]. Especially, the result of Putin’s party United Russia is doubt-
ful. Consider the empirical distribution of electoral votes on the percentages of
votes in the interval [0, 100]. Typically, such a distribution is bell-shaped and is
smoothly curved in the center and at both tails. Due to the evident manipulation
of votes the distribution of Putin’s party is heavily skewed to the right (large
percentages), and shows small up-and-down spikes above 30% around the values
35, 40, 45, . . . , 100%, cf. Fig. 6.
The lesson learned from the last type of data manipulation that sometimes
simply plotting of histograms (empirical distributions) is an effective “starter”
for further investigations on suspectable data.

no of electoral districts

Fig. 6. Manipulation of Russian elections in 2011, image: [28] layout modified.

3 Spy Out
One can raise the question whether or not Cheating by copying text or formulas
from neighbors during examinations at school is an entry into the world of spy
 Data Fraud Detection: A First General Perspective 21

out or plagiarism, and is abnormal or not. No doubt that kind of spying out is
happening at a very low fraud level.
From a historical perspective massive Spy out is related to ancient and
modern populations and their military actions against other countries. It may be
considered as a prerequisite of any war and struggle, and aims for getting better
information about the enemy for improving the success of their own attacks.
Think only of the empires Greek and Persia, Egypt and Hittites, or Roman
and Germanic. The secret services of all modern countries around the globe
are expected to deliver information about the rest of the world. After 9/11 the
National Security Agency (NSA), USA, has scaled-up its computing facilities
for spying out everybody’s data on an extreme large scale. The question arises
whether “Big brother is watching you” is realized or not. Here any kind of world-
wide communication by phone, email, Twitter or any other media is affected.
Even encrypting is not at all safe and a guarantee for people’s privacy. The fact
that even the German chancellor’s telephone is tapped by US authorities is a
nasty perspective.
Let us leave this domain and turn to economics and business. Telephone fraud
became roughly a problem at the end of the eighties. Legal telephone cards with
or without a credit were stolen and manipulated by betrayers which made phone
calls possible free of any charge. Corruption related to data fraud was recently
reported by the German Press Agency and printed in Der Tagesspiegel, a leading
newspaper in Berlin, Germany. Employees of the sanctuary health assurance
company Debeka were accused of having tried to get data of state employees
candidates from corruptive state employees targeting for new contracts, [5].
As buying and selling is an intrinsic part of daily life of consumers, tricksters
and betrayers consider trade as their domain for data fraud. Each trade has a
phase “selecting and putting an item into the shopper basket” and “paying”.
Paying is mostly done by cash, alternatively by EC or credit cards. Forget cyber
money here due to its minimal market share. Card Fraud is concerned with
card theft or the misuse of bank account and PIN data. Therefore, we have the
following four classes of fraud, cf. [2]:

– Theft
– Duplicate Generation
– Skimming
– Pishing.

Theft is a real fraud if a credit card like VISA, AMERICAN EXPRESS

or MASTER is stolen and used for money withdrawals. However, when a card
theft is combined with Card Cloning (in the sense of duplicate production)
things become more tricky and dangerous, and increase the card owner’s risk of
big monetary losses. The thief simply can withdraw money from various sites.
Skimming starts at the other end of the buying chain just to say so. Before
a consumer pays by cash he has to collect money, for instance, from a teller
machine (ATM). Tricksters install a small web cam and a thin keyboard at the
ATM’s site for catching the bank account or credit card number together with
22 H.-J. Lenz

Banking Fraud

EC Card Credit Card Web/Internet

Card Theft Click Cheat

Duplicate use
Identity Theft Phishing
(Phishing)

Fig. 7. Fraud of banking and internet banking.

the PIN from an innocent client. Phishing refers to sending faked emails to
users with a subject like “Full mail box” or “Special offer”, and asking them for
delivering their account, password (PIN) and transaction number (TAN). In a
similar way, novice surfers are redirected to false (criminal) links or attracted
by special (low priced) offers for visiting specially prepared web pages. When
such sites are entered the trickster fishes the bank account, PIN or name, credit
card number and safety code for starting his own criminal transactions. Let us
add Advertising Click Fraud where Pay per click systems are manipulated
by artificially generating clicks by the betrayer for illegally making money at the
expense of the advertising company, [2].
For an overview by a diagram of data fraud related to card usage or Inter-
net banking have a look at Fig. 7. There exists a big bundle of Multivari-
ate Explorative Data Analysis Techniques for detecting EC or credit card
fraud, especially methods like Generalized Linear Regression and Classification.

4 Plagiarism

As noted above Plagiarism is considered to be the dishonest use of data of a

second party. It has many facets and exists in many fields of human life like arts,
business and science. The motives are either prestige or mostly profit. A short
overview is given in the diagram in Fig. 8 below.
We present famous cases of plags: Pirating of documents (“Galileo Forgery”)
and doctoral dissertation plag of the former German minister of defense, K.-T.
von Guttenberg. Both cases happened in Germany, and the forgery was proven
in 2012 and 2011, respectively.
 Data Fraud Detection: A First General Perspective 23

Data Plagiarism

Art Science Business

Product Piracy
Painting A …Z

Counterfeiting
Composing Money

Poetics

Fig. 8. Selected domains of plagiarism.

Fig. 9. Discrepancies between the original (“Sidereus Nuncius”) and faked book, image:
A Galileo Forgery.

Galileo’s book “Sidereus Nuncius” was published in 1610, and it is a fact

that 80 legal copies are known to exist worldwide. In 2005, surprisingly, a fur-
ther (faked) copy including until then eighty unknown (faked) ink drawings was
offered by art dealers on the international market. The German expert Horst
Bredekamp, Humboldt Univ. Berlin, used the drawings from the faked copy in
a book he published entitled “Galilei, der Künstler”. Finally, in 2012 a British
historian proved the forgery done by the Italian M.M. De Caro. His detection
caused Bredekamp et al. to edit the book “A Galileo Forgery” where he compiles
the events and explains what happened. It is interesting to note that besides of
various material analytical techniques Pattern Matching of symbols (stars)
and characters (L) helped discriminating between the original and faked docu-
ment, cf. Fig. 9.
One of the most spectacular data plagiarism claims corresponding to doc-
toral dissertations was caused by the work of K.-T. von Guttenberg, the former
24 H.-J. Lenz

minister of defense of Germany. Today his dissertation can be considered one

of the best investigated plagiarism. It may be even considered as a yardstick
for further improvements of academic plagiarism detection software. While most
of the plag software uses simple Substring Matching techniques it was the
idea of B. Gipp (2011), [4], to use a kind of meta analysis based on information
retrieval methods for developing his Citation based Plag Detection (CbPD)
technique. Several steps are built-in: Identify citation patterns by order, non-
lexical proximity and distinctiveness of in-text citations. This idea enabled him
to detect multi-lingual and multi-source detection plagiarism by cross-reference
of the citation structures of an original and its plag. All computational output
is carefully visualized. Especially, the hit rate of detecting disguised plag like
paraphrases and cross-language plag could be clearly increased. It is a great
advantage of Gipp’s approach that his procedure produces higher confidence
levels than the existing matching algorithms for academic plagiarism detection,
[4,30]. Simple string matching or substring search work as follows, cf. the plag
platforms citeplag, Vroniplag etc. Let A be an alphabet, and s1 ∈ Am and
s2 ∈An literal text strings where m ≤ n. Prove s1 ⊆ s2 or sim(s1 , s2 ) > slow !
We present a barcode representation of the thesis produced by manual inspection
done by the members of the GuttenPlag Wiki project, see Fig. 10. The semantics
of coloring is as follows:

– red bars: multiple sources plag pages

– black: single source plag pages
– white: no plag pages
– blue: content and bibliography pages.

It was found that 64 % of all lines of the text was plagiarized. More details of
v. Guttenberg’s plagiarism case can be unscrambled by using the plag location
and visualization prototype CitePlag developed by Gipp(2013). The software
offers five citation-based approaches using two documents as input - a plag can-
didate and the original - [30]. We present only one illustration here. In Fig. 11
the left image shows the dense citation patterns produced by the Bibliographic
Coupling method in v. Guttenberg’s thesis. The image to the right shows the
concept of citation chunking. Numbers represent matching citations occurring
in both documents, and the letter x indicates non-matches.

Fig. 10. Barcode of Guttenberg’s thesis by plag type, image: [31].

 Data Fraud Detection: A First General Perspective 25

Fig. 11. Citation chunking of Guttenberg’s thesis and its concept, images: [30].

– Bibliographic Coupling is a similarity measure between two reference lists

– Citation Chunking is a citation pattern matching irrespective of the order of
matching citations
– Greedy Citation Tiling identifies longest citation patterns consisting entirely
of matching citations in the exact same order
– Longest Common Citation Sequence searches for the longest string of citations
matching in both documents in identical order
– Longest Common Sequence of Distinct Citations includes only the first occur-
rence of a matching citation. It ignores repeated citations of the same source
regardless of occurring in the same order in both documents or not.

5 Data Manipulation and Fabrication

As mentioned above Data Manipulation and Fabrication overlap from a
methodological point of view. Of course, the data generation process is quite
different. Therefore it makes sense to present fraud detectors for both fraud
kinds together.
Data manipulation is the dishonest change of the content of existing own or
third party data, irrespectively, whether the content is encapsulated in text doc-
uments or not, i.e. tables, diagrams or photos. In most cases numbers are manip-
ulated. Data fabrication is the criminal production of artificial figures driven by
gaining power, prestige or profit (“Gier”).
In the following we mainly focus on data fraud in science. There exists a lot of
studies on the various types of fraud. A recent field study based on interviews and
questionnaires was published in Nature in 2005, and was authored by Martinson,
Anderson and de Vries, [32]. They addressed N = 7760 scientists who got a grant
from the National Institutes of Health, USA. The final sample size was n = 3247,
i.e. the response rate is only slightly less the 42%. The field study was designed
as an anonymous self-report based on a standardized questionnaire with (10 + 6)
items of strong and medium fraud types. The study differed between early and
mid-career researchers. Figure 12 summarizes the quite disappointing aggregated
results.
26 H.-J. Lenz

Martinson et al Study, 2005 Bad storage

Change of design etc.

Dropping obs.

using inadeqaute designs

ignoring data etc.

secret mongering

doubtful authorship

Fig. 12. Types of data manipulation, compiled in the Martinson et al. study, image: [32].

Empirical data is no longer accessible, designs are doubtful, and inappro-

priately application of statistical methods like sub-setting, sequential testing or
manipulation of p-values. From the author’s point of view the Martinson et al.
report supports the conjecture of time-stability of such findings in psychology,
social science and medicine. Further contributions supporting this hypothesis
are due to J.P.A. Ioannidis and L. John, [33,34]. Again, remind yourself that
data manipulation exists everywhere at any time in science. For instance, the
Japanese stem-cell researcher H. Obokata, Kobe, was forced in 2014 to retract
her co-authored papers in Nature because of four main allegations raised about
her research, [35]:
– Irregularities of published images
– Identical text copied from another own paper without reference (self plag)
– Inconsistencies between published papers and later author’s explanation
– Non Reproduceability of experimental results.
Next we present some few cases where statistical methods are not correctly
used leading to surprising results. One has to confess that the borderline between
non professional usage of statistics and data manipulation is fuzzy.
– Although the diastolic and systolic blood pressure are stochastical dependent,
doctors use them independently of each other for their diagnoses worldwide.
– Nowadays, doctors generally consider the PSA value or the related 1 or 5% -
confidence intervals as an important tumor marker of the prostate carcinoma.
As the calcium, phosphate and PSA value are feedback controlled by the
human body, it is doubtful to consider one-dimensional confidence intervals,
thus ignoring correlation.
– Fixing a (nominal) probability of the error of first kind, α , and testing for
various hypotheses based on the same data set increases strongly the effective
underlying α. For instance, checking 20 hypotheses and fixing αnom = 5%
leads to αef f ≈ 64%. In the long run every hypothesis is “accepted”. There
exists a Bonferroni correction, but this must be handled with care, too.
 Data Fraud Detection: A First General Perspective 27

5.1 Benford’s Law

In the following we focus on numerical data fraud including manipulation and fab-
rication of figures. It should be stressed that we shall present only a real subset
of statistical and related methods. Clustering, classification, Generalized Linear
Models or even case-based reasoning truly belong to any anti-fraud tool box. For
instance, the last methodology is treated in the context of fraud detection in [46].
Consider a “homogeneous” data set or corpus like revenue and expenditure
transactions in business, assurance claims in health care, main economic indi-
cators of an UNO membership country etc. It was Newcomb (1881) who first
described the phenomena when detecting unexpectedly many usage spots at the
digit 1 of some tables of logarithms at hand, [36]. Benford (1938) gave the first for-
mal proof, [37], and later on, Hill (1995) added more technical details, [38]. Very
interesting examples from a broad range of domains can be found in [39].
Simply speaking, the Benford’s Law states that the distribution of the lead-
ing numeral D1 with range 1, 2, . . . , 9 of figures from a well defined corpus C
obeying the law is given by PC (D1 = d)) = log(1 + 1/d). The formula can be
generalized to the first k digits, i.e. PC (D1 = d1 , D2 = d2 , ..., Dk = dk )) =
log(1 + 1/(d1 d2 ...dk )). Of course, not all data sets are distributed according to
the Benford’s Law. Sets of identifiers used in database systems or a taxonomy or
house numbers are counter examples. Pinkham (1961) proved that for the distri-
bution fX to be a Benford probability distribution it is equivalent to be scale and
basis invariant [40]. Let us illustrate the law, and assume a betrayer manipulated
a set of bookings by adding faked ones with amounts between Euro 6100 − 6900.
A plot of the empirical and Benford’s Law signals the forgery, see Fig. 13.

5.2 Data-Model Conforming Tests (DMCTs)

Functional relationships between main business indicators are mostly based on

the four arithmetic operations. If we assume that the included variables are
superimposed by observational or measurement errors they can be modeled as

Fig. 13. Manipulated bookings with too many items between Euro 6100 − 6900.
28 H.-J. Lenz

random variables. Consequently, their functional dependency is captured by a

(generalized) regression model with errors in the variables, [45]. In the linear
case (addition and subtraction) the unobservable (true) values of the variables
can be estimated exactly using the Method of Generalized Least Squares
(GLS). However, if the variables are linked by multiplication and division a
Taylor Approximation of first kind developed around the mean vector becomes
necessary.
Let (x, z)0 ∈ Rp,q be noisy observation vectors of the state equation sys-
tem x = ξ + u, and balance equation system z = ζ + v = H(ξ) + v
where H :Rp → Rq . Note, that we use small letters for random vec-
tors. If the balance equations model linear relationships they reduce to the
matrix equation ζ = Hξ. For example, in the  univariate case we have
Revenues
P rof it = Revenues − Expenditures = Hξ = (1, −1) repre-
Expenditures
senting a linear equation. The fundamental economic relation T urnover =
Quantity × P rice = H(Quantity, P rice) is of non-linear type.
As proven in [45], (ξ, ζ) ∈ Rp,q can be estimated by GLS with maximum
precision under the hypothesis of Gaussian noise, a linear system and uncorre-
lated noise, i.e. (u, v) ∼ N (0, Σuv ). From the computationally point of view the
estimation problem reduces to minimizing a quadratic form:
 
ˆ T −1 u
ξGLS = argmax{(uv) Σuv } (1)
v
and
ζ̂GLS = H ξˆGLS . (2)
The variances on the diagonal of the covariance matrix Σuv are assumed to be
completely known due to prior information while the correlations are set to zero.
This means that for each variable the observational or measurement error should
be known. To some reasonable extent this can justified by the principle Mini-
mal Specifity saying that the covariance matrix Σuv has a block structure, and
all correlations (off-diagonal-elements) vanish. The estimators above have mini-
mal estimation variance according to the Gauss-Markov Theorem, [48]. If the
relationship is linear but non-normality must be assumed ξˆGLS , ζ̂GLS are still best
linear unbiased estimators. However, if products or ratios exist as operators, the
Gauss-Markov Theorem is only approximately true. In any case the estimates have
some convenient characteristics for detecting data-model non-conformity, [45]:

ˆ ζ̂) fulfill – except for numerical imprecision – the system of balance equa-
1. (ξ,
tions and
2. Σ̂ξ̂ ≤ Σx and Σ̂ζ̂ ≤ Σz where relational operator “≤” is to be applied to each
single component.
3. “Large” deviations between an observational value and its corresponding esti-
mate are a hint to non-conformity and to reject the data-model consistency.
This can be statistically tested, [45]
4. Error free variables are not changed.
 Data Fraud Detection: A First General Perspective 29

Fig. 14. Reduced DuPont-System as a spreadsheet.

Let us consider a simple illustrating example. We consider the famous

DuPont-Model which is presented in Fig. 14. We focus on the (linear) bal-
ance equation Sales = Cost + P rof it. Note that we used German labels for the
variables of the model. More formally, ζ = ξ1 + ξ2 . Let the measurements be
imprecise. The observations and the absolute errors of the three quantities are
as follows: Sales(z) = 100 ± 5, Cost(x1 ) = 80 ± 4 and P rof it(x2 ) = 30 ± 1.5.
Evidently, the measurements (x1 , x2 , z) do not satisfy the balance equation
because 100 = 80+30. GLS estimation using the software Quantor ([49]), delivers
the following consistent estimates and estimation errors: ζ̂ = 110 ± 3, ξˆ1 = 85 ± 3
and ξˆ2 = 25.6±0.9. We confirm 110 ≈ 85+25, 6. As mentioned above data-model
consistent estimates reduce the imprecision of a data set (or leave it unchanged)
given levels of 90, 95, 99%-confidence. For instance, εz = 5 > ε̂ζ̂ = 3. Alternative
approaches of GLS are Fuzzy Logic or MCMC Simulation. Details can be
found in [2].

5.3 Inlier Detection

The Benford Law makes clear that a trickster who tries to manipulate numerical
data must be careful and skillful doing so because the law imposes logical restric-
tions on a “human number generator” as we have seen. Generally, further tests
on data manipulation and fabrication exist implying more and more restrictions
on fudged figures.
Some betrayers fearing to be detected because of generating too large num-
bers, prefer to do the opposite, i.e. produce “too many” numbers near the aver-
age. Such a value is called Inlier in Statistics. Roughly speaking, inliers represent
a dense cluster of data items around the mean related to a different density.
A log-score approach of inlier detection under the quite strong assumption
of independence among the random variables considered is due to Weir and
Murray (2011) and leads to a “quick and dirty” treatment of the problem, [41].
The procedure is as follows.
30 H.-J. Lenz

ϕ (x)

‘near to the Mean’

Fig. 15. Potential inlier range under a Gaussian regime.

Procedure Inlier Test

{Weir and Murray (2011)};
Input: Problem size (n,p),
confidence level (1-alpha),
data matrix X_(nxp);
Output: Scores szi^2, ln szi^2 for i=1,2,?,n
begin
Standardize data by z=(x-my)/sigma
for all p variables and n test objects;

Compute squared score z_i^2 by summing z_ij

over all p variables and n objects;

Perform chi^2 test;

Plot ln-scores ln z_i^2 against i=1,2,...,n
end.

We illustrate the approach by a fictive case study: Consider the energy con-
sumption (electric power and gas ) of Company X. Thus we have p = 2 variables
of interest. The company runs five factories in each of five districts. Altogether,
we have n = 25 objects of interest. As the management has suspicion whether or
not the factories of a specific district falsify their figures the score test supports
assessing the risk of data manipulation.
Figure 16 gives a hint that in district no 3 inliers are present, and that the
reported data may be manipulated. While the overall mean of the five districts
¯ 2 = +0, 29, we observe the mean of district no 3 to be exceptionally small,
is lnz
lnz3 = −1, 53. In such cases the management of company X should start trouble
2

shooting for finding out the causality of what has happened.

 Data Fraud Detection: A First General Perspective 31

District D i , Factory Bj

Fig. 16. Hint of inlier generation at all factories of district 3.

5.4 Outlier Tests

Now we turn to the problem of detecting outliers. The motive of tricksters pro-
ducing outliers either by manipulation or fabrication of data is typical profit
making when participating in investment banking, stock market selling or pri-
vate borrowing.
A naive approach for detecting outliers is given by the “popular” 3-sigma
Rule. Let us assume that the amounts of transactions, say, have a normal dis-
tribution, i.e. X ∼ N (μ, σ 2 ). Note, that generally the moments of the Gaussian
distribution are unknown and must be estimated from the sample x1 , x2 , . . . , xn ,
mostly assuming identically and independently distributed (i.i.d) observations.
The hypothesis of the test is H0 : x generated by N (μ, σ 2 ). The alternative hypoth-
esis is H1 : x not generated by N (μ, σ 2 ). Under a Gaussian regime the confidence
level is (1 − α) = 0, 9973 for a 3σ-confidence interval. Consequently, the rejection
area of the related outlier test given a suspectable observation x ∈ R is reject H0 if

|x − μ|/σ > 3. (3)

As the unknown parameters must be estimated from the sample which possi-
bly includes outliers the Masking Effect is caused, [42]. It leads to the masking
of outliers by distorting the estimated mean and standard deviation. The mask-
ing effect can be quantified as follows. The 3σ-rule is ineffective to locate outliers
with the same sign at a rate pI = 1/(1 + λ2 ), [42]. For example, let λ = 3. It
follows pI = 10%. We visualize the masking effect in Fig. 17.
Evidently, the outlier x11 can only be certainly located if the true value of
(μ, σ) is known. Even using robust estimation by substituting the standard devia-
tion s by the median (MED) of the absolute deviations around the overall median
x̃ as Hampel (1985) proposed, i.e. M AD(xν )ν=1,2,...,n −MED(|xν − x̃|)ν=1,2,...,n ,
no convincing improvement of the outlier detection is recognizable.
32 H.-J. Lenz

Fig. 17. Masking effect of outlier detection.

In our context, a more efficient outlier location rule is the c-MAD Rule pro-
posed by [42]. The idea is to determine the proper (1 − αn ) percentile c of the
distribution of M AD(xν )ν=1,2,...,n by Monte-Carlo simulation. Reject x ∈ R if

|x − x̃| > cn,αn M AD(xν )ν=1,2,...,n (4)

where α = αn = 1 − (1 − α̃)1/n and cn,αn is simulated by solving the inequality

Pn,αn (X ∈
/ {x ∈ R||x − x̃| > cn,αn M AD(xν )ν=1,2,...,n }) ≥ 1 − α̃ (5)

where α̃ ∈ (0.5, 1) is given. Davies and Gather(1993) showed by example that

for a nominal value of α̃ = 5% the critical values of cn should be set equal to
c20 = 3, 02, c50 = 3, 28, and c100 = 3, 47, [44]. This means that except for very
small sample sizes the 3σ-rule is misleading.
A straightforward generalization of outlier detection in multi dimensional
data spaces is to use the Mahalanobis distance, cf. [45]. Assume X∼ N (μ, Σ)
with μ ∈ Rp and Σ ∈ Rp×p known. Then reject x ∈ Rp if

(x − μ)T Σ−1 (x − μ) > χ2p,1−α . (6)

Of course, the same problems as in the univariate case arise:

– unknown parameters (μ, Σ)

– non normality (skewness, mixed distributions)
– outlier robustness.

Alternative approaches to be investigated are given by an order statistics or

convex hull confidence regions approach. The bottleneck here will be the curse
of high dimension.
 Data Fraud Detection: A First General Perspective 33

6 Outlook and Perspectives

Data Fraud Detection is a hot topic not only since computers and the Internet
dominate our daily life. We conjecture that the relation fraud rate ∝ 1/ethical
attitude is true and there seems some evidence of continuously decreasing ethics
in economics, business and society. All important domains like industry, science,
public service, press, clinical and pharmaceutical research, health care, religious
communities etc. are affected by data fraud. Therefore there exists a great need
for improving methodologies and tools for data fraud detection. In science, one
step into the right direction of increasing the transparency among the scientific
community is by the notification of cross-referencing of ongoing related research.
An example for such an authority is the start-up ResearchGate at Berlin,
Germany.
Furthermore, independent, international scientific data centers are needed
where published data must be deposited. The Principles of Repeteability and
Reproduceability are to be obeyed in any case. Observational and Experimental
Selection Bias should become a taboo. This implies checking the correctness and
soundness of experimental designs, collecting data schemes, and applying sound
statistical methods. However, such an inspection of data sets of a third party is
not a very inspiring task for any researcher!
Finally, we have to admit that there exist vague boundaries between data
fraud, fudge, falsification, appraisal, cheat, deception, and scouting. But this
should not stop data fraud hunting at all. The scene reminds a bit of the rela-
tionship between betrayers and detectives. As the saying goes:“In the long run
we get them all!” Or as Di Trocchio put it: “Fraud has been since ever an art.
Recently it has become a science.” [6].

References
1. Akkaya, A.D., Tiku, M.L.: Robust estimation and hypothesis testing under short-
tailedness and inliers. Test 14(1), 129–150 (2005)
2. Müller, R.M., Lenz, H.-J.: Business Intelligence. Springer, Heidelberg (2013)
3. http://de.wissenschaftlichepraxis.wikia.com/wiki/Untersuchungen zu Datenfal
schung und schlechter Wissenschaft. Accessed: 19 June 2014
4. Gipp, B.: Citation-based Plagiarism Detection - Applying Citation Pattern Analy-
sis to Identify Currently Non-Machine-Detectable Disguised Plagiarism in Scientific
Publications. Doctoral Dissertation, Technical University Magdeburg (2013)
5. German Press Agency (dpa): DEBEKA Polizei im Haus. Der Tagesspiegel, Nr.
22106, July 17, (2014) 15
6. di Trocchio, F.: Der große Schwindel Betrug und Fälschung in der Wissenschaft.
Campus-Verlag, New York (1994)
7. Sheldrake, R.: Sieben Experimente, die die Welt verändern. Scherz Verlag, Berlin
(1996)
8. Broad, W.J., Wade, N.: Betrayers of the Truth. Oxford Paperback Reference,
Oxford (1985)
9. Westfall, R.S.: Newton and the fudge factor. Sci. 179(4079), 751–758 (1973)
10. World System. http://de.wikipedia.org/wiki/Almagest. Accessed 18 July (2014)
34 H.-J. Lenz

11. Plotemy. http://en.wikipedia.org. Accessed 18 July 2014

12. Almagest. http://ibiblio.org. Accessed 18 July 2014
13. Hipparchos. http://myastrologybook.com. Accessed 18 July (2014)
14. Galilei, G.: http://wundervollesrom.com. Accessed 18 July 2014
15. Galileo’s document on heliocentric System. http://medienwerkstatt-online.de.
Accessed 18 July 2014
16. Galileo’s Telescope. http://www.museum.vic.gov.au/scidiscovery/images/mn0063
09w150.jpg. Accessed 18 July 2014
17. Galileo’s inclined plane. http://sciencedemonstrations.fas.harvard.edu/icb/icb.
do?keyword=k16940&pageid=icb.page80863&pageContentId=icb.pagecontent
341734&state=maximize&view=view.do&viewParam name=indepth.html.
Accessed 18 July 2014
18. Godfrey, K.: Newton Portrait. National Portrait Gallery, London (1702)
19. Nature: Austria’s most serious report of scientific misconduct in recent memory
must be handled properly. (The scandalous behavior of Dr. H. Strasser at MUI,
Innsbruck, Austria) Nature 454, pp. 917–918, 21 August 2008
20. Ch. Siedenbiedel: Die Libor-Bande. Frankfurter Allgemeine Sonntagszeitung
(FASZ), pp. 21–22, no 6, 10 February 2013
21. Tilburg Univ.: Prof. Diederik Stapel suspended. Press release Tilburg University,
September 7, (2011). Accessed 17 September 2011
22. Howden, D.: St. Castle: Greece admits deficit figures were fudged to secure Euro
entry. The Independent, 16 November 2004
23. Gennies, S.: ADAC gibt Manipulationen zu. Der Tagesspiegel, vol. 15, no. 21933,
20 January 2014
24. S. Alvarez, J. Huber: Mit dem Zweiten trickst man besser. Der Tagesspiegel, vol.
24, no. 22402, 13 July 2014
25. Sanderson, R.: The scandal at the Vatican bank. Financial Time maga-
zine, 6 December 2013. http://www.ft.com/cms/s/2/3029390a-5c68-11e3-931e-
00144feabdc0.html. Accessed 21 July 2014
26. Organ Transplantationen. Der Spiegel, pp. 42–44, no 3 (2013)
27. Hartmann, G.: Die Doping-Uni vertuscht ihre Doping-Vergangenheit. Zeit Online
section Sport, 7 February 2013
28. Ziegler, G.M.: Keine Wahl. In: Der Tagesspiegel, Nr. 21816, vol. 31, 21 September
2013
29. Thornton, D., van Capelleveen, G., van Hillegersberg, J., Mueller, R.M.: Outlier-
based health insurance fraud detection for u.s. medicaid data. In: Proceeding of
CD on Special Session on Information Systems Security - ISS, ICEIS 2014, Lisboa
(2014)
30. Gipp, B., Meuschke, N., Breitinger, C., Pitman, J., Nürnberger, A.: Web-based
demonstration of semantic similarity detection using citation pattern visualization
for a cross language plagiarism case. In: Proceeding of (CD) on Special Session on
Information Systems Security - ISS, ICEIS 2014, Lisboa (2014)
31. GuttenPlag Wiki )2011). http://de.Guttenplag.wikia.com/wiki/GuttenPlag Wiki.
Accessed 22 July 2014
32. Martinson, B.C., Anderson, M.S., de Vries, R.: Scientists behaving badly. Nature
435, pp. 737–738, 9 June 2005. http://www.vub.ac.be/phd/doctoralschools/lsm/
docs/435737a.pdf. Accessed 20 February 2014
33. John. L.: Seven Shades of Grey. Psychological Science, April 2012
34. Ioannidis, J.P.A.: Why most published research findings are false. PLoS Med. 2(8),
0696–0701 (2005)
 Data Fraud Detection: A First General Perspective 35

35. Martin, A.: Five allegations against riken stem-cell researcher in Japan. Japan
Realtime Technology, 12 March 2014
36. Newcomb, S.: Note on the frequency of use of the different digits in natural num-
bers. Am. J. Math. 4(1), 39–40 (1881)
37. Benford, F.: The law of anomalous numbers. Proc. Am. Phil. Soc. 78, 551–572
(1938)
38. Hill, T.P.: A statistical derivation of the significant-digit law. Stat. Sci. 10, 354–363
(1995)
39. Nigrini, M.J.: Benford’s Law: Applications for Forensic Accounting, Auditing, and
Fraud Detection. Wiley, New York (2012)
40. Pinkham, R.S.: On the distribution of first significant digits. Ann. Math.Stat.
32(4), 1223–1230 (1961)
41. Weir, C., Murray, G.: Fraud in clinical trials detecting it and preventing it. Signif.
8(4), 164–168 (2011)
42. Davies, L., Gather, U.: Robust statistics. In: Gentle, J., Härdle, W., Mori, Y.
(eds.) Handbook of Computational Statistics Concepts and Methods, pp. 655–695.
Springer, Heidelberg (2004)
43. Hampel, F.: The breakdown points of the mean combined with some rejection
rules. Technometrics 27, 95–107 (1985)
44. Davies, L., Gather, U.: The identification of multiple outliers. J. Am. Stat. Assoc.
88, 782–801 (1993)
45. Lenz, H.-J., Röel, E.: Statistical quality control of data. In: Gritzmann, P., Hettich,
R., Host, R., Sachs, E. (eds.) Operations Research 1991, pp. 341–346. Springer,
Heidelberg (1991)
46. Wheeler, R., Aitken, S.: Multiple algorithms for fraud detection. Knowl. Based
Syst. 13(2–3), 93–99 (2000)
47. Windolf, P.: Korruption, betrug und ‘corporate governance’ in den USA -
anmerkungen zu enron. Leviathan 31(2), 185–218 (2003)
48. Anderson, T.W.: An Introduction to Multivariate Statistical Analysis. Wiley, New
York (1958)
49. Schmid, B.: Bilanzmodelle. Simulationsverfahren zur Verarbeitung unscharfer Teil-
informationen. ORL-Bericht No.40, ORL Institut, Universität Züich (1979)
50. Mosler, K., Lange, T., Bazovkin, P.: Computing zonoid trimmed regions in dimen-
sion d > 2. Comput. Stat. Data Anal. 53, pp. 2500–2510 (2009)
Databases and Information
Systems Integration
 ERP System Integration:
An Inter-organizational Challenge
in the Dynamic Business Environment

Tommi Kähkönen(&), Andrey Maglyas, and Kari Smolander

Software Engineering and Information Management,

Lappeenranta University of Technology,
P.O. Box 20, 53851 Lappeenranta, Finland
{Tommi.Kahkonen,Andrey.Maglyas,Kari.Smolander}@lut.fi

Abstract. The number of ERP system integration challenges has recently

increased as ERP systems need to be integrated not only with the other business
information systems inside the organization but also with the systems of busi-
ness partners. We examined the development and integration of a customized
ERP system used by a global manufacturing enterprise by interviewing 21
industrial experts. By using Grounded Theory as the data analysis approach,
four groups of factors influencing ERP system integration were revealed.
Business environment sets the constraints for integration that is governed in a
dynamic organizational landscape, hindered by political agendas. Besides the
ERP system vendor, many other ERP development network partners contribute
to ERP system integration. The complexity of system level can further distract
the integration efforts. Our findings can help managers to guide their strategic
decision making on integration issues.

Keywords: Information systems
Enterprise systems
Enterprise resource

planning
ERP
Integration
ERP development network
Stakeholders

Affecting factors
Moderating factors
Grounded theory

1 Introduction

“If you investigate IT and are searching for an easy integration between systems, there is no
such thing. Or if there is, I’m very interested in hearing more about it.”

–Enterprise architect, adopting organization

This statement from one of our interviewees clearly describes the motivation of this
study and highlights its relevance to practice. Indeed, integration remains a continuous
challenge in organizations. Decades ago, as a partly solution to integration problems
ERP systems were introduced to integrate the core business processes of an organization
[1]. An ERP implementation is difficult because of variety of reasons. It is an expensive
project and a socio-technical challenge that involves both social interactions between
many stakeholders and technical aspects in development or customization of the ERP
system [2]. After choosing the ERP system to be implemented, the project is constant

© Springer International Publishing Switzerland 2015

J. Cordeiro et al. (Eds.): ICEIS 2014, LNBIP 227, pp. 39–56, 2015.
DOI: 10.1007/978-3-319-22348-3_3
40 T. Kähkönen et al.

balancing between customization of the package and re-engineering of business pro-

cesses to fit into the package [3]. ERP systems tend to change the organizational culture
and way how people have been used to do their work [4]. ERP projects are prone to
failures and some of them even led to total catastrophes [5].
As the field of ERP systems matured, it was soon realized that with an ERP system,
a company could improve its operations and get rid of numerous legacy systems, but
the need for other business information systems interacting with the ERP system did
not vanish [6, 7]. Because of the collaborative nature of modern business, the
boundaries between systems have become fuzzier as systems cross the organizational
borders and integrate with business partners’ systems [8]. Because of the extended role
of the ERP system as the backbone enterprise business suite that connects with cus-
tomers and business partners [9], integration becomes an important consideration
during the ERP development.
However, it has been noted that ERP systems cannot be easily integrated with other
systems [10, 11]. Moreover, it has been suggested that methods for enterprise systems
integration have not been aligned with the advances on integration technologies [7].
Some studies state that integration is not well understood as a concept [12, 13]. A lit-
erature review on Information System (IS) integration research pointed out that we do
not know much about moderating factors of IS integration [12]. In our previous study,
we analyzed the existing literature on ERP system integration and concluded that
integration has not been often studied by means of systematic research approaches [14].
We attempt to fill the gap and contribute to the knowledge on ERP system inte-
gration. In this paper, our interest is to study ERP system integration from the per-
spective of ERP Development network (EDN), by which we mean the involved
organizations in the ERP system development. We used Grounded Theory [15] to
observe and understand the practice of ERP system integration in a global manufac-
turing enterprise. The following research questions were set: (1) what are the factors
that have impact on ERP system integration? And (2) how can these factors be
grouped?
The next section introduces the key concepts used in this study: ERP system
integration and ERP development network. Section 3 presents the research approach
and Sect. 4 reports the findings by describing the four groups of factors influencing
ERP system integration. The novelty of the findings is discussed in Sect. 5 before
concluding the paper.

2 Background

2.1 ERP System Integration

Integration is a general term that has various dimensions and meanings in the domain of
information systems. According to Linthicum [16], integration has technical, business
process and strategic perspectives and it includes data exchange between systems,
standardization of business processes and also cooperation and coordination between
human actors. Integration can happen inside a single organization or it can cross orga-
nization boundaries, which can be considered external integration [17]. Gulledge [13]
 ERP System Integration: An Inter-Organizational Challenge 41

clarified the concept of integration related to enterprise systems by dividing integrationto

“big I”, in which business processes are integrated by a single software application such
as ERP, and “little I”, in which enterprise systems are linked together by different
approaches, such as database-to-database and application server integration.
When examining integration from the perspective of an ERP system, it can be
concluded that integration consists of various activities. Integration of business func-
tions is the goal of an ERP implementation as the ERP system enables data flow
between business processes [8]. However, numerous other information systems, such
as Decision Support Systems (DSS) and Manufacturing Execution Systems (MES) are
still needed, and application-level integration of ERP and these systems is often nec-
essary [18], [19]. The functionality of an ERP is often enhanced by bolt-on applica-
tions, such as Customer Resource Management (CRM), and Warehouse Management
System (WMS) [20]. Because the purpose of a contemporary ERP is to provide the
backbone for business collaboration, external integration with business partners’ sys-
tems is unavoidable [21]. Another form of ERP system integration is to provide
interfaces for customers and clients to access the system through the Web or on mobile.
This type of integration is called portal-oriented application integration where an
interface is built to display the desired information needed by the intended user group
[16]. In this paper we understand ERP system integration as an activity that builds
interfaces and manages interconnections between the ERP and other internal and
external systems during the ERP system development as a collaborative effort by the
ERP development network.

2.2 ERP Development Network

Many groups of stakeholders are involved in ERP projects [22]. Besides the adopting
organization, an ERP vendor can have the key role in the project by providing support
and tools for development [23]. Consultants are often hired to ERP projects to solve
problems that occur during the implementation [24]. The ERP community has been
defined as a group consisting of an ERP vendor, consultant and implementing orga-
nization and it has been suggested that understanding the relationships and interactions
within this group would be a key milestone in the ERP research [25]. Koch [26] uses
the term “ERP network” in his work but mainly focuses on the complexity of orga-
nizational structures of ERP vendors.
However, it is often the case that this network of stakeholders involved in ERP
development is even more complex if all the involved organizations are taken into
account. The “flagship” organization, such as SAP or Microsoft can have a major role
in the ERP development when a packaged ERP is adopted. The network also includes
supply chain partners, suppliers of supporting software (databases, operating systems
and tools), as well as vendors of the existing systems that are integrated with the ERP
system. Multiple levels from the key organizations are involved, including the upper
management, business process owners, mid-level managers, the IT-department, busi-
ness representatives and end-users. Furthermore, the network is dynamic, which means
that it constantly changes its shape during the ERP development. In this regard,
42 T. Kähkönen et al.

we define the ERP Development Network (EDN) as a dynamic group of stakeholders

from different levels of all the involved organizations that are needed for ERP-related
problem solving during the ERP system development.

3 Research Approach

Qualitative research methods are essential for studies in information systems develop-
ment (ISD) and software development, because of the central role of human behavior in
them and due to the fact that they introduce, besides technological challenges, also
numerous organizational and management issues [27]. Grounded Theory (GT), origi-
nally developed by Glaser and Strauss in 1967, was chosen as the research method for
this study because ERP projects are complex and they include cooperation and col-
laboration of various stakeholders. As an inductive research method that is based on
collecting research data primarily from interviews with practitioners, GT is suitable for
approaching complex organizational phenomena [28]. ERP development is a socio-
technical endeavor making the role of network of stakeholders and human interactions
evident [2]. Respectively, ERP system integration is not purely a technological chal-
lenge but includes also collaboration and knowledge sharing among various stake-
holders [29].
Our specific focus on the integration challenges in ERP development networks
required in-depth knowledge of different stakeholders involved in the ERP project.
Therefore, we needed to approach the subject with an iterative inquiry into the EDN
and with investigation of the challenges presented from different viewpoints. Without
having a predefined theoretical model in mind, we investigated the EDN from the
viewpoint of one stakeholder to another, iteratively collecting and analyzing the data,
which GT supported well. This far GT has not been widely utilized to investigate the
integration in ERP projects. However, we deemed it especially suitable when inves-
tigating broad phenomena, such as ERP system integration, in depth.
GT is a qualitative research method that allows to develop theory iteratively based
on data that is systematically collected and analyzed [30]. Data is usually collected by
interviewing or observing one or several cases, but other sources of evidence like
written documentation or other archive material can be used as well [31]. GT is
considered to be useful for creating context-based and process-oriented descriptions of
organizational phenomena and it provides, in its Strauss and Corbin version, relatively
clear guidelines for the data analysis [15]. The benefit of GT is that it allows a
researcher to trace back to the original sources of data in order to observe how the
theory has been developed and how different instances of data have emerged into
concepts and relationships between them [30].
The data analysis in Strauss and Corbin’s version of GT consists of three coding
procedures: open, axial, and selective coding. In open coding, the transcribed data is
first labelled with codes that capture the meaning of the current piece of data. The most
important procedure in open coding is constant comparison between the pieces of data
in order to find similarities and differences. In axial coding, the connections between
categories are formed. Basically, this is the interpretation of codes, categories, and
properties developed in open coding with the goal of refining the constructs and
 ERP System Integration: An Inter-Organizational Challenge 43

making them more abstract and theoretical [31]. In selective coding, the goal is to
choose a core category and interpret its relationships to other categories and explain it
as a theory.
As data is collected and analyzed iteratively, the main question is when to stop the
process. As a theory emerges, more focus can be needed on some particular aspects of
it. At the same time, categories, dimensions, and properties become more refined as
more data collected. The situation when a researcher finds out that any new set of data
will not bring significant new codes, categories and/or relationships is called theoretical
saturation [30].

3.1 Context Description

The adopting organization (from now on referred as AO) is a large and global man-
ufacturing enterprise with an annual turnover over 9 billion euros. AO decided to build
a fully-customized ERP system for sales and logistics in order to replace several legacy
systems and also to overcome the year 2000 problem without having to make the
necessary updates to all the systems. The implementation started in the middle of 1990s
and during that time, the existing ERP packages did not have the desired functionality
to support business processes of the domain and control the complex supply chain in
AO’s specific business field. The ERP project went through major challenges,
including redesigning the insufficient system architecture and a merger of companies.
Eventually, the project greatly exceeded the intended budget. However, the system is
currently in a global use and it was widely considered as successful in the interviews.
The system is still under a constant development in 2014. The vendor of the system has
remained the same from the beginning and has a long-term relationship with AO.
Major parts of the development have been recently outsourced to Asia by the vendor to
reduce development costs. Benchmarking against ERP products in the market is
constantly being done, but for the time being, AO has decided to keep the system to
handle its core business processes.
Integration of the ERP system has been a challenging endeavor during the early
phases of the project, requiring a vast amount of resources, expertise and strict pro-
cesses, and also being the major consideration of the current development. The system
is integrated with a packaged ERP from SAP that is used for administrative processes
such as financial controlling and human resources. According to AO’s global ERP
strategy the customized ERP system is taken into use in any new facility in order to
achieve synergy benefits. This requires integrating the system with operative systems in
facilities. In order to let customers and partners to access the relevant information, a
web interface to the system has been built. Creating an infrastructure to support mobile
use to access the system with mobile devices has also been under consideration. To
ease the supply chain collaboration, e-business standardization with competitors and
business partners within the same domain has been considered. Integration with cus-
tomers’ ERPs and supply chain partners’ systems, including systems of warehouse and
transportation operators as well as customs systems has also been made.
44 T. Kähkönen et al.

3.2 Data Collection and Analysis

The data collection included two interview rounds. In the first round we conducted 17
theme-based interviews between February and May 2013. Instead of determining a
large number of fixed questions addressing specific areas of interest, the questions in
the first round were open-ended, focusing on interviewees’ experiences in the ERP
project. The more detailed questions were asked based on responses of interviewees.
For example, major challenges and successes experienced in the ERP development
were asked. This way, we were able to get a rich set of data for further investigation.
The first round started with discussions with our key contact person from the upper
management in AO. The goals of the research project were briefly presented to him in
order to identify the right persons to interview. The snowballing technique [30] in
which the next interviewee is a referral from the previous one was used for selecting the
interviewees. Rather than interviewing random persons, we navigated through the ERP
development network from one interviewee to another in order to get different view-
points to the same issues. The interviewees had different positions, ranging from upper
management to mid-level management and developers, and included people from the
AO, an ERP vendor and a company providing the middleware to the system. Due to
the long duration of the ERP system development, the roles and responsibilities of the
interviewees have been constantly changing. Some of the interviewees have been
intensively involved during the early implementation of the system whereas the others
are currently working with the system. The duration of interviews ranged from 26 to
100 min, the average being 53 min.
The second data collection round was conducted between May and June 2014. This
round focused more deeply on integration issues. In total, six experts from AO were
chosen, based on the recommendations of the contact person. The question set included
more detailed questions about integration issues, standards, technologies, organizations
and stakeholders dealing with the issues. The list of interviewees’ organizations and
roles is presented in Table 1.

Table 1. The roles and organizations of the interviwees and durations of the interviews.
Representatives of adopting organization Representatives of vendor and middleware
provider
ID Role R1 R2 ID Role R1
AO1 Business-IT negotiator 62 100 V1 Software manager 48
AO2 IT manager of business area 49 65 V2 Service owner 32
AO3 Programme manager 32 – V3 Continuous service manager 56
AO4 Enterprise architect 38 – V4 Infrastructure manager 56
AO5 Representative of sales 58 – V5 Project manager 29
AO6 IT support manager 32 – V6 Lead Software Developer 29
AO7 Representative of logistics 31 – V7 Service manager 52
AO8 Project manager 43 – MP1 Middleware manager 73
AO9 Manager of E-business and – 83 MP2 Technical consultant 73
integration
AO10 Head of E-business and integration – 60 AO = Adopting organization
AO11 Business support manager of a – 83 V = Vendor
business area MP = Middleware provider
AO12 Director of business process – 34 R1, R2 = duration of the first and the second round
development interviews in minutes
 ERP System Integration: An Inter-Organizational Challenge 45

Open and Axial Coding. After conducting the interviews, they were transcribed to
text format and analyzed by using ATLAS.ti as the coding tool. The first step in GT is
to open code the data by conceptually labelling the data based on its interpreted
meaning. We grouped the open codes into categories. A category gives the context for
the code and provides the data with more concrete meaning. For example, structural
change is ambiguous if it appears without a category, but providing the code with the

Table 2. Caterogires and their relationships.

Category Description Examples of codes Relationships to other
categories
ERP EDN consists of organizations -actor – database vendor -carries out the ED
development involved in the ERP system -AO–vendor relationship
network development -middleware provider–vendor
(EDN) relationship
-conflicting objective
-evolution
-global network
AO The company that took the ERP -governance model -is a part of the EDN
system into use -organizational culture -collaborates with other
-structural change organizations in the
-enterprise architecture EDN
-political agendas -operates in the BE
Vendor The company that implemented -involvement -is a part of the EDN
the ERP system -role in integration
-role in decision making
Middleware The company involved to -involvement -is a part of the EDN
provider redesign the system -expertise
(MP) architecture
Business Customers and other value chain -different types -are a part of the EDN
partners (BP) partners of AO -setting integration requirements
Individuals Besides the organizations, the -role in ERP development -are a part of the EDN
EDN consisted of Individuals -critical action
-viewpoint on system’s success
ERP Included the development -specifying the system -partly targets
development activities of the ERP system -testing integration
(ED) -deploying the system
-practices supporting integration
Integration Consisted of different forms of -target – internal systems -is mainly determined
integration -target – external systems by AO
-web interface -is done during the ED
-standardization -is affected by BP
-business processes
ERP system AO’s customized ERP system -scope -is the outcome of the
(ES) for sales and logisticsa -architecture ED
-flexibility -is a part of the IA
-changed scope
IT-architecture Other information systems and -packaged ERP system -is determined by AO
(IA) technologies of AO -facility systems -has connections to BP’s
-middleware systems
-integration platform
Business The business environment and -business processes support of -determines BP
environment domain where AO operated ERP products -can set constraints for
(BE) -change ED
-cost cutting of AO
ERP strategy AO had a certain strategy for its -road-mapping -is determined by AO
ERP systems -exception (local ERPs) -targets the ERP system
-synergy benefits
a
From now on referred as the system.
46 T. Kähkönen et al.

category Adopting organization clarifies the meaning. The code Adopting organiza-
tion: structural change makes it more meaningful.
In axial coding, the relationships between categories are identified and new cate-
gories may be formed based on them. Open and axial coding are not necessarily
sequential steps in the analysis process, but are often done concurrently. In total, 12
categories were created during open and axial coding. Table 2 shows the categories
created during open and axial coding, some of the codes and relationships between
categories.
There were indications of theoretical saturation in the analysis of the last inter-
views. The data did not produce new codes and already observed phenomena and
patterns repeated.
Selective Coding. In selective coding the core category is chosen and the relationships
of other categories to this category are interpreted. The whole data can be revisited and
looked from the perspective of the core category to support and explain the emerging
theory. Integration was chosen as the core category, because integration issues
emerged from the data as an important matter, and it appeared that integration was a

Fig. 1. Selective coding: creation of factors from codes and categories.

 ERP System Integration: An Inter-Organizational Challenge 47

major challenge during the project phase of the ERP system development as well as
being one of the current challenges.
By analyzing the relationships between other categories and integration, factors that
have influence on integration were created. A factor can be based on a single code or it
can be based on one or several codes in the same or different categories. For example,
the factor “Consultants” is based on the codes in EDN and Middleware provider. When
naming factors, general names were considered, for example “EDN: actor – database
provider” was renamed to “Suppliers of supporting software”. Finally, factors were
divided into four groups. Figure 1 show factors emerged from codes and categories by
displaying the most important codes that the factors are based on.

4 Factors Influencing ERP System Integration

This section describes how ERP system integration was affected by the identified
factors, divided into four groups: Organizational landscape, EDN partners, System
level and Business environment.

4.1 Organizational Landscape

Organizational landscape consists of the following factors that had impact on ERP
system integration: ERP strategy, enterprise architecture, political agendas, practices
supporting integration and nature of an integration project.
ERP Strategy. AO had an ERP strategy, which considered the two main ERP sys-
tems: the system handled the sales and logistics functionality and the packaged ERP
that was utilized for administrative processes. This consequentially meant that an
integration between these systems had to be made. New managers often questioned the
strategy of having two main ERP systems. AO aimed to deploy these two ERP systems
to any new facility worldwide in order to reach synergy benefits. However, sometimes
the local ERP systems that were used in facilities were remained. This made the IT
architecture more complex and led to further problems, such as difficult integration
scenarios. AO had a 5-year roadmap for the system, but constant changes in the
business and in the organization made it difficult to follow.
“You never reach the ideal world, you end up in having lots of [different systems] here and
there, maybe all the possible ERP vendors in some way. Then you have this company-level
roadmap and it constantly evolves.” –AO1

Enterprise Architecture. Because of the global organization and the challenging

supply chain, the Enterprise Architecture (EA) of AO appeared to be very complex.
When new features were introduced into the system, an internal architecture check was
done first to see if there was a duplicate feature in the IT-architecture and if the new
functionality could be achieved through integration. However, a single and integrated
system used by different business units with different needs caused challenges when
specifying new functionality for the system. Moreover, it was suggested that EA should
48 T. Kähkönen et al.

have more central role in the early phases when making development decisions in order
to reduce the complexity.
“Whenever any kind of change to IT is planned, then the architecture is involved. It always
looks to the future, so that it is ensured that the change is being done in as correctly as
possible” –AO4

Political Agendas. The organizational structure of the AO was constantly changing

during the development of the system. For example, a big merger of two companies
took place when the system was not yet taken into use. This changed the governance
model and power structures inside AO: different functional areas became under a
changed leadership. This led to decisions to take some of the functionality away from
the ERP system, to be implemented in other systems, which required additional inte-
grations with the ERP system. As a result, the original scope of the system changed.
This was mentioned as one of the major challenges in the current state of the system
because of the increased the costs. It was also pointed out that the increased complexity
of systems introduced delays when querying order status information from the supply
chain. Furthermore, the major structural change also prolonged the integration projects
between AO and customers – it took years until these projects were carried out again
under a new management that favored e-business.
“[Logistics] started making separate islands, they wanted to “freeze” the system to a certain
point and started to include all kinds of additional systems there. It has been ongoing for ten
years now and we have ended up to serious problems and the costs have increased in that area.
[Consultants] have evaluated the systems and made this great finding that it’s a spaghetti and a
new transportation management system needs to be built there.” –AO1

Practices Supporting Integration. The system replaced several legacy systems. The
parallel run with the legacy systems could take from several months to one year of
time. Appropriate practices to manage the master data while running the system in
parallel with the legacy systems needed to be in place. When deploying the system to
facilities, supporting practices for roll-out approaches and testing of the facility inte-
gration were developed through trial and error. The first deployment of the system
failed due to performance issues. It appeared that the pilot system was tested in an
environment that did not match the real environment and the scope of the system was
not fully realized in the beginning of the development. The importance of testing of the
facility integration was emphasized as a critical factor in deployments.

“The more successful the testing sessions are between the facility system and the ERP, the
better everything will start off. In that sense the testing of the facility integration is absolutely
the key”–AO5

Nature of an Integration Project. Different integration projects were constantly

being done while developing the system further. Some of these projects were done by a
dedicated integration team of AO, without the need of the vendor to be involved.
Moreover, if a similar integration project had already been done before, the next one
was carried out more rapidly. Sometimes the vendor had to be involved to make
changes to the system, because of the integration requirements. The duration of
 ERP System Integration: An Inter-Organizational Challenge 49

integration projects ranged from several weeks to years, depending on the nature of the
project. However, sometimes a separate integration project was not established. In
these projects, the need for integration was realized in later stages and that caused
problems when allocating resources for carrying out the integration and testing. The
integration requirements, especially the need of integration testing, were sometimes
overestimated in these projects.

“Now and then you underestimate things like the demands of system integration.” –AO5

4.2 EDN Partners

The organizations in EDN (others than AO) influenced integration with varying
emphasis. In the context of this implementation, the following organizations had an
impact on ERP system integration: consultants, ERP vendor, business partners,
standardization partners and suppliers of supporting software.
Consultants. Because of the serious performance and scalability issues encountered,
the middleware provider was involved in the beginning of the ERP system imple-
mentation to redesign the system architecture. The original 2-tier architecture was
replaced with a middleware solution based on transaction processing monitors. This
made the system architecture more scalable for a broader user base and enabled the
integration of business functions. The relationship between the middleware provider
and the vendor appeared to be crucial when redesigning the system architecture. The
cooperation with the vendor was considered as challenging in the beginning, but after
the initial conflicts, an improved system architecture was developed.

“Practically, [the vendor] didn’t have a clue of how to make it work […] and when we looked
at it, it seemed that the way of implementing the system and the use of object model was
completely wrong.” –MP2

ERP Vendor. The vendor had the key role in ERP system integration. Because of the
long-term relationship with AO, the vendor built many of the systems used in facilities
and had the required knowledge on these systems when integrating systems. Further-
more, the vendor’s knowledge on AO’s business was proven to be a major facilitator in
cooperation. Even though both AO and vendor had a rather positive viewpoint on their
relationship, neither of the partners had always been satisfied with this relationship. AO
had even considered of buying the source code of the system from the vendor, but
according to one interviewee, “it did not turn out to be a realistic option”. The other
interviewee considered the relationship as a “forced marriage”. Deploying the system
to facilities was carried out in a close cooperation between AO and vendor.

“We are in a close cooperation daily, we are making things together. [The relationship] is not
a traditional customer-vendor.” –V6.
“We’ve had the benefit of very skilled representatives from the vendor side, with a long history
with [the system] and system integration. This is worth its weight in gold, and more.” – AO5
50 T. Kähkönen et al.

Business Partners. Because of the complexity of the supply chain, many third party
companies were involved, including freight forwarders, harbor operators and ware-
houses. The system had been integrated with the systems of these parties. This was
occasionally considered challenging. It was pointed out that a sudden need to integrate
an external system with the ERP system can occur. Besides the supply chain partners,
AO worked with customers more intensively. In the early phases of the project, a web
interface was built to allow customers to view the certain information of the system.
Customer integration projects with the most important customers were ongoing. These
projects were often initiated by customers that wanted to improve their own business
processes or made an update to their ERP system. This then called for appropriate
actions by AO that tried to fulfill the integration needs of its customers and other value
chain partners.

”And later came – it was not originally specified as a requirement of the system – this
transportation cost management system came there.” –AO7

Standardization Partners. Standardization partners appeared to be another EDN

group that had an impact on ERP system integration. AO has participated in e-business
standardization efforts within the domain in order to ease the collaboration with
business partners. Through these standards, messaging between the system and external
systems was made easier.

“[It] is a separate messaging standard which has been built for [our] industry. [Our company]
is one of the companies developing that. There are also all our biggest competitors involved in
that work.” –AO9

Suppliers of Supporting Software. The representative of middleware provider esti-

mated that vendor’s choice for the base technologies of the system was slightly affected
by database vendors and their technologies the vendor was initially familiar with. AO
relied on the supplier’s expertise in this matter and the project ended up in difficulties
because of the non-scalable system architecture designed by the vendor. Consultants
from the middleware provider company small in size, were not able to convince AO to
choose their technology until later when architectural problems occurred.

“You should never believe in the sales speeches of salesmen, the organization should have
sangfroid to test the options.” –MP1

4.3 System Level

On system level the amount of ERP system customization, ERP system architecture,
integrative systems’ characteristics and complexity of the IT architecture influenced
ERP system integration.
Amount of ERP System Customization. By having a fully customized ERP, AO
could have a total control over the system and its integration capabilities. AO was not
affected, for example, by the version updates made by the ERP flagship organization,
 ERP System Integration: An Inter-Organizational Challenge 51

such as SAP. In addition, the control of development, being free of licensing costs and
the advanced functionality provided through customization were highlighted as bene-
fits. The vendor saw that in some situations, the new application logic could be directly
built into the system instead of adding additional application layers. However, devel-
opment of a customized system introduced some of the specific challenges, such as
performance issues due to the non-scalable system architecture. Moreover, the devel-
opment was expensive and there were no other parties driving the development as it is
often the case with packaged ERP systems.

“[AO] couldn’t have had a better system what they got when they made a glove to a hand […] I
have never seen such advanced functionality anywhere, you can just drag a shipping container
and drop it to a ship” –V4

ERP System Architecture. The system had to be flexible enough to enable integra-
tion of various different systems. Moreover, it had to enable the rearrangements as
some of the logistics functionality previously provided by the system was replaced by
external integrations. However, it appeared that replacing certain functionality in the
system was not always straightforward, because of the system architecture. More
modular architectural design that would better enable these modifications, was
emphasized. Also, the system messaging between internal systems was not standard-
ized, but instead, messages had a proprietary format made by the vendor. This has led
to a situation, in which more intermediary was needed and also required the vendor to
be involved in integration.

“[The system] has enabled many things that we have been doing over the years to increase our
competitiveness and supported the organizational changes. We have been able to rearrange the
services by fluently combining different machine lines and production pipelines according to
how we want to arrange our business.” –AO6

Characteristics of Integrative Systems. Because of the heterogeneity of the facility

systems, such as some of them even programmed in Cobol in 1970s, the system
integration approach was different in each location due to varying functionality of the
facility system in question. The system had to be modified based on the characteristics
of the facility system in question. Different approaches for roll-outs were established to
deploy the system.

“In many cases it will require big changes to [the system], depending on the facility. Some
facilities don’t require many changes […] The ease of the roll-out may vary greatly between
facilities.” –S2

Complexity of the IT Architecture. The IT architecture of AO was a complex one.

Besides the two main ERP systems, it consisted of tens of facility systems, local ERP
systems, integration platforms and many other internal and external systems. Besides
the AO’s attempt to reduce the complexity, some islands of technology were formed.
The complexity of the IT architecture introduced difficult integration scenarios. For
example, sending an invoice from one office to another involved the system, the global
SAP ERP and a local SAP in the target location, and exchanging the information
52 T. Kähkönen et al.

between these systems was problematic. The project that was started four years ago for
internal invoicing was still ongoing.

“So we have [the system], then [one business area] has huge amount of different systems […]
We have a global SAP for finance purposes. Then we have a local SAP in some of our sites. And
we’re trying to get all these different systems to work together. It’s a total headache. It’s a
mission impossible, I would say.” –AO9

4.4 Business Environment

The business environment had an effect on the other three groups, thus it had an
indirect impact on integration. By defining the domain where AO operated, it also
determined the EDN partners to collaborate with. Economic situation and business
process support of ERP products were identified as factors that had impact on other
groups but eventually they had an impact on integration.
Economic Situation. Because of changing business conditions, AO constantly had to
cut down the development costs. It was suggested that business processes of AO could
be developed further, for example through mobile applications that would provide
interfaces to the users and business partners to access the system. However, economic
situation postponed the development of these features.

“[Mobile interfaces] have a quite big price tag, so that when you have to save in costs, they are
probably the first thing to drop out.” –V6

Business Process Support of ERP Products. During the time when AO made the
decisions about the system, business process support of available ERP products in
the market was not comprehensive. This was seen as the major driver that led to the
decision to make a fully customized ERP system. The business process support of ERP
products affected the amount of customization of the system which in turn affected ERP
system integration.

“[Logistics] is quite complicated and SAP didn’t manage this back then. I still feel that it
wouldn’t.” –AO10

4.5 Summary of Results

Figure 2 displays the four groups of factors that have influence on ERP system inte-
gration. Based on the ERP strategy of AO, the system was integrated with the packaged
ERP system and was to be taken into use in any new facility in order to reach synergy
benefits. When planning new features to the system, enterprise architecture was eval-
uated to investigate if the new functionality could be achieved through integration.
Political agendas taking place in AO led to more fragmented IT architecture and
additional integrations. Practices supporting integration were developed in order to
 ERP System Integration: An Inter-Organizational Challenge 53

Fig. 2. Factors influencing ERP system integration.

enable smooth deployments to facilities. The nature of an integration project deter-

mined how easily the integration was carried out. Sometimes a separate project was not
established because the integration requirements were not realized.
ERP system integration was affected by five EDN partners. Suppliers of supporting
software slightly affected the decisions made on integration technologies. Consultants
were needed to re-design the system architecture. AO’s close relationship with the
vendor as well as vendor’s expertise have aided especially the internal integration.
Business partners introduced additional systems to be integrated with and some of the
integration requirements emerged directly from them. Standardization partners were
involved when further enhancing the supply chain collaboration.
By having a fully customized system, some of the integration challenges that would
emerge with a packaged ERP system could be bypassed. Sometimes integration
introduced changes to the ERP system architecture, which had to enable the rear-
rangements. When deploying the system to a facility, the characteristics of the facility
systems determined how easily the integration was carried out. The complexity of the
IT architecture introduced challenging integration scenarios when trying to transfer the
data between several different ERP systems.
Business environment had an indirect impact on integration by defining the domain
of AO and the business partners and standardization partners to cooperate with. It also
had an impact on the economic situation. Because of the cost cutting of AO, business
process improvement through a mobile infrastructure was postponed. Business envi-
ronment also defined the business process support of ERP products, which was the
main reason for AO to build a customized ERP system.

5 Discussion

The analysis of the EDN of a global manufacturing enterprise highlights the

socio-technical nature of the ERP system integration and the role and impact of dif-
ferent organizations involved in it. The role of EDN has not been often emphasized in
54 T. Kähkönen et al.

the studies on ERP system integration. This study reveals that EDN should be taken
into account because the ERP system cannot be isolated from its environment and
business partners. Considering EDN is especially important when managing integration
of the ERP system with other information systems.
The previous studies on affecting factors on ERP implementation and studies on
ERP success factors are partially overlap with our findings and confirm them. For
instance, a socio-technical model for ERP implementation has been proposed in [32].
In this model, ERP implementation process is affected by the external environment and
the organization itself. This model does not discuss about EDN, which may be
explained by the early publication time of the study. Also, our further analysis revealed
that Business environment has no direct effect to integration. Instead, this factor rep-
resents a meta-factor affecting the other factors. This also highlights that ERP inte-
gration challenges do not happen in isolation and therefore should be studied in the
context of EDN.
Although ERP success factors have been studied comprehensively, e.g. in [11, 33],
they are often organizational and lack the EDN perspective. The relationship between
the client (AO) and the vendor has been identified as a success factor in ERP imple-
mentation [33]. Our findings suggest its importance also in ERP system integration. In
addition, our findings suggest that the vendor-consultant relationship can have a key role
when solving integration issues. Moreover, it has been noted that enterprise application
integration (EAI) success factors are similar to general ERP success factors – successful
EAI needs to consider the factors on the levels of business, organization, technology,
project as well as environmental factors [12, 34]. Hoverer, this classification does not
either address the role of EDN. Besides the role of the vendor, our evidence suggests the
important roles of consultants and business partners in ERP system integration. It means
that further studies on ERP system integration should not be limited to understanding the
relationships between the AO and the vendor only, because this perspective lacks the
input from other partners involved in resolving integration issues.
As in all qualitative studies, the findings of this study cannot be easily generalized.
The findings are related only to the specific case and the generalization of the findings
of the study is theoretical [35], i.e. they generalize specific observations to theoretical
concepts. With these concepts we can explain the events in the studied organization and
we also strongly believe that these factors influencing ERP system integration are
similar in other contexts as well. In future, we will study different EDNs and apply the
results of this study in different contexts, e.g. in cases were a packaged ERP is adopted.
Moreover, investigating how integration is governed in EDNs, especially how decision
making on integration evolves in them and what is the role of enterprise architecture in
integration governance are the topics of our future interests.

6 Conclusions

We have proposed empirically grounded factors that had an effect on ERP system
integration in a large manufacturing enterprise. The factors are divided into four
groups: Business environment, Organizational landscape, ERP development network
partners, and System level. The findings reflect the nature of ERP system integration as
 ERP System Integration: An Inter-Organizational Challenge 55

socio-technical endeavor that involves interaction and collaboration between stake-

holders. This perspective is often ignored by ERP integration studies that often focus
on the adopting organization only. Through our findings, managers can better under-
stand the nature of integration in the context of ERP development networks and guide
their strategic decisions on integration issues.

Acknowledgements. This study was funded by Academy of Finland grant #259454.

References
1. Beheshti, H.M.: What managers should know about ERP/ERP II. Manage. Res. News 29(4),
184–193 (2006)
2. de Albuquerque, J., Simon, E.: Dealing with socio-technical complexity: towards a
transdisciplinary approach to IS research. In: Proceedings of the European Conference on
Information Systems (ECIS), 2007, pp. 1458–1468 (2007)
3. Law, C.C.H., Chen, C.C., Wu, B.J.P.: Managing the full ERP life-cycle: considerations of
maintenance and support requirements and IT governance practice as integral elements of
the formula for successful ERP adoption. Comput. Ind. 61(3), 297–308 (2010)
4. Liang, H., Xue, Y.: Coping with ERP-related contextual issues in SMEs: a vendor’s
perspective. J. Strateg. Inf. Syst. 13(4), 399–415 (2004)
5. IDG Consumer & SMB: 10 Biggest ERP Software Failures of 2011. http://www.pcworld.
com/article/246647/10_biggest_erp_software_failures_of_2011.html
6. Lehmann, H., Gallupe, B.: Information systems for multinational enterprises - some factors
at work in their design and implementation. J. Int. Manage. 11(2), 163–186 (2005)
7. Xu, L.D.: Enterprise systems: state-of-the-art and future trends. IEEE Trans. Ind. Inform. 7
(4), 630–640 (2011)
8. Hsu, P.-F.: Integrating ERP and e-business: resource complementarity in business value
creation. Decis. Support Syst. 56, 334–347 (2013)
9. Hvolby, H.-H., Trienekens, J.H.: Challenges in business systems integration. Comput. Ind.
61(9), 808–812 (2010)
10. Doedt, M., Steffen, B.: Requirement-driven evaluation of remote ERP-system solutions: a
service-oriented perspective. In: 34th IEEE Software Engineering Workshop (SEW), 2011,
pp. 57–66 (2011)
11. Momoh, A., Roy, R., Shehab, E.: Challenges in enterprise resource planning
implementation: state-of-the-art. Bus. Process Manage. J. 16(4), 537–565 (2010)
12. Chowanetz, M., Legner, C., Thiesse, F.: Integration: an omitted variable in information
systems research. In: European Conference on Information Systems (ECIS) 2012
Proceedings (2012)
13. Gulledge, T.: What is integration? Ind. Manag. Data Syst. 106(1), 5–20 (2006)
14. Kähkönen T., Smolander, K.: ERP integration - a systematic mapping study. In: 15th
International Conference on Enterprise Information Systems (ICEIS) 2013 Proceedings
(2013)
15. Corbin, J., Strauss, A.: Grounded theory research: procedures, canons, and evaluative
criteria. Qual. Sociol. 13(1), 3–21 (1990)
16. Linthicum, D.S.: Next generation application integration: from simple information to Web
services. Addison-Wesley, Boston (2004)
56 T. Kähkönen et al.

17. Barki, H., Pinsonneault, A.: A model of organizational integration, implementation effort,
and performance. Organ. Sci. 16(2), 165–179 (2005)
18. Shafiei, F., Sundaram, D., Piramuthu, S.: Multi-enterprise collaborative decision support
system. Expert Syst. Appl. 39(9), 7637–7651 (2012)
19. Tao, Y.-H., Hong, T.-P., Sun, S.-I.: An XML implementation process model for enterprise
applications. Comput. Ind. 55(2), 181–196 (2004)
20. Watts, C.A., Mabert, V.A., Hartman, N.: Supply chain bolt-ons: investment and usage by
manufacturers. Int. J. Oper. Prod. Manag. 28(12), 1219–1243 (2008)
21. Møller, C.: ERP II: a conceptual framework for next-generation enterprise systems?
J. Enterp. Inf. Manag. 18(4), 483–497 (2005)
22. Skok, W., Legge, M.: Evaluating enterprise resource planning (ERP) systems using an
interpretive approach. Knowl. Process Manage. 9(2), 72–82 (2002)
23. Somers, T.M., Nelson, K.G.: A taxonomy of players and activities across the ERP project
life cycle. Inf. Manage. 41(3), 257–278 (2004)
24. Metrejean, E., Stocks, M.H.: The role of consultants in the implementation of enterprise
resource planning systems. Acad. Inf. Manag. Sci. J. 14(1), 1–25 (2011)
25. Sammon, D., Adam, F.: Decision Making in the ERP Community. In: European Conference
on Information Systems (ECIS) 2002 Proceedings (2002)
26. Koch, C.: ERP - a moving target. Int. J. Bus. Inf. Syst. 2(4), 426–443 (2007)
27. Seaman, C.B.: Qualitative methods in empirical studies of software engineering. IEEE
Trans. Softw. Eng. 25(4), 557–572 (1999)
28. Charmaz, K.: Constructing Grounded Theory. SAGE Publications, London (2006)
29. Welker, G.A., van der Vaart, T., van Donk, D.P.: The influence of business conditions on
supply chain information-sharing mechanisms: a study among supply chain links of SMEs.
Int. J. Prod. Econ. 113(2), 706–720 (2008)
30. Strauss, A., Corbin, J.: Basics of Qualitative Research: Techniques and Procedures for
Developing Grounded Theory, 3rd edn. SAGE Publications, Thousand Oaks (2008)
31. Urquhart, C., Lehmann, H., Myers, M.D.: Putting the ‘theory’ back into grounded theory:
guidelines for grounded theory studies in information systems. Inf. Syst. J. 20(4), 357–381
(2010)
32. Somers, T., Nelson, K., Ragowsky, A.: Enterprise Resource Planning (ERP) for the next
millennium: development of an integrative framework and implications for research. In:
AMCIS Proceedings (2000)
33. Ngai, E.W.T., Law, C.C.H., Wat, F.K.T.: Examining the critical success factors in the
adoption of enterprise resource planning. Comput. Ind. 59(6), 548–564 (2008)
34. Lam, W.: Investigating success factors in enterprise application integration: a case-driven
analysis. Eur. J. Inf. Syst. 14(2), 175–187 (2005)
35. Lee, A.S., Baskerville, R.L.: Generalizing generalizability in information systems research.
Inf. Syst. Res. 14(3), 221–243 (2003)
Using Complex Correspondences for Integrating
Relational Data Sources

Valéria Pequeno1(B) , Helena Galhardas2 , and Vânia M. Ponte Vidal3

1
INESC-ID, Taguspark, Oeiras, Portugal
[email protected]
2
Instituto Superior Técnico, Universidade de Lisboa and INESC-ID,
Taguspark, Oeiras, Portugal
[email protected]
3
Universidade Federal do Ceará, Fortaleza, Brazil
[email protected]

Abstract. Data Integration (DI) is the problem of combining a set of

heterogeneous, autonomous data sources and providing the user with a
unified view of these data. Integrating data raises several challenges, since
the designer usually encounters incompatible data models characterized
by differences in structure and semantics. One of the hardest challenges is
to define correspondences between schema elements (e.g., attributes) to
determine how they relate to each other. Since most business data is cur-
rently stored in relational databases, here present a declarative and for-
mal approach to specify 1-to-1, 1-m, and m-to-n correspondences between
relational schema components. Differently from usual approaches, our
(CAs) have semantics and can deal with outer-joins and data-metadata
relationships. Finally, we demonstrate how to use the CAs to generate
mapping expressions in the form of SQL queries, and we present some
preliminary tests to verify the performance of the generated queries.

Keywords: Schema matching · Correspondence assertions · Data

integration · Relational model

1 Introduction
A DI system aims at integrating a variety of data obtained from different data
sources, usually autonomous and heterogeneous, and providing a unified view of
these data, often using an integrated schema. The integrated schema makes a
bridge between the data sources and the applications that access the DI system.
Data in a DI system can be physically reconciled in a repository (materialized
data integration approach), or can remain at data sources and is only consoli-
dated when a query is posed to the DI system (virtual data integration approach).
A data warehouse system [1] is a typical example of the first approach. As exam-
ples of the second approach, we can cite federated information systems [2] and
mediator systems [3]. In the present work, both scenarios can be used, but in
this paper we will focus on the materialized integration approach.
c Springer International Publishing Switzerland 2015
J. Cordeiro et al. (Eds.): ICEIS 2014, LNBIP 227, pp. 57–74, 2015.
DOI: 10.1007/978-3-319-22348-3 4
58 V. Pequeno et al.

One of the hardest problems to solve in DI is to define mappings between

the integrated schema (the target) and each data source schema, known as the
schema mapping problem. It consists of two main tasks: i) schema matching to
define/generate correspondences (a.k.a. matches) between schema elements (e.g.,
attributes, relation, XML tags, etc.); and ii) schema mapping to find data trans-
formations that, given data instances of a source schema, obtain data instances
of the target schema.
The result of schema matching is a set of correspondences that relate elements
of a source schema to elements of the target schema, where an element can be
a relation name or attribute in the relational model. These correspondences can
be described using a Local-as-view (LAV), a Global-as-view (GAV), or a Global
and Local-asview (GLAV) language. In summary, in a LAV approach, each data
source is described as a view over the integrated schema. In a GAV approach,
the integrated schema is expressed as a view over the data sources. Finally, the
GLAV combines the expressive power of both GAV and LAV. Once the schema
matching is performed, the correspondences are used to generate the schema
mappings. For example, a schema mapping can be codified through an SQL
query that transforms data from the source into data that can be stored in the
target.
Extensive research on schema matching has been carried out in recent
years [4,5]. The majority of the works on this subject identifies 1-1 correspon-
dences between elements of two schemas. For example, a 1-1 correspondence
can specify that element title in one schema matches element film in another
schema, or that relation genre matches relation category1 . This kind of
schema matching is known in the literature as basic matching. Good surveys
can be found in [6,7].
While basic matching is common, it leaves out numerous correspondences of
practical interest, in particular when we consider DI systems. Thus, more com-
plex matches are necessary. A complex matching specifies 1:n, m:n, or n:1 corre-
spondences between elements of two schemas. For example, it may specify that
totalPrice corresponds to unitPrice * quantity; or that name matches con-
catenate(firstName, lastName), where concatenate is a function that applies
to two strings and returns a concatenated string; or even that the average
departmental salary avgWage corresponds to grouping the salaries (salary)
of all employees (emp) by department (dept). Works in [8,9] are examples of
approaches that deals with complex matches.
Some researchers go beyond dealing with complex matches and add semantics
to the correspondences to improve the overall matching quality. In the Sect. 2,
we explain more about complex matches and show a motivation example. The
remainder of the paper is structured as follows. In Sect. 3, we present the neces-
sary background in Correspondence Assertions (CAs), the formalism used in this
work to specify correspondences between elements of schemas. In Sect. 4, we pro-
pose new CAs to deal with join operators and metadata. Section 5 shows how to

1
We use bold to represent attribute names and uppercase to represent relation
names.
 Using Complex Correspondences for Integrating Relational Data Sources 59

generate mapping expressions from CAs. Section 6 shows some preliminary tests
to evaluate our approach. Section 7 describes the related work. Finally, Sect. 8
concludes and describes future work.

2 Motivating Example
Consider a motivating example with the source schemas S1 and S2 in Fig. 1,
which contain information about movies. S1 keeps a catalog of movies with
information about different types of media (dvd, blue rays, etc.) in which the
movies are available. The names of the relations and attributes are mostly self-
explanatory. Some non-self-explanatory attributes in S1 have the following mean-
ing: id is the movie identifier, year is the year of a movie, film is the title of
a movie, number is the tape identifier, name can be a producer or a director
name, and role can be producer or director. FK1 and FK2 are foreign keys.
We use the notation FK(R:L, S:K ) to denote a foreign key, named FK, where
R and S are relation names and L and K are list of attributes from R and S,
respectively, with the same length. FK1 is the foreign key of tape that refers
to movie and FK2 is the foreign key of movieMakers that refers to movie.
S2 stores general information about movies and the places (in different cities)
where movies are being shown. We assume that S1 can store older movies than
S2 . Some non-self-explanatory attributes in S2 have the following meaning: rate
is the classification of the movie with regard the audience, location and city
are, respectively, the cinema and the name of the city where the movie is shown,
and time is the date when the movie is shown.

Fig. 1. Example of source schemas and a integrated schema.

The integrated schema M, also shown in Fig. 1, provides a unified user view
of movies currently shown in cinemas of Lisbon. It is populated by information
from schemas S1 and S2 . The relation movie stores movies shown currently at
a cinema. The relation filmMakers keeps information about professionals of
show businesses. The relation schedule contains information about the schedule
of movies shown in Lisbon. The relation remakes keeps the years of movies for
60 V. Pequeno et al.

which there is at least one remake. The relation rating stores the classification of
movies with regard to suitability audience. Some non-self-explanatory attributes
in M have the following meaning: description is the summary of a movie,
nvYear is the year of the most recent version of a movie, ovYear is the year
of the older versions of a movie, and quantity is the total of movies with the
same rating.
Given the schemas S1 , S2 , and M, we can consider the correspondences
between the source schemas S1 and S2 , and the target schema M. As an exam-
ple, we can state that M.schedule corresponds to S2 .showTime, because both
relations store information regarding the same real world concept2 . However,
in this correspondence, it is not clear that M.schedule only keeps schedules
about movies shown in Lisbon. The additional information: M.schedule corre-
sponds to S2 .showTime when S2 .showTime.city = “Lisbon”, specifies better
the matching.
The works reported in [10–12] and [5](chap. 3) propose schema matching
approaches that can specify correspondences to deal with situations as required
in the example The reader can see more proposals to add semantics to schema
matching in [8,10,12,13]. However, the following situations have not been fully
covered yet:
1. Correspondences Between Relations Involving Join Conditions Other than
Equality of Attributes. Consider the relation M.remakes that keeps a list of
remakes with the years of the oldest versions. Knowing that S2 .film keeps
current movies and S1 .movie may contain older versions of the same movie,
we want to indicate which of the current movies are remakes and store this
information in M.remakes. The correspondence between these relations can
be specified as: M.remakes corresponds to S2 .film join S1 .movie where
S2 .film.title = S1 .movie.film and S2 .film.year > S1 .movie.year. Usual
schema matching approaches cannot specify this correspondence, because join
conditions are not explicitly defined in schema matching. Moreover, join paths
are normally automatically discovered in the schema mapping phase [14],
and the algorithms used can only find equi-join conditions, so they cannot
automatically discover the condition S2 .film.year > S1 .movie.year. Hence,
we need a schema matching approach that makes it possible to specify the
join between relations and allows general join conditions containing operators
different from equality.
2. Correspondences Between Relations Involving Outer-joins (Full, Left, or
Right). We want to indicate how M.movie is related to source schemas S1
and S2 . M.movie and S2 .film represent the same concept of the real world
(i.e., both relations store current movies shown at some cinema). However, it
is not enough to specify that M.movie matches S2 .film, because there are
attributes in S1 .movie (namely, category and summary) that contain infor-
mation required in the schema of M.movie. Hence, we should specify that
2
We use a path representation: an attribute A of a given relation R in a given database
schema D is referred to as D.R.A. For simplicity, we omit the database schema when
the context is clear.
 Using Complex Correspondences for Integrating Relational Data Sources 61

M.movie is related to both S1 .movie and S2 .film. However, it is not correct

we simply match M.movie to S1 .movie because S1 .movie can store movies
that are not being shown in a cinema anymore and M.movie can store recent
movies that are not available in dvds yet. In summary, we should specify that:
M.movie corresponds to S2 .film left outer-join S1 .movie on S2 .film.title =
S1 .movie.film and S2 .film.year = S1 .movie.year. Note that the condition
S2 .film.title = S1 .movie.film and S2 .film.year = S1 .movie.year guaran-
tees that we refer to a same movie stored in both S1 .movie and S2 .film.
Again, we cannot specify this type of correspondence since joins (and their
variants) are not explicitly defined in current schema matching approaches.
3. Correspondences Between Data and Metadata. Consider the relations
S1 .movieMakers and M.filmMakers. Both keep information about the
relationship between a movie, a producer, and a director. We want to
indicate that M.filmMakers corresponds to S1 .movieMakers since they
represent the same concept in the real world. In addition, we want to spec-
ify the correspondences between the attributes of these relations. Know-
ing that S1 .movieMakers.name can be a producer name or a director
name, we would like to specify that M.filmMakers.producer corresponds
to S1 .movieMakers.name when S1 .movieMakers.role = “producer”
and that M.filmMakers.director corresponds to S1 .movieMakers.name
when S1 .movieMakers.role = “director”. However, we cannot spec-
ify these correspondences using traditional schema matching approaches,
because these correspondences involve semantics not covered yet by these
approaches. Actually, we can only specify that M.filmMakers.producer
matches to S1 .movieMakers.name and M.filmMakers.director matches
to S1 .movieMakers.name.

In order to deal with these situations, we propose to use a formalism based on

CAs [10,15]. Using CAs, we can declaratively specify basic and complex match-
ings with semantics. We propose to adapt CAs to be able to express schema
matching between relational schemas, as well as to extend this formalism with
new types of CAs to deal with joins, outer-joins, and data-metadata relation-
ships. Finally, we demonstrate how mapping expressions in the form of SQL
queries can be generated from CAs.

3 Background
In this section, we present the basic terminology used in this paper. We also
review the different classes of CAs, and adapt them to the Relational Data
Model (RDM).

3.1 Basic Concept and Notation

We assume that the reader is familiar with the relational concepts. We denote
a relation schema as R(A1 , A2 , . . . , An ), and a foreign key as FK(R:L, S:K ).
We say that FK relates R and S.
62 V. Pequeno et al.

A relational schema is a pair S= (R, Ω), where R is a set of relation schemas

and Ω is a set of relational constraints such that: (i) Ω has a unique primary key
for each relation schema in R; (ii) if Ω has a foreign key of the form FK(R:L,
S: K ), then Ω also has a constraint indicating that K is the primary key of S.
Given a relation schema R(A1 , A2 , . . . , An ) and a tuple variable t over R, we
use t[Ai ] to denote the projection of t over Ai .
Let S= (R, Ω) be a relational schema and R and T be relation names of
relation schemas in R. We denote  = FK1 • FK2 • · · · • FKn−1 a path from R
to T iff there is a list R1 , . . . , Rn of relation schemas in S such that R1 = R,
Rn = T, and FKi relates Ri and Ri+1 . We say that tuples of R reference tuples
of T through .

3.2 Correspondence Assertions

We use Correspondence Assertions (CAs) in order to express schema matchings
between schema elements. CAs are formal expressions of the general form ψ:
T ← S, where ψ is the name of the CA, T is an expression formed by elements
of the target schema, and S is an expression formed by elements of a source
schema. The symbol “←” means “is matched from”.
In accordance to [10], there are four types of CAs: Relation Correspondence
Assertion (RCA), Attribute Correspondence Assertion (ACA), Summation Cor-
respondence Assertion (SCA), and Grouping Correspondence Assertion (GCA).
RCAs and SCAs specify the relationship between relations of distinct schemas,
while ACAs and GCAs specify the relationship between attributes of relations
of distinct schemas. We now shortly describe each type of CA, adapting them to
the RDM. In the remainder of this Section, consider: Si = (Ri , Ωi ) be relational
schemas for 1 ≤ i ≤ n, with Ri being relation names of relation schemas in Ri .
Definition 1. Let σ be a selection over R2 . A Relation Correspondence Asser-
tion RCA is an expression of one of the following forms:

n
1. ψ: S1 [R1 ] ← S2 [R2 ] 4. ψ: S1 [R1 ] ← i=1 Si [Ri ]
2. ψ: S1 [R1 ] ← S2 [R2 σ] n
3. ψ: S1 [R1 ] ← S2 [R2 ] − S3 [R3 ] 5. ψ: S1 [R1 ] ← i=1 Si [Ri ] 

In Definition 1, we say that ψ matches R1 and Ri , for 1 ≤ i ≤ n. RCAs

express the various kinds of semantic equivalent relationships. Two relations R1
and R2 are semantically equivalent if they represent the same real concept and
there is a 1-to-1 correspondence between their instances. ψ1 , shown in Fig. 2, is
an example of a RCA.
ψ1 specifies that M.schedule is semantically equivalent to S2 .showTime
when the condition S2 .showTime.city = “Lisbon” is satisfied. This means that
only a subset of tuples of S2 .showTime, those that satisfy the condition are
involved in the match.
Before we define an Attribute Correspondence Assertion (ACA), we need
introduce the concept of attribute expression, as follows:
 Using Complex Correspondences for Integrating Relational Data Sources 63

Fig. 2. Examples of correspondence assertions.

Definition 2. Let R2 and T be relation names in R2 , with A being an attribute

of R2 and B an attribute of T. Let also  be a path from R2 to T. An attribute
expression E over R2 is an expression with one of the following forms:

1. S2 [R2 ] • A 2. S2 [R2 ] • /B. 

Definition 3. Let Ai be attributes of R1 (for 1 ≤ i ≤ n). Let also E j , for

1 ≤ j ≤ m, be attribute expressions over R2 . An Attribute Correspondence
Assertion (ACA) is an expression of one of the following forms:

1. ψ: S1 [R1 ] • A1 ← E1 3. ψ: S1[R1 ] •A1 ← (E1 , p1 ); . . . ; (Em ; pm );v

2. ψ: S1 [R1 ] • A1 ← ϕ(E1 , E2 , . . . , Em ) 4. ψ: S1[R1 ](A1 , . . . , An ) ← (E1 , . . . ,En )

Where ϕ is a function over attributes of R2 , pj (for 1 ≤ j ≤ m) are boolean

conditions over attributes of R2 , and v is a value. We say that ψ matches R1
and R2 . 

ACAs specify the relationship between the attributes of relations that are
matched by a RCA. They allow to define 1:1, 1:n, n:1, or m:n relationships
between attributes of relations of different schemas. For example see the ACA ψ2
presented in Fig. 2. It specifies the correspondence between M.schedule.movie
and S2 .film.title through a path from showTime to film.

Definition 4. Let σ be a selection over R2 . Let also A’i attributes of R2 (for

1 ≤ i ≤ m). A Summation Correspondence Assertion (SCA) is an expression of
one of the following forms:

1. ψ: S1 [R1 ] ⇐ groupby(S2 [R2 ](A1 , A2 , . . . , Am ))

2. ψ: S1 [R1 ] ⇐ groupby(S2 [R2 σ](A1 , A2 , . . . , Am ))
3. ψ: S1 [R1 ] ⇐ normalise(S2 [R2 ](A1 , A2 , . . . , Am ))
4. ψ: S1 [R1 ] ⇐ normalise(S2 [R2 σ](A1 , A2 , . . . , Am )) 

In Definition 4, we say that ψ matches R1 and R2 . SCAS specify 1:n, n:1, m:n
relationships between relations with distinct schemas. Here we use the symbol
“⇐” instead of “←” in order to emphasize that the correspondence is not 1:1
as is usual in the most part of schema matching approaches. SCAS are used
to describe the summary of a relation whose tuples are related to the tuples of
another relation by gathering them into logical groups. This means that a SCA
has only the necessary information to indicate which grouping field is involved in
64 V. Pequeno et al.

the relationship and the process used to grouping the tuples. ψ3 shown in Fig. 2
is a simple example of a SCA.
GCAs specify the relationship 1:1, 1:n, n:1, or m:n between attributes of
relations that are matched by a SCA.
Definition 5. Let A be an attribute of R1 . Let also E i , for 1 ≤ i ≤ m, be
attribute expressions over R2 . A Grouping Correspondence Assertion (GCA) is
an expression of one of the following forms:

1. ψ: S1 [R1 ] • A ⇐ E1 5. ψ: S1 [R1 ] • A ⇐ γ(ϕ(E1 , E2 , . . . , Em ))

2. ψ: S1 [R1 ] • A ⇐ ϕ(E1 , E2 , . . . , Em )
6. ψ: S1 [R1 ] • A ⇐ γ(E1 , p).
3. ψ:S1[R1 ]•A ⇐ (E1 , p1 );. . . ;(Em ; pm );v
4. ψ: S1 [R1 ] • A ⇐ γ(E1 ). 7. ψ: S1 [R1 ]•A ⇐ γ(ϕ(E1 , E2 , . . . , Em ), p)

Where ϕ is a function over attributes of R2 , pj (for 1 ≤ j ≤ m) are boolean

conditions over attributes of R2 , v is a value, and γ is one of the aggregate
functions: sum (summation), max (maximum), min (minimum), avg (average),
or count. We say that ψ matches R1 and R2 . 
Consider the relations S2 .film and M.rating. ψ4 , represented in Fig. 2, spec-
ifies that M.rating.quantity corresponds to the counting of all distinct values
of S2 .film.rate.
Definition 6. Let S1 , S2 , . . . , Sn and T be relational schemas; R1 be a relation
schema of T, and R2 a relation schema of some Si , 1 ≤ i ≤ n. Let also E j
(for 1 ≤ j ≤ m) be expressions as defined in Definition 2. A schema matching
between schemas S1 , S2 , . . . , Sn and the schema T is a set M of CAs such that:
1. if M has an ACA ψ such that ψ matches R1 and R2 , then M has a RCA ψ 
that matches R1 and R2 .
2. if M has a GCA ψ such that ψ matches R1 and R2 , then M has a SCA ψ 
that matches R1 and R2 .
3. if M has a RCA ψ such that ψ matches R1 and R2 , then M has an ACA
ψ  : S1 [R1 ](A1 , . . . , An ) ← (E1 , . . . , En ) that matches R1 and R2 . 

4 Specifying New CAs

In Sect. 1, we identified the following types of relationships between schemas
elements that are not properly handled in current schema matching approaches:
(1) matches involving explicit join conditions; (2) matches involving outer-joins;
and (3) matches involving data-metadata. Join (and outer-join) relationships
can express one-to-one or many-to-many correspondences between the relations
involved. Matches involving data-metadata can express many-to-many corre-
spondences between the relations involved. So, we extend our previous defini-
tions of RCA and SCA in order to better specify these types of matchings. In
the following text consider Si relational schemas, Ri relation schemes of Si (for
1 ≤ i ≤ 3), θ a join condition between R2 and R3 , and Aj attributes of R2 (for
1 ≤ j ≤ n)
 Using Complex Correspondences for Integrating Relational Data Sources 65

Definition 7. A Relation Correspondence Assertion (RCA) is an expression of

one of the following forms:

1. Expressions shown in Definition 1 4. ψ: S1 [R1 ] ← S2 [R2 ]   S3 [R3 ]θ

2. ψ: S1 [R1 ] ← S2 [R2 ]  S3 [R3 ]θ 5. ψ: S1 [R1 ] ← S2 [R2 ]   S3 [R3 ]θ
3. ψ: S1 [R1 ] ← S2 [R2 ]  S3 [R3 ]θ 

Definition 8. A Summation Correspondence Assertion (SCA) is an expression

of one of the following forms:

1. Expressions shown in Definition 4 4. ψ: S1 [R1 ] ⇐ S2 [R2 ]  S3 [R3 ]θ

2. ψ: S1 [R1 ] ⇐ S2 [R2 ]  S3 [R3 ]θ 5. ψ: S1 [R1 ] ⇐ S2 [R2 ]  S3 [R3 ]θ
3. ψ: S1 [R1 ] ⇐ S2 [R2 ]  S3 [R3 ]θ 6. ψ:S1[R1 ] ⇐ metadata(S2 [R2 ](A1 ,. . . ,An )) 

Consider the three examples about join, outer-join, and data-metadata cor-
respondences described in Sect. 1. The correspondence between M.remakes and
both S1 .movies and S2 .film can be specified by the SCA ψ5 shown in Fig. 3. ψ5
specifies that M.remakes corresponds to a join between S1 .movies and S2 .film
where the join condition: S2 .film.title= S1 .movie.film and S2 .film.year >
S1 .movie.year is satisfied.

Fig. 3. Examples of CAs involving joins, outer-joins and data-metadata.

The correspondence between M.movie and both S2 .film and S1 .movie can
be specified by the RCA ψ6 , shown in Fig. 3. ψ6 specifies that M.movie corre-
sponds to a left outer-join between S2 .film and S1 .movie.
The correspondence between M.filmMakers and S1 .movieMakers can be
specified by the SCA ψ7 , shown in Fig. 3. ψ7 specifies that M.filmMakers cor-
responds to grouping S1 .movieMakers by the attribute id, being that a data-
metadata translation should be performed (i.e., some data should be converted
into metadata).
Once the schema matching is finished, the CAs generated can be used, for
example, to generate mapping expressions that convert data sources into data
target. We propose that the mapping expressions are automatically generated in
the form of SQL queries, which are used to load the relations (the materialized
views) of the integrated schema.
66 V. Pequeno et al.

5 From CAs to Mapping Expressions

In our proposal, the process to create queries to transform data from a schema
to another one consists of three steps:

1. Indicate the source schemas and the integrated schema using a high-level data
model. In our case, we use the RDM.
2. Define the CA that formally specify the relationships between the integrated
schema and the source schemas.
3. Generate a set of queries based on the CAs generated in step 2, in order to
populate the relations of the integrated schema.

In order to illustrate our approach, consider the integrated schema M and

the sources schemas S1 and S2 shown in Fig. 1.
Now, we should define CAs between M and S1 , and CAs between M and
S2 . In our work, the CAs are specified using a GAV approach rather than a
LAV one. One of reasons for our choices was due to the GAV approach makes
the query answering easier than LAV one, both in materialized and in virtual
integration approaches.
The process to generate the CAs consists of the following steps:

1. To each relation RT of the target T do:

(a) Identify the correspondences at a relation level (i.e., if there is a RCA or
a SCA matching a target relation RT and some source relation RS).
(b) Identify the correspondences at an attribute level: (1) identify the ACAs
between the attributes of RT and RS (if there is a RCA between RT and
RS); (2) identify the GCA between the attributes of RT and RS (if there
is a SCA between RT and RS).
(c) Determine which RCA and SCA can be combined to form a single CA.

In the current work, CAS were manually specified. However, we can use tradi-
tional schema matching tools (e.g.,[8,9]) as a starting point to find basic match-
ings. Then these basic matchings can be enriched through our formalism (using
the CAs).
Some examples of RCAs, ACAs, SCAs, and GCAs between elements of M
and the source schemas S1 and S2 can be found in Figs. 3 and 4.
The final step in the process of creating queries to transform data from
a schema to another is the generation of the queries. In our proposal, they
are defined based on the definition of the schemas and the CAs. Here we use
SQL syntax of MySQL, since MySQL is an open source database that allows to
combine the information from many databases in a single query. However, our
CAs can be used to generate queries in any SQL syntax or even other federating
queries languages as SchemaSQL [16].
Let M be a set of CAs that defines a matching between the source schemas
S1 , S2 and the integrated schema G, that is, M satisfies the conditions stated
in Definition 6. Algorithm 1 shows the procedure to automatically generate the
statements of SQL queries from the CAs in M.
 Using Complex Correspondences for Integrating Relational Data Sources 67

Fig. 4. Examples of ACAs and GCAs.

The Algorithm 1 generates a set of SQL queries, one for each relation schema
RT in the integrated schema. First it spans all ACAs and GCAs that relates
attributes of RT , and puts the correct value in lists S, J, and LA, in accordance
to the type of the CA. S keeps the relation schemas that will be included in the
FROM clause, J keeps the join conditions that will be included in the WHERE
clause, and LA keeps the attributes that will be included in the SELECT clause.
The procedure G SQL ACA(), shown in Algorithm 2, spans the ACAs, while the
procedure G SQL GCA(), shown in Algorithm 3, spans the GCAs. After, the
algorithm spans the RCAs and SCAs, of RT , in order to create the SQL query
to load RT , using templates in Table 1. In accordance to type of CA besides S,
J, and LA, other variables are needed to keep the join conditions that will be
included in the ON clause (θ), the relation schema that will be included in (inner,
outer, left, or right) JOIN clause (RJ), and the grouping attributes that will be
included in the GROUP BY clause (G). Due to space limitations, Algorithms
1, 2, and 3, as well as the Table 1, do not cover the whole set of CAs as defined
in Definitions 3, 5, 7, and 8.
In Algorithms 2, we assumed that ϕ() is a pre-defined SQL function or a
user-defined function on SQL. In Algorithms 3, JAux, and Aux are lists used
when it is necessary to create temporary tables in SQL. This occurs when the
SQL query is created from a SCA of metadata. JAux stores the joins that will
be included in the WHERE clause of the temporary table, while Aux keeps the
relation schema that will be the alias of the temporary table.
In Table 1, Att() is a function that returns the list of attribute names of a
relation schema. We use the short word outer join to emulate a UNION of a
LEFT JOIN and a RIGHT JOIN, since MySQL does not support directly full
outer-joins.
The SQL queries generated by our algorithms can be used to compute the
data target once, and to recompute them at pre-stablished times in order to
maintain the target data up-to-date (this approach is named rematerialization).
Generally, a more efficient approach is to periodically modify only part of the tar-
get data to reflect updates in data sources (this approach is named incremental
maintenance). Rematerialization is adequate, for example, when the integrated
schema is firstly populated, or in situations involving complex operations.
68 V. Pequeno et al.

Algorithm 1 . Generate the SQL query to load a integrated schema G from the
sources.
for all relation schema N = RT in G do
Let ψR be a RCA or SCA of RT
if ψR is a RCA then G SQL ACA( RT ,ψR )
else G SQL GCA( RT ,ψR )
append [RT ] to S
switch ψR do
case ψR : G[RT ] ← S[R]
if J = [ ] then use template T1
else use template T3
case ψR : G[RT ] ← S1 [R1 ] ∪ S2 [R2 ]
append [R2 ] to RJ
θ keeps join conditions formed by primary key attributes of R1 and R2
if J = [ ] then use template T7
else use template T8
case ψR : G[RT ] ← S1 [R1 ]  S2 [R2 ]θ
append [R2 ] to RJ
if J = [ ] then use template T5
else use template T6
case ψR : G[RT ] ⇐ metadata(S[R](A))
append [A] to AL
if J = [ ] then use template T13
else use template T4
case ψR : G[RT ] ⇐ normalise(S[R](A))
append [A] to G
Use template T15
end switch
end for

Algorithm 2. G SQL ACA().

Input: RT , ψR
Let S[R] be a source’s relation in ψR
for all attribute AT in RT do
while ∃ψA /ψA is an ACA of AT relating it to some atribute of S[R] do
if ψA : G[RT ] • AT ← S[R] • A1 then append [A1 ] to LA
if ψA : G[RT ] • AT ← ϕ(S[R] • A1 , S[R] • A2 , . . . , S[R] • Am ) then append
[ϕ(A1 , A2 , . . . , Am )] to LA
end while
end for

Figure 5 presents the SQL query to transform data from S1 .movie and S2 .film
to M.movie from the RCA ψ6 and ACA ψ8 , ψ9 , ψ10 , and ψ11 . The “select” clause
(in line 2) is derived based on ACAS ψ8 , ψ9 , ψ10 and ψ11 . The “from” clause (in
line 3) implements a join operation as specified by RCA ψ6 . The “on” clause (in
line 4) is based on the join condition indicated in the end of ψ6 .

Fig. 5. Query definition to populate M.movie from S2 .film and S1 .movie.

 Using Complex Correspondences for Integrating Relational Data Sources 69

Algorithm 3. G SQL GCA( ).

Input: RT , ψR
Let S[R] be a source’s relation in ψR
for all attribute AT in RT do
while ∃ψA /ψA is an GCA of AT relating it to some atribute of S[R] do
if ψA : G[RT ] • AT ← S[R] • A1 then append [A1 ] to LA
if ψA : G[RT ] • AT ← S[R] • /Bk then
for all FK in  do
Let R1 and R2 be relation schemas related by FK
Let [a1 , . . . , an ] and [b1 , . . . , bn ] be lists of key attributes of, respectively, R1 and R2
append [R1 , R2 ] to S
append [a1 = b1 , . . . , an = bn ] to J
append [B] to LA
end for
if ψA :G[RT ] • AT ← (S[R] • A1 , p1 ) and ψR is of metadata then
append [A1 ] to LA
append [Temp R] to Aux
append [p1 ] to JAux
end while
end for

Table 1. Templates to generate SQL Statements induced by RCAS and ACAS.

Figure 6 presents the definition of the query to transform data from

S1 .movieMakers to M.filmMakers. For this query, we have to define a
nested select statement to each case-base GCA that relates attributes of
S1 .movieMakers to attributes of M.filmMakers. Each nested select state-
ment must be joined through an outer-join in order to guarantee both: i)
that duplicate tuples will be merged properly, and 2) not duplicate tuples will
be stored in M.filmMakers. Thus, the clauses “from” (line 3), “outer join”
(line 7), and “on” (line 12) correctly implement the data-metadata relationship
specified by the SCA ψ7 . The “on” clause (line 12) is based on the attribute
indicated in ψ7 . The first nested select statement (lines 4 to 6) is defined based
on the GCAS ψ12 and ψ14 . The second nested select statement (lines 8 to 11) is
similar to the first one, but now it is based on ψ13 and ψ14 . The “select” clause
in line 2 is based on the left-hand side of GCAs ψ12 , ψ13 and ψ14 .
70 V. Pequeno et al.

Fig. 6. Query definition to populate M.filmMakers from S1 .movieMakers.

6 Empirical Evaluation
We have performed some preliminary tests to verify that our approach is
tractable for reasonably sized input.

6.1 Study Case Scenario

For our evaluation, we create a case study to simulate a situation close to the real
world. We need to integrate information of three different sources: IES, FSP, and
CDV to get a more complete information about Brazil’s universities. IES contains
data about Brazil’s universities (name, city, state, etc.). It has a single relation
(ies 2011) with 26 attributes and 2366 tuples3 . FSP contains data about the
ranking of the Brazil’s universities (ranking, university, grade, etc.). It has two
relations, but only one of them (ranking, with 13 attributes and 191 tuples)
was used in the evaluation4 . CDV contains data about the living cost of some
cities of Brazil. It has a single relation (livingCost) with 5 attributes and 84
tuples5 .
The integrated schema, named G, contains the necessary structure to keep
the information required by the designer. It contains 8 relations with a total of
31 attributes and 8 foreign keys, as can be saw in the Fig. 7.

6.2 Method
We measure the performance of the data translation (i.e., the run time of
the queries to load the schema G). For our tests, we have used a Macbook
Pro/2.3GHz Intel Core (4GB of RAM and 499Gb of HD) running OSx 10.9.5.
All databases were locally stored in this machine using the MySQL 5.6.
We first manually defined the CAs, with the aid of a tool implemented by
us. For this case study, we defined 3 RCAs, 5 SCAs of normalize, 31 ACAs,
3
IES data was extract from http://www.dados.gov.br/dataset/instituicoes-de-
ensino-superior.
4
FSP data was extract from http://ruf.folha.uol.com.br/2014/rankingdeuniversi
dades/.
5
CDV data was extract from http://wwwcustodevida.com.br/brasil.
 Using Complex Correspondences for Integrating Relational Data Sources 71

Fig. 7. The integrated schema G.

and 8 GCAs, being a total of 52 CAs. Using the Algorithm 1, we generate 8

SQL queries: 5 queries of group by, 2 simple select-from queries, and 1 more
complex query that simulates the outer join operator. Some queries use stored
functions defined to looks for the value of a primary key in a target relation
based on the attribute value of a source relation.
For data translation test, we measured the time that MySQL took to load
each target relation using the queries generated by the Algorithm 1. Due to the
run time of SQL queries can change depending on internal and external factors,
we ran each query by 50 times and took the average to each 10 executions. All
tests were performed locally in a same machine and only the MySQL server
and MySQLWorkbench were running at the time. The result of the test can be
observed through the chart shown in Fig. 8.

Fig. 8. Run time query by quantity of executions chart.

72 V. Pequeno et al.

6.3 Discussion
We noted that the run time to most queries are more or less constant and below
that 60 ms. It is not a surprise that the queries with higher execution time were
those to load fsp ranking (about 200ms) and university (about 590 ms),
because both has more tuples to load than the others. Considering the number
of tuples of the university (more than 2000) and that the query generated is a
bit complex (includes left-join, right-join, union all, and 5 stored functions), we
believe that 590 ms is a good performance.

7 Related Work
Schema matching is an important step of the data integration process. Typically,
1:1 correspondences between two different schemas are manually defined using
a GUI or are (semi-) automatically discovered using matchers (usually through
heuristics). Each correspondence, in general, only specifies which elements refer
to a same attribute or relation in the real world [17]. AgreementMaker [4], and
OII Harmony [9] are some examples of tools for schema matching. Agreement-
Maker [4] can match schemas and ontologies using schema information as well as
instance-level data to generate the correspondences. OII Harmony [9] combines
multiple matchers algorithms based on natural-language processing to identify
correspondences between schemas.
Correspondences such as those defined/generated in [4,9] do not provide all
necessary information for discovering expressions to transform data sources in
data target (i.e., the mapping expressions), the next phase in the schema map-
ping process. Richer models for specifying correspondences between schemas
were proposed by [8,11–13] and [5](chap. 3). These approaches allow to define
one-to-one or many-to-one attribute correspondences (i.e., association between
attributes of two schemas). COMA++ [8] is a generic prototype for schema
and ontology matching, schema-based and instance-based, and support a semi-
automatic or manual enrichment of simple 1:1 correspondences into more com-
plex mapping expressions including functions to support data transformations.
[13] describes the IMAP system, which semi-automatically discovers complex
matches, using different kinds of information such as domain knowledge, and
domain integrity constraints to improve matching accuracy. [5](chap. 3) and [11]
allow to express conditional correspondences (i.e., the value of an attribute A
is the same of an attribute B if a given condition is satisfied). More closely to
our approach is the work in [12]. In [12], the authors allow to manually specify
one-to-one correspondence assertions between elements of Entity Relationship
models. Although they cannot specify many-to-many matches, their correspon-
dences have some semantic and allow to specify relationships such as: equiva-
lence, union, intersection, and selection.
[10] specify one-to-one and many-to-many basic, complex, and semantic
matches between elements of object-relational schemas. They can specify most
part of the correspondences specified in [12] and other more complex. For exam-
ple, they can deal with aggregate functions, denormalisations, and grouping (i.e.,
group by in SQL). Joins and outer-joins are implicitly defined based on the
 Using Complex Correspondences for Integrating Relational Data Sources 73

integrity constraints or match functions6 . A distinguished feature of the app-

roach proposed in [10] is that it allows to match, in the same correspondence,
relations and attributes of two or more schemas. Yet, the information they pro-
vide is not sufficient, since they do not explicitly enable the specification of join
paths and its variants, nor to deal with data-metadata relationships.
Data-metadata translations between elements of different relational schemas
have been studied extensively. SchemaSQL [16] and FIRA/FISQL [18] are the
most notable works on this subject. SchemaSQL [16] is a SQL-like metadata
query language that uses view statements to restructure one column of val-
ues of a relation into metadata in another one. FISQL [18] is a sucessor of
SchemaSQL and it is equivalent to the query algebra FIRA. Both SchemaSQL
and FIRA/FISQL were proposed to provide interoperability in relational multi-
database systems. Our SCA of metadata was based on the promote metadata
operator of FIRA.

8 Conclusions

This paper focused on present CAs that deal with 1:1 and m:n matchings between
schemas components, including correspondences involving aggregations, joins,
and metadata. We emphasize that, in our approach, the CAs can specify basic
and complex correspondences with semantics. Using CAs, we shown how SQL
queries can be automatically generated to populate relations (views) of a global
schema.
We presented some preliminary tests to evaluate the performance of the
queries generated from CAs. We intend to realize more tests to evaluate the
performance to different types of queries and other datasets.
We currently are working in as specifying complex correspondences between
relational schemas and (RDF). Some initial work was published in [19]. We intent
extend the initial proposal with the CAs presented here.

Acknowledgements. This work was partially supported by national funds through

FCT - Fundação para a Ciência e a Tecnologia, under the project PEst-OE/EEI/
LA0021/2013, DataStorm Research Line of Excellency funding (EXCL/EEI-ESS/
0257/2012) and the grant SFRH/BPD/76024/2011. We are especially grateful to Diego
Cardoso (UFC, Brazil) for the implementation of the algorithms.

References
1. Kimball, R., Ross, M., Thornthwaite, W., Mundy, J., Becker, B.: The Data Ware-
house Lifecycle Tookit, 2nd edn. Wiley, Indianapolis (2008)
2. Popfinger, C.: Enhanced Active Databases for Federated Information Systems.
PhD thesis, Heinrich Heine University Düsseldorf (2006)

6
Match functions are functions that determine if two different instances represent the
same concept in the real world.
74 V. Pequeno et al.

3. Langegger, A., Wöß, W., Blöchl, M.: A semantic web middleware for virtual data
integration on the web. In: Bechhofer, S., Hauswirth, M., Hoffmann, J., Koubarakis,
M. (eds.) ESWC 2008. LNCS, vol. 5021, pp. 493–507. Springer, Heidelberg (2008)
4. Cruz, I.F., Antonelli, F.P., Stroe, C.: Agreementmaker: efficient matching for large
real-world schemas and ontologies. Proc. VLDB Endow. 2(2), 1586–1589 (2009)
5. Bellahsene, Z., Bonifati, A., Rahm, E. (eds.): Schema Matching and Mapping.
Data-Centric Systems and Applications. Springer, Heidelberg (2011)
6. Rahm, E., Bernstein, P.A.: A survey of approaches to automatic schema matching.
VLDB J. 10(4), 334–350 (2001)
7. Shvaiko, P., Euzenat, J.: A survey of schema-based matching approaches. In: Spac-
capietra, S. (ed.) Journal on Data Semantics IV. LNCS, vol. 3730, pp. 146–171.
Springer, Heidelberg (2005)
8. Massmann, S., Raunich, S., Aumueller, D., Arnold, P., Rahm, E.: Evolution of the
COMA match system. In: The 6th Intl. Workshop on Ontology Matching. (2011)
9. Mork, P., Seligman, L., Rosenthal, A., Korb, J., Wolf, C.: The Harmony integration
workbench. J. Data Semant. 11, 65–93 (2008)
10. Pequeno, V.M., Pires, J.C.M.: Using perspective schemata to model the ETL
process. In: ICMIS 2009, pp. 332–339. World Academy of Science, Engineering
and Technology (2009)
11. Bohannon, P., Elnahrawy, E., Fan, W., Flaster, M.: Putting context into schema
matching. In: VLDB, pp. 307–318 (2006)
12. Vidal, V.M.P., Lóscio, B.F.: Updating multiple databases through mediators. In:
ICEIS 1999, pp. 163–170 (1999)
13. Dhamankar, R., Lee, Y., Doan, A., Halevy, A.Y., Domingos, P.: IMAP: Discovering
complex mappings between database schemas. In: ACM SIGMOD, pp. 383–394
(2004)
14. Yan, L.L., Miller, R.J., Haas, L.M., Fagin, R.: Data-driven understanding and
refinement of schema mappings. In: ACM SIGMOD, pp. 485–496. ACM (2001)
15. Pequeno, V.M., Aparı́cio, J.N.: Using correspondence assertions to specify the
semantics of views in an object-relational data warehouse. In: ICEIS 2005, pp.
219–225 (2005)
16. Lakshmanan, L., Sadri, F., Subramanian, I.: SchemaSQL - a language for inter-
operability in relational multi-database systems. In: VLDB, pp. 239–250. Morgan
Kaufmann (1996)
17. Doan, A., Halevy, A., Ives, Z.: Principles of Data Integration. Morgan Kaufmann,
Waltham (2012)
18. Wyss, C.M., Robertson, E.L.: Relational languages for metadata integration. ACM
Trans. Database Syst. 30, 624–660 (2005)
19. Pequeno, V.M., Vidal, V.M.P., Casanova, M.A., Neto, L.E.T., Galhardas, H.: Spec-
ifying complex correspondences between relational schemas and rdf models for gen-
erating customized R2RML mappings. In: IDEAS 2014, pp. 96–104. ACM (2014)
 Towards an Efficient and Distributed DBSCAN
Algorithm Using MapReduce

Ticiana L. Coelho da Silva(B) , Antonio C. Araújo Neto, Regis Pires Magalhães,

Victor A.E. de Farias, José A.F. de Macêdo, and Javam C. Machado

Computing Science Department, Federal University of Ceara, Fortaleza, Brazil

{ticianalc,javam}@ufc.br,
{antonio,regispires,victorfarias,jose.macedo}@lia.ufc.br

Abstract. Clustering is a major data mining technique that groups a

set of objects in such a way that objects in the same group are more
similar to each other than to those in other groups. Among several types
of clustering, density-based clustering algorithms are more efficient in
detecting clusters with varied density and different shapes. One of the
most important density-based clustering algorithms is DBSCAN. Due to
the huge size of generated data by the widespread diffusion of wireless
technologies and the complexity of big data analysis, new scalable algo-
rithms for efficiently processing such data are needed. In this chapter we
are particularly interested in using traffic data for finding congested areas
in a city. For this purpose, we developed a new distributed and efficient
strategy of DBSCAN algorithm that uses MapReduce to detect dense
areas based on the input parameters. We conducted experiments using
real traffic data of a brazilian city, Fortaleza, and compared our approach
with the centralized and the MapReduce-based approaches. Our prelim-
inary results confirmed that our approach is scalable and more efficient
than the other ones. We also present an incremental version of DBSCAN
considering the MapReduce version of it.

Keywords: DBSCAN · MapReduce · Traffic data

1 Introduction
One of the most important density-based clustering algorithms known in litera-
ture is DBSCAN (Density-based Spatial Clustering of Application with Noise) [4].
Its advantages over other clustering techniques are that it groups data into clus-
ters of arbitrary shape, it does not require a priori number of clusters, and it deals
with outliers in the data set. However, DBSCAN is more computationally expen-
sive than other clustering algorithms, such as k-means, for example. Moreover,
DBSCAN does not scale when executed on large data sets in a single machine.
Recently, many researchers have been using cloud computing infrastructure in
order to solve scalability problems of some traditional centralized clustering algo-
rithms [2]. Thus, the strategy to parallelize DBSCAN in shared-nothing computer
clusters is an adequate solution to solve such problems [9].
c Springer International Publishing Switzerland 2015
J. Cordeiro et al. (Eds.): ICEIS 2014, LNBIP 227, pp. 75–90, 2015.
DOI: 10.1007/978-3-319-22348-3 5
76 T.L. Coelho da Silva et al.

Clearly, the provisioning of an infrastructure for large scale processing

requires software that can take advantage of the large amount of machines and
mitigate the problem of communication between them. It has been increasing
the amount of techniques to manage computer clusters, among which stands out
the paradigm MapReduce [3] and its open source implementation Hadoop [15],
used to manage large volumes of data across clusters of computers.
Hadoop framework is attractive because it provides a simple programming
model that makes it easier for users to implement relatively sophisticated dis-
tributed programs. The MapReduce programming model is recommended for
parallel processing of large data volumes in computational clusters [8]. It is also
scalable and fault tolerant. The MapReduce platform divides a task into small
activities and materializes its intermediate results locally. When a fault occurs
in this process, only failed activities are re-executed.
This chapter is based on [10] which aims at identifying, in a large data set,
traffic jam areas in a city using mobility data. In this sense, a parallel version
of DBSCAN algorithm, based on the MapReduce paradigm, was proposed as a
solution. Related works, such as [6] and [2] also use MapReduce to parallelize
DBSCAN and they are explained on Sect. 3.
The main contributions of this work are: (1) Our partitioning strategy is less
costly than the one proposed on [2]. In [2], the authors present a grid-based
partitioning strategy. Our approach is traffic data aware and it partitions the
data set with regard to one attribute (in our experiments we used the streets’
name); (2) To gather clusters of different partitions, our merge strategy does
not need data replication as [6] and [2]. Moreover, our approach guarantees the
same result of the centralized DBSCAN; (3) We proved that our distributed
DBSCAN algorithm is correct on Sect. 4; (4) An incremental and distributed
version of DBSCAN.
The structure of this chapter is organized as follows. Section 2 introduces
basic concepts needed to understand the solution of the problem. Section 3
presents our related work. Section 4 addresses the methodology and implemen-
tation of this work, that involves the solution of the problem. The discussion of
the incremental version is presented in Sect. 5. The experiments are described in
Sect. 6. Section 7 presents our conclusion.

2 Preliminary

2.1 MapReduce

The need for managing, processing, and analyzing efficiently large amounts of data
is a key issue in the Big Data scenario. To address these problems, different solu-
tions have been proposed, including the migration/building applications for cloud
computing environments, and systems based on Distributed Hash Table (DHT)
or structure of multidimensional arrays [12]. Among these solutions, there is the
MapReduce paradigm [3], designed to support the distributed processing of large
data sets on clusters of servers and its open source implementation Hadoop [15].
Towards an Efficient and Distributed DBSCAN Algorithm Using MapReduce 77

The MapReduce programming model is based on two primitives of functional

programming: Map and Reduce. The MapReduce execution is carried out as fol-
lows: (i) The Map function takes a list of key-value pairs (K1 , V1 ) as input and a list
of intermediate key-value pairs (K2 , V2 ) as output; (ii) the key-value pairs (K2 , V2 )
are defined according to the implementation of the Map function provided by the
user and they are collected by a master node at the end of each Map task, then
sorted by the key. The keys are divided among all the Reduce tasks. The key-value
pairs that have the same key are assigned to the same Reduce task; (iii) The Reduce
function receives as input all values V2 from the same key K2 and produces as out-
put key-value pairs (K3 , V3 ) that represent the outcome of the MapReduce process.
The reduce tasks run on one key at a time. The way in which values are combined
is determined by the Reduce function code given by the user.
Hadoop is an open-source framework developed by the Apache Software Foun-
dation that implements MapReduce, along with a distributed file system called
HDFS (Hadoop Distributed File System). What makes MapReduce attractive
is the opportunity to manage large-scale computations with fault tolerance. The
developer only needs to implement two functions called Map and Reduce. There-
fore, the system manages the parallel execution and coordinates the implementa-
tion of Map and Reduce tasks, being able to handle failures during execution.

2.2 DBSCAN

DBSCAN is a clustering algorithm widely used in the scientific community. Its

main idea is to find clusters from each point that has at least a certain amount of
neighbors (minP oints) within a specified range (eps), where minP oints and eps
are input parameters. Finding values for both the parameters might be difficult,
once it depends on the data and on the information one desires to discover. The
following definitions are used in DBSCAN algorithm and they will be used in
the Sect. 4:

– Card(A): cardinality of the set A.

– Neps (o): p ∈ Neps (o), if and only if the distance between p and o is less or
equal than eps.
– Directly Density-Reachable (DDR): o is DDR from p, if o ∈ Neps (p) and
Card(Neps (p)) ≥ minP oints.
– Density-Reachable (DR): o is DR from p, if there is a chain of points
{p1 , ..., pn } where p1 = p and pn = o, such as pi+1 is DDR from pi and
∀i ∈ {1, ..., n − 1}.
– Core Point: o is a Core Point if Card(Neps (o)) ≥ minP oints.
– Border Point: p is a Border Point if Card(Neps (p)) < minP oints and p is
DDR from a Core Point.
– Noise: q is Noise if Card(Neps (q)) < minP oints and q is not DDR from any
Core Point.

DBSCAN finds for each point o, Neps (o) in the data. If CardNeps (o) ≥
minPoints, a new cluster C is created and it contains the points o and
Neps (o). Then each point q ∈ C that has not been visited is also
78 T.L. Coelho da Silva et al.

checked. If Neps (q) ≥ minPoints, each point r ∈ Neps (q) that is not
in C is added to C. These steps are repeated until no new point is
added to C. The algorithm ends when all points from the data set are visited.

3 Related Work
P-DBSCAN [7] is a density-based clustering algorithm based on DBSCAN for
analysis of places and events using a collection of geo-tagged photos. P-DBSCAN
introduces two new concepts: density threshold and adaptive density, that is used
for fast convergence towards high density regions. However P-DBSCAN does not
have the advantages of the MapReduce model to process large data sets.
Another related work is GRIDBSCAN [14], that proposes a three-level clus-
tering method. The first level selects appropriate grids so that the density is
homogeneous in each grid. The second stage merges cells with similar densi-
ties and identifies the most suitable values of eps and minP oints in each grid
that remain after merging. The third step of the proposed method executes the
DBSCAN method with these identified parameters in the data set. However,
GRIDBSCAN is not suitable for large amounts of data. Our proposed algorithm
in this work is a distributed and parallel version of DBSCAN that uses MapRe-
duce and is suitable for handling large amounts of data.
The paper [6] proposes an implementation of DBSCAN with a MapReduce
of four stages using grid based partitioning.
The paper also presents a strategy for joining clusters that are in different
partitions and contain the same points in their boundaries. Such points are
replicated in the partitions and the discovery of clusters, that can be merged
in a single cluster, is analyzed from them. Note that the number of replicated
boundary points can affect the clustering efficiency, as such points not only
increase the load of each compute node, but also increase the time to merge the
results of different computational nodes.
Similar to the previous work, [2] proposes DBSCANMR that is a DBSCAN
implementation using MapReduce and grid based partitioning. In [2], the parti-
tion of points spends a great deal of time and it is centralized.
The data set is partitioned in order to maintain the uniform load balancing
across the compute nodes and to minimize the same points between partitions.
Another disadvantage is the proposed strategy is sensitive to two input parame-
ters. How to obtain the best values for those parameters is not explained in the
paper. The DBSCAN algorithm runs on each compute node for each partition
using a KD-tree index. The merge of clusters that are on distinct partitions is
done when the same point belongs to such partitions and it is also tagged as a
core point in any of the clusters. If it is detected that two clusters should merge,
they are renamed to a single name. This process also occurs in [6].
This work has similar approach to the papers [6] and [2] because they con-
sider the challenge of handling large amounts of data and use MapReduce to
parallelize the DBSCAN algorithm. However, this chapter presents a data par-
titioning technique that is data aware, which means partitioning data according
to their spatial extent. The partitioning technique proposed by [2] is centralized
Towards an Efficient and Distributed DBSCAN Algorithm Using MapReduce 79

and spends much time for large amount of data as we could see on our exper-
iments. Furthermore, unlike this work to merge clusters, [6] and [2] proposed
approaches that require replication, which can affect the clustering efficiency.
The cluster merging phase in this work checks if points previously considered as
a noise point becomes a border or core point.

4 Methodology and Implementation

Mobility data has been fostered by the widespread diffusion of wireless technolo-
gies, such as call details records from mobile phones and GPS tracks from naviga-
tion devices, society-wide proxies of human moving behaviour. These data opens
new opportunities for discovering the hidden patterns and models that charac-
terize the trajectories humans follow during their daily activity. This research
topic has recently attracted scientists from several fields, being not only a major
intellectual challenge, but also a very important issue to domains such as urban
planning, sustainable mobility, transportation engineering, public health, and
economic forecasting [5].
The increasing popularity of mobility data has become the main source for
evaluating the traffic situation to support drivers’ decisions related to displace-
ment in a city in real time. Traffic information in big cities can be collected from
GPS devices or from traffic radars, or even gathered from tweets. This information
can be used to complement the data generated by cameras and physical sensors
in order to guide municipality actions in finding solutions to mobility problems.
Through these data it is possible to analyze where the congested areas are within
a city in order to discover which regions are more likely to have traffic jams. Such
discovery may assist the search for effective reengineering traffic solutions in the
context of smart cities. This is an application for our propose approach.
Normally the traffic data indicates the name of the street, the geographic
position (latitude and longitude), the average speed of vehicles at the moment,
among others. We address in this chapter the problem of discovering density
areas, that may be represent a traffic jam. We also consider that frequent updates
can happen in the data set. After insertions or deletions to the dynamic data-
base, the clusters discovered by DBSCAN has to be updated. So we discuss
how the incremental approach of DBSCAN [11] algorithm can be applied on our
distributed approach.
In this section, we focus on the solution of finding density areas or clusters
from raw traffic data on MapReduce. We formulate the problem as follows:

Problem Statement. Given a set of d-dimensional points on the traffic data

set DS = {p1 , p2 , ..., pn } such that each point represents one vehicle with the
average speed, the eps value, the minimum number of points required to form a
cluster minPoints and a set of virtual machines VM = {vm1 , vm2 , ..., vmn } with
a MapReduce program; find the density areas on the traffic data with respect
to the given eps and minP oints values. In this chapter, we only consider two
dimensions (latitude and longitude) for points. Furthermore, each point has the
information about the street it belongs to.
80 T.L. Coelho da Silva et al.

Fig. 1. The phases of distributed and parallel DBSCAN execution.

4.1 MapReduce Phases and Detection of Possible Merges

This section presents the implementation of the proposed solution to the prob-
lem. Hereafter, we explain the steps or phases to parallelize DBSCAN using the
MapReduce programming model as we can see on Fig. 1.

– Executing DBSCAN Distributedly. This phase is a MapReduce process.

Map and Reduce functions for this step are explained below.
1. Map function. Each point from the data set is described as a pair
key, value, such that the key refers to the street and the value refers to a
geographic location (latitude and longitude) where the data was collected.
As we could see on Algorithm 1.
2. Reduce function. It is presented on Algorithm 2. This function receives
a list of values that have the same key, i.e. the points or geographical
positions (latitude and longitude) that belongs to the same street. The
DBSCAN algorithm is applied in this phase using the KD-tree index [1].

Algorithm 1. First MapReduce - Map.

Input: Set of points of the data set T
1 begin
2 for p ∈ T do
3 createP airp.street name, p.Lat, p.Lon

The result is stored in a database. This means that the identifier of each
cluster and the information about their points (such as latitude, longitude, if
it is core point or noise) are saved.
– Computing Candidate Clusters. Since there are many crossroads between
the streets in the city and the partitioning of data is based on the streets, it
is necessary to discover what are the clusters of different streets that intersect
each other or may be merged into a single cluster. For example, it is common
in large cities have the same traffic jam happening on different streets that
Towards an Efficient and Distributed DBSCAN Algorithm Using MapReduce 81

Algorithm 2. First MapReduce - Reduce.

Input: Set P of pairs k, v with same k, minPoints, eps
1 begin
2 DBSCAN(P, eps, minPoints)
3 Store results on database;

intersect to each other. In other words, two clusters may have points at a
distance less than or equal to eps in such a way that if the data were processed
by the same reduce or even if they were in the same partition, they would be
grouped into a single cluster. Thus, the clusters are also stored as a geometric
object in the database and only the pairs of objects that have distance at
most eps will be considered in the next phase that is the merge phase. Tuples
with pairs of candidate clusters for merge are passed with the same key to the
next MapReduce. As we could see on Algorithm 3.

Algorithm 3. Find merge candidates clusters.

Input: Set C of Clusters
Output: V is a set of merge candidates clusters
1 begin
2 for Ci ∈ C do
3 Create its geometry Gi ;
4 for all geometries Gi and Gj and i <> j do
5 if Distance(Gi , Gj ) ≤ eps then
6 V ← Ci , Cj 

7 return V ;

Algorithm 4. Second MapReduce - REDUCE.

Input: Set V of pairs Ci , Cj  of clusters candidates to merge
1 begin
2 for Ci , Cj  ∈ V do
3 if CanM erge(Ci , Cj ) then
4 Rename Cj to Ci in V

– Merging Clusters. It is also described by a MapReduce process. Map and

reduce functions for this phase are explained below.
1. Map function. It is the identity function. It simply passes each key-value
pair to the Reduce function.
82 T.L. Coelho da Silva et al.

2. Reduce function. It receives as key the lowest cluster identifier from all
the clusters that are candidates to be merged into a single cluster. The
value of that key corresponds to the other cluster identifiers that are
merge candidates. Thereby, if two clusters must be merged into a single
cluster, the information about points belonging to them are updated. The
Algorithms 4 and 5 are executed in this phase.
Lines 2 to 5 from Algorithm 5, we check and update the neighbors of
each point pi ∈ Ci and pj ∈ Cj . This occurs because if Ci and Cj are
merge clusters candidates, there are points in Ci and Cj that the distance
between them is less or equal than eps. On the lines 6 to 11, the algorithm
verifies if there is some point pi ∈ Ci and pj ∈ Cj that may have become
core points. This phase is important, because Ci and Cj can merge if
there is a core point pi ∈ Ci and another core point pj ∈ Cj , such that
pi is DDR from pj as we can see on lines 12 to 15 on the Algorithm 5. In
the next section, we present a proof that validate our merge strategy.
This work considers the possibility that a noise point in a cluster may
become a border or core point with the merge of clusters different of
our related works. We do that on the line 16 of Algorithm 5 calling the
procedure updateNoisePoints(). Considering that pi ∈ Ci , pj ∈ Cj and
pi ∈ Neps (pj ), if pi or pj were noise points before the merge phase and
it occurred an update on Neps (pi ) and Neps (pj ), pi or pj could not be
more a noise point, but border point or core point. That is checked on
the procedure updateNoisePoints().
Our merge phase is efficient, because it does not consider replicated data as
our related work. Next, we prove that our strategy merges clusters candidates
correctly.

4.2 Validation of Merge Candidates

Theorem 1. Let C1 be a cluster of a partition S1 , C2 be a cluster of a partition
S2 , and S1 ∩ S2 = ∅. Clusters C1 and C2 should merge if there are two points
p1 ∈ C1 and p2 ∈ C2 that satisfy the following properties:

– Distance(p1 , p2 ) ≤ eps;
– Neps (p1 ) ≥ minP oints in S1 ∪ S2 ;
– Neps (p2 ) ≥ minP oints in S1 ∪ S2 ;

Proof. First, we can conclude that there are at least two points p1 ∈ C1 and
p2 ∈ C2 such that the distance between them is less than or equal to eps,
otherwise it would be impossible for any points from C1 and C2 to be placed in
the same cluster in the centralized execution of DBSCAN because they would
never be Density-Reachable (DR). Moreover, such condition is necessary to allow
the merge between two clusters. Still considering the points p1 and p2 , we have
the following possibilities:

1. p1 and p2 are core points in C1 and C2 respectively;

2. p1 is core point in C1 and p2 is border point in C2 ;
Towards an Efficient and Distributed DBSCAN Algorithm Using MapReduce 83

Algorithm 5. CanMerge.
Input: Clusters Ci , Cj candidates to merge
1 begin
2 for pi ∈ Ci do
3 for pj ∈ Cj do
4 if ((pi ∈ Neps (pj )) then
5 setAdjacent(pi , pj )

6 for pi ∈ Ci do
7 if (Card(Neps (pi )) ≥ minP oints) then
8 pi .isCore ← true

9 for pj ∈ Cj do
10 if (Card(Neps (pj )) ≥ minP oints) then
11 pj .isCore ← true

12 for pi ∈ Ci do
13 for pj ∈ Cj do
14 if (Adj(pi , pj ) ∧ pi .isCore ∧ pj .isCore) then
15 merge ← true
16 updateNoisePoints();

17 if (merge) then
18 for pj ∈ Cj do
19 pj .cluster ← i

20 return merge

3. p1 is border point in C1 and p2 is core point in C2 ;

4. p1 and p2 are border points in C1 and C2 respectively;

Analyzing the first case, where p1 is a core point, by definition

Card(Neps (p1 )) ≥ minP oints. Considering the partitions S1 ∪ S2 , we observe
that p2 ∈ Neps (p1 ). When being visited during the execution of DBSCAN, the
point p1 would reach point p2 directly by density. As p2 is also a core point,
all others points from C2 could be density reachable from p1 . Thus, the points
from clusters C1 and C2 would be in the same cluster. Similarly, we can state
the same for point p2 , as it could reach by density all points from C1 through
point p1 . In this case, C1 and C2 will be merged.
The analysis is analogous in the second case, where only p1 is a core point.
Thus, when visiting p1 its neighbors, particularly p2 , will be expanded. Consid-
ering the space S1 ∪ S2 , suppose that Neps (p2 ) < minP oints. So, the point p2 is
not expanded when visited and point p1 will not reach the core point belonging
to cluster C2 , that is a p2 neighbor. In this case, C1 and C2 will not be merged.
Similarly one can analyze the third case and reach the same conclusion of
the second case.
84 T.L. Coelho da Silva et al.

For the fourth and last case, consider that none of the two points p1 and p2
have more than or equal to minP oints neighbors, i.e., Neps (p1 ) < minP oints
and Neps (p2 ) < minP oints in S1 ∪ S2 . When visited, neither will be expanded,
because they will be considered border points. In the case that only one of
them has more than minP oints neighbors in S1 ∪ S2 (the previous case), we
could see that such condition is not enough to merge the clusters. Therefore, the
only case in which such clusters will merge is when Neps (p1 ) ≥ minP oints and
Neps (p2 ) ≥ minP oints in S1 ∪ S2 or in the first case (p1 and p2 are core points).

5 Incremental Version DBSCAN

Due to the high rate of updates in some data sets, the results of a data mining
algorithm execution may change very often, and so the analisys. Also, when it
comes to large volumes of data, it is infeasible to rerun the algorithm at every
data set update.
Incremental data mining means applying data mining algorithms on incre-
mental database, i.e., that changes with time. The goal of incremental data min-
ing algorithms is to minimize the scanning and calculation effort for updated
records [13].
Typically, updates are applied on the data set periodically. Thus, the clusters
computed by the clustering algorithm have to be updated as well. Due to the
very large size of the databases, it is highly desirable to perform these updates
incrementally [11]. The paper [11] examines which part of an existing clustering
is affected by an update of the database in a centralized way.
On this chapter we present how the algorithms for incremental updates of a
clustering can be applied after insertions and deletions based on our distributed
DBSCAN.

5.1 Insertions

Consider that we have the clusters derived from the original data set using our
DBSCAN algorithm. Let C be the clustering using DBSCAN, such that C =
{S0 , ..., Sm , Out} where Out is called the set of outliers and ∀i, i ∈ {0, ..., m} , Si
be a DBSCAN cluster.
For each new point inserted on the data set, a geometric object is stored
on the database. We can imagine that each new point is a cluster Si , so
C  = {S0 , ..., Sn }. We should compare the elements of C and C  in order to
find possible merges between them, or if the new points just create new clusters
and some outliers. Thus, we can apply the Computing candidate clusters and
Merging clusters phases described on the Sect. 4.1 to find the merges between
C and C  or to find the new clusters. For Sj ∈ C  that do not merge with any
cluster or create a new cluster, Sj becomes an outlier on the whole data set.
Note that for incremental DBSCAN we just reuse only two phases of our pro-
posed distributed and parallel DBSCAN. It is cheaper than executing DBSCAN
from scratch for the updated data set.
Towards an Efficient and Distributed DBSCAN Algorithm Using MapReduce 85

5.2 Deletions
As well as insertions of new entries on the data set might happen, the deletion of
old entries is also possible. Considering this, one can easily see that the deletion
of some particular points or even of a subset of the data set might change the
clustering result. For example, a core point might become a border point or
a noise point and break a cluster into two. Similar cases are considered and
presented on [11], where the authors proved the correctness of their strategy.
When it comes to a distributed version of DBSCAN, the deletion of entries in
the data set is a special case due to the phisycal distribution of the data across
a computer cluster.
Considering a previous clustering processing, for each point in the data set,
whether it belongs to a cluster or it is an outlier. As stated in [11], a point
removal affects only a part of the data set, that belongs to the same cluster
of the point to be removed. Thus, in a distributed scenario, each cluster would
be assigned to a specific machine, and if the point to be removed belongs to a
cluster Si , the processing needs to be done only on the machine that contains
this cluster. Such strategy allows multiple parallel point removals from multiple
clusters with no interference between them. Furthermore, the processing on the
machine is similar to the one proposed originally on [11], but with the benefits
of executing it at the same time for several clusters.

6 Experimental Evaluation
The experimental evaluation was performed in a private cloud environment at
the Federal University of Ceara. In total, 11 virtual machines running Ubuntu
12.04, each with 8GB of RAM and 4 CPU units, were deployed for this purpose.
The Hadoop version used on each machine was 1.1.2 and the environment vari-
ables were set according to the Table 1. Each scenario was carried out 5 times
and reported the average, maximum and minimum observed values of execution
time.
The data sets used in the experiments were related to avenues from Fort-
aleza city in Brazil and the collected points were retrieved from a website that
obtains the data from traffic radar. Each entry contains the avenue’s name, the
geographic position of vehicles (latitude and longitude), as well as its speed at
a specific moment of time. In the context of the problem, what our approach
identifies groups with high density of points from the data set that have low
speeds. The results can be used to detect traffic jam areas in Fortaleza city.
The first test varied the eps value and kept the same amount of points in each
street, i.e., of the original data set that corresponds to 246142 points. The eps
values used were 100, 150, 200, 250, 300 and 350, while keeping the value of 50
to minP oints. As it was expected, the Fig. 2 shows that with the increase of eps,
the processing time has also increased, because more points in the neighborhood
of a core point could be expanded.
The Fig. 3 illustrates the variation of the number of points from the input
data set related to the processing time. The number of points used was 246142,
86 T.L. Coelho da Silva et al.

Fig. 2. eps variation.

Fig. 3. Data set size variation.

Table 1. Hadoop configuration variables used in the experiments.

Variable name Value

hadoop.tmp.dir /tmp/hadoop
fs.default.name hdfs://master:54310
mapred.job.tracker master:54311
mapreduce.task.timeout 36000000
mapred.child.java.opts -Xmx8192m
mapred.reduce.tasks 11
dfs.replication 5

324778, 431698 and 510952 points. As expected, when the number of points
processed by DBSCAN is increased, the processing time also increases, showing
that our solution is scalable.
The Table 2 presents a comparison of our approach execution time and cen-
tralized DBSCAN execution time. We varied the number of points that was
246142, 324778, 431698 and 510952 points for eps = 100 and minP oints = 50.
In all cases, our approach found the same clusters that centralized DBSCAN on
the data set, but spent less time to process as we expected.
The Figs. 4 and 5 show a comparison between our approach and DBSCANMR
[2] that has the partitioning phase centralized, different of our approach. On the
Towards an Efficient and Distributed DBSCAN Algorithm Using MapReduce 87

Table 2. Comparing the execution time of our approach and centralized DBSCAN for
eps=100 and minPoints = 50.

Data set Our Approach [ms] Centralized [ms]

246142 197263 1150187.6
324778 254447 2002369
431698 330431 3530966.6
510952 409154 4965999.6

Fig. 4. Varing the data set size and comparing our approach with DBSCANMR.

Fig. 5. Varing eps and comparing our approach with DBSCANMR.

both experiments, our solution spent less processing time than DBSCANMR,
because DBSCANMR spends great cost to build the grid during the centralized
partitioning phase.
Moreover, DBSCANMR strategy is sensitive to two input parameters that
are the number of points that could be processed on memory and a percent-
age of points in each partition. For these two parameters, the paper does not
present how they could be calculated and what are the best values. We did the
experiments using the first one equals to 200000 and the second as 0,49.
We did not compare our approach with the other related work [6]. However,
we believe that our approach presents better results than [6], because our merge
strategy does not need data replication, that can affect the clustering efficiency for
a large data set.
The Fig. 6 presents the points plotted for 246142 points of data set. Note
that each color represents an avenue of Fortaleza. On the Fig. 7, we can see
88 T.L. Coelho da Silva et al.

Fig. 6. The data set with 246142 points plotted.

Fig. 7. Clusters found after run our approach for 246142 points (eps=100 and min-
Points = 50).

the clusters found by our approach using eps = 100 and minP oints = 50. Each
cluster found is represented by a different color on the Fig. 7. Note that the
merge occurred where there are more than one avenue that crosses each other
as we expected.

7 Conclusions

Clustering is a major data mining technique that groups a set of objects in such
a way that objects in the same group are more similar to each other than to
those in other groups. Among many types of clustering, density-based clustering
algorithms are more efficient in detecting clusters with varied density and dif-
ferent shapes. One of the most important density-based clustering algorithms is
the DBSCAN.
In this chapter we present a distributed DBSCAN algorithm using MapRe-
duce to identify congested areas within a city using a large traffic data set. We
also discuss an incremental version for DBSCAN using our distributed DBSCAN.
This incremental version is cheaper than running the distributed DBSCAN from
scratch.
Our approach is more efficient than DBSCAN-MR as confirmed by our exper-
iments, while varying the data set size and the eps value. We also compare our
Towards an Efficient and Distributed DBSCAN Algorithm Using MapReduce 89

approach to a centralized version of DBSCAN algorithm. Our approach found

the same clusters as the centralized DBSCAN algorithm, moreover our approach
spent less time to process, as expected.

References
1. Bentley, J.L.: Multidimensional binary search trees used for associative searching.
In: Communications of the ACM, vol. 18, pp. 509–517. ACM (1975)
2. Dai, B.-R., Lin, I.-C.: Efficient map/reduce-based dbscan algorithm with optimized
data partition. In: 2012 IEEE 5th International Conference on Cloud Computing
(CLOUD), pp. 59–66. IEEE (2012)
3. Dean, J., Ghemawat, S.: Mapreduce: simplified data processing on large clusters.
Commun. ACM 51(1), 107–113 (2008)
4. Ester, M., Kriegel, H.-P., Sander, J., Xu, X.: A density-based algorithm for discov-
ering clusters in large spatial databases with noise. In: KDD, vol. 96, pp. 226–231
(1996)
5. Giannotti, F., Nanni, M., Pedreschi, D., Pinelli, F., Renso, C., Rinzivillo, S.,
Trasarti, R.: Unveiling the complexity of human mobility by querying and mining
massive trajectory data. VLDB J. Int. J. Very Large Data Bases 20(5), 695–719
(2011)
6. He, Y., Tan, H., Luo, W., Mao, H., Ma, D., Feng, S., Fan, J.: Mr-dbscan: an efficient
parallel density-based clustering algorithm using mapreduce. In: 2011 IEEE 17th
International Conference on Parallel and Distributed Systems (ICPADS), pp. 473–
480. IEEE (2011)
7. Kisilevich, S., Mansmann, F., Keim, D.: P-dbscan: a density based clustering algo-
rithm for exploration and analysis of attractive areas using collections of geo-tagged
photos. In: Proceedings of the 1st International Conference and Exhibition on Com-
puting for Geospatial Research & Application, p. 38. ACM (2010)
8. Lin, J., Dyer, C.: Data-intensive text processing with mapreduce. Synth. Lect.
Hum. Lang. Technol. 3(1), 1–177 (2010)
9. Pavlo, A., Paulson, E., Rasin, A., Abadi, D.J., DeWitt, D.J., Madden, S., Stone-
braker, M.: A comparison of approaches to large-scale data analysis. In: Proceed-
ings of the 2009 ACM SIGMOD International Conference on Management of data,
SIGMOD 2009, pp. 165–178, New York, NY, USA. ACM (2009)
10. Coelho da Silva, T.L., Araujo, A.C.N., Magalhaes, R.P., de Farias, V.A.E., de
Macêdo, J.A., Machado, J.C.: Efficient and distributed dbscan algorithm using
mapreduce to detect density areas on traffic data. In: Proceedings of International
Conference on Enterprise Informational Systems, ICEIS 2014
11. Ester, M., Kriegel, H.-P., Sander, J. and Wimmer, M., Xu, X.: Incremental clus-
tering for mining in a data warehousing environment. In: The VLDB Journal –
The International Journal on Very Large Data Bases (1998)
12. Sousa, F.R.C., Moreira, L.O., Macêdo, J.A.F., Machado, J.C.: Gerenciamento de
dados em nuvem: Conceitos, sistemas e desafios. Simpsio Brasileiro de Banco de
Dados. SBBD 2010, pp. 101–130 (2010)
13. Chakraborty, S., Nagwani, N.K.: Analysis and study of Incremental DBSCAN clus-
tering algorithm. In: International Journal of Enterprise Computing and Business
Systems 2011, pp. 101–130
90 T.L. Coelho da Silva et al.

14. Uncu, O., Gruver, W.A., Kotak, D.B., Sabaz, D., Alibhai, Z., Ng, C.: Gridbscan:
grid density-based spatial clustering of applications with noise. In: IEEE Interna-
tional Conference on Systems, Man and Cybernetics, 2006. SMC 2006, vol. 4, pp.
2976–2981. IEEE (2006)
15. White, T.: Hadoop: the definitive guide. O’Reilly, Sebastopol (2012)
 Discovering Frequent Patterns
on Agrometeorological Data with TrieMotif

Daniel Y.T. Chino1(B) , Renata R.V. Goncalves2 , Luciana A.S. Romani3 ,

Caetano Traina Jr.1 , and Agma J.M. Traina1
1
Institute of Mathematics and Computer Science, University of São Paulo,
São Carlos, Brazil
{chinodyt,caetano,agma}@icmc.usp.br
2
Cepagri-Unicamp, Campinas, Brazil
[email protected]
3
Embrapa Agriculture Informatics, Campinas, Brazil
[email protected]

Abstract. The “food safety” issue has concerned governments from

several countries. The accurate monitoring of agriculture have become
important specially due to climate change impacts. In this context, the
development of new technologies for monitoring are crucial. Finding pre-
viously unknown patterns that frequently occur on time series, known as
motifs, is a core task to mine the collected data. In this work we present
a method that allows a fast and accurate time series motif discovery.
From the experiments we can see that our approach is able to efficiently
find motifs even when the size of the time series goes longer. We also
evaluated our method using real data time series extracted from remote
sensing images regarding sugarcane crops. Our proposed method was
able to find relevant patterns, as sugarcane cycles and other land covers
inside the same area, which are really useful for data analysis.

Keywords: Time series · Frequent motif · Remote sensing image

1 Introduction
One of the issues being pursued by the database researchers is how to take
advantage of the large volume of data daily generated by the plethora of sen-
sors placed in many environments and systems. The information gathered by
the sensors are usually stored in time series databases, being a rich source for
decision making for the owners of such data. One of the main tasks when mining
time series is to find the motifs present therein, that is, to find patterns that
frequently occurs in time series. That is, the motifs provide key information to
mine association rules aimed at spotting patterns indicating that some events
frequently lead to others.
Existing applications involving time series, such as stock market analysis,
are not yet able to record all the factors that govern the data stored in the time
series, such as political and technological factors. On the contrary, climate vari-
ations nowadays have most of its governing factors being recorded, using sensing
c Springer International Publishing Switzerland 2015
J. Cordeiro et al. (Eds.): ICEIS 2014, LNBIP 227, pp. 91–107, 2015.
DOI: 10.1007/978-3-319-22348-3 6
92 D.Y.T. Chino et al.

equipments such as satellites and ground-based weather stations. However, the

diversity of data available makes it hard to discover complex patterns that can
support more robust analyzes. In this scenario, finding motifs has an even greater
importance.
An example of time series, is the one extracted from remote sensing imagery
containing Normalized Dierence Vegetation Index (NDVI) measurements. The
NDVI time series present the vegetative strength of the plantation [14]. To follow
the development of a crop is strategic for agribusiness practices in Brazil, since
agriculture is the country’s main asset. The accurate monitoring of agriculture in
the whole world have become more and more important specially due to climate
change impacts. The “food safety” issue has concerned governments from several
countries and the development of new technologies for monitoring, as well as the
proposition of mitigation and adaptation measures, are crucial. In this sense,
remote sensing can be an important tool to improve the fast detection of changes
in the land cover besides to aid at monitoring the crop cycle.
Since the volume of time series databases as well as the length of the series is
growing at a very fast pace, it is mandatory to develop algorithms and methods
that can deal with time series in the scenario of big data. In this paper we
present the TrieMotif, a new method to extract motifs from time series and a
new algorithm to index them in a trie data structure that performs well over
large time series, which is up to 3 times faster than the state-of-the-art method.
We evaluated TrieMotif over both synthetic and real data time series, obtaining
very promising results.
This paper is organized as follows. Section 2 summarizes the main concepts
used as basis to develop our work. Section 3 describes the TrieMotif algorithm
and Sect. 4 discusses its evaluation. Section 5 shows the TrieMotif performance
on real data obtained from remote sensing images. Section 6 concludes this paper.

2 Background and Related Works

A time series motif is a pattern that occur frequently. They were first defined in
[11] and a generalized definition was given in [2]. In this section we recall these
definitions and notations, as they will be used in this paper. First we begin with
a definition of time series:

Definition 1. Time Series: A time series T = t1 , . . . , tm is an ordered set of m

real-valued variables.

Since we want to find patterns that frequently occur along a time series, we will
not work with the whole time series, we are aiming only at parts of a time series,
which are called subsequences and are defined as follows.

Definition 2. Subsequence: Given a time series T of length m, a subsequence

Sp of T is a sampling of length n < m of contiguous positions from T beginning
at position p, that is, Sp = tp , . . . , tp+n−1 for 1 ≤ p ≤ m − n + 1.
 Discovering Frequent Patterns on Agrometeorological Data with TrieMotif 93

30

20

10

-10

-20
Sq
Trivial Match
-30
0 20 40 60 80 100 120 140 160 180
Time

Fig. 1. The best matches of a subsequence Sq are probably the trivial matches that
occur right before or after Sq .

In order to find frequent patterns, we need to define a matching between patterns.

Definition 3. Match: Given a distance function D(Sp , Sq ) between two sub-

sequences, a positive real number R (range) and a time series T containing a
subsequence Sp and a subsequence Sq , if D(Sp , Sq ) ≤ R then Sq is called a
matching subsequence of Sp .

On subsequences of the same time series, the best matches are probably subse-
quences that are slightly shifted. Matching between two overlapped subsequences
is called a trivial match. Figure 1 illustrates the idea of a trivial match. The triv-
ial match is defined as follows:

Definition 4. Trivial Match: Given a time series T , containing a subsequence

Sp and a matching subsequence Sq , we say that Sq is a trivial match to Sp either
if p = q or if there is no subsequence Sq beginning at q  such that D(Sp , Sq ) > R,
and either q < q  < p or p < q  < q. That is, if two subsequences overlaps, there
must exist a subsequence between them that is not a match.

These definitions allow defining the problem of finding Frequent K-Motif. First,
all subsequences are extracted using a sliding window. Then, since we are inter-
ested in patterns, each subsequence is z-normalized to have zero mean and one
standard deviation [6]. The K-Motif is defined as follows.

Definition 5. Frequent K-Motifs: Given a time series T , a subsequence of

length n and a range R, the most significant motif in T (1-Motif ) is the subse-
quence F {1} that has the highest count of non-trivial matches. The K th most sig-
nificant motif in T (K-Motif ) is the subsequence F {K} that has the highest count
of non-trivial matches, and satisfies D(F {K} , F {i} ) > 2R, for all 1 ≤ i < K.

The Nearest Neighbor motif was defined by Yankov et al. [17], and it represents
the closest pair of subsequences. In our proposed work, we focus on the Fre-
quent K-Motif problem and we will be referring to them as K-Motif. Since the
K-Motifs are unknown patterns, a brute-force approach would compare every
94 D.Y.T. Chino et al.

-1

-2

-3
0 20 40 60 80 100 120
Time

Fig. 2. A time series subsequence of size n = 128 is reduced to a PAA representation

of size w = 8 and then is mapped to a string of a = 3 symbols aabbcccb.

subsequence with each other. It has quadratic computational cost on the time
series size, since it requires O(m2 ) distance function evaluations.
An approach to reduce the complexity of this problem employs dimension-
ality reduction and discretization of the time series [11]. The SAX (Symbolic
Aggregate approXimation) technique allows time series of size n to be repre-
sented by strings of arbitrary size w (w < l) [10]. For a given time series, SAX
consists of the following steps. Firstly, the time series is z-normalized, so that the
data follow normal distribution [4]. Next, the normalized time series is converted
into the Piecewise Aggregate Approximation (PAA) representation, decreasing
the time series dimensionality [5]. The time series is then replaced with w values
corresponding to the average of the respective segment. Thus, in the PAA repre-
sentation, the time series is divided into w continuous segments of equal length.
Finally, the PAA representation is discretized into a string with an alphabet of
size a > 2. Figure 2 shows an example of a time series subsequence of size n = 128
discretized using SAX with w = 8 and a = 3. It is also possible to compare two
SAX time series using the MINDIST function. The MINDIST lower bounds the
Euclidean distance [12], warranting no false dismissals [3].
Chiu et al. proposed a fast algorithm based on Random Projection [2]. Each
time series subsequence is discretized using SAX. The discretized subsequences
are mapped into a matrix, where each row points back to the original position of
the subsequence on the time series. Then, the algorithm uses the random projec-
tion to compute a collision matrix, which counts the frequency of subsequence
pairs. Through the collision matrix, the subsequences are checked on the origi-
nal domain seeking for motifs. Although fast, the collision matrix is quadratic
on the time series length, requiring a large amount of memory. Also using the
SAX discretization, Li and Lin [8,9] proposed a variable length motif discovery
based on grammar induction. Catalano et al. [1] proposed a method that works
on the original domain of the data. The motifs are discovered in linear time and
constant memory costs using random sampling. However, this approach can lead
to poor performance for long time series with infrequent motifs [13].
 Discovering Frequent Patterns on Agrometeorological Data with TrieMotif 95

Several works proposed to solve the motif discovery problem taking advantage
of tree data structures. Udechukwu et al. proposed an algorithm that uses a suffix
tree to find the Frequent Motif [15]. The time series are discretized considering
the slopes between two consecutive measures. The symbols are chosen according
to the angle between the line joining the two points and the time axis. Although
this algorithm do not require to set the length of the motif, the algorithm is
affected by noise. A suffix tree was also used to find motifs in multivariate time
series [16]. Keogh et al. solved a problem similar to the motif discovery using a
trie structure to find the most unusual subsequence in a time series [7].
The TrieMotif algorithm is up to 3 times faster and requires up to 10 times
less memory than the current state of the art, the Random Projection approach,
because TrieMotif selects only the candidates that are most probably a match
to a motif. Using this approach, TrieMotif reduces the number of unnecessary
distance calculations.

3 The TrieMotif Algorithm

In this section we present the TrieMotif algorithm. Since a motif is an unknown
pattern, one of the problem of finding them is the need to compare every sub-
sequence with every other. On this context, we intend to reduce the number of
subsequence comparisons by discarding subsequences that are discrepant from
each other. Let Sq be a possible motif, matching subsequences might have values
similar to Sq and therefore they might be on a region near Sq , as shown in Fig. 3.
The TrieMotif algorithm defines this area by setting an upper and a lower limit
to Sq , this way, we only compare possible matching subsequences. The TrieMotif
algorithm consists of three stages:
– First, all subsequences are extracted from the time series and converted into
a symbolic representation;
– The subsequences are indexed using a Trie and a list of possible non trivial
matches (candidates) are generated for each subsequence using the Trie index;
– The distances between the motif candidates on the original time series are
calculated.
On the first stage, all subsequences Si of size n are extracted using a slid-
ing window and are z-normalized. These subsequences pass through a dimen-
sionality reduction process to reduce the computational cost on the next stage.
A subsequence of size n can be represented as a sequence of size w, via the PAA
algorithm. On the next step of this stage, subsequences are converted into a
symbolic representation. This representation is obtained by dividing the inter-
val of the subsequence values into a equal size bins, where each bin receives
a symbol. Each value in the subsequence is converted into the symbol of the
corresponding bin. Figure 4 shows an example that converts a subsequence S of
size m = 128 and values between [−2.83, 1.36] into a string of size w = 8 using
an alphabet a of 3 symbols. Initially the subsequence passes through a dimen-
sionality reduction via PAA with w = 8. Then, assuming a = 3, three bins are
96 D.Y.T. Chino et al.

2.5 2

2
1
1.5
0
1

0.5
-1
0
-2
-0.5

-1 -3
Sq -20 0 20 40 60 80 100 120
-1.5
0 10 20 30 40 50 60 70 80 Time

Fig. 3. Subsequences Si that satisfy Fig. 4. A time series subsequence of

D(Sq , Si ) ≤ R are possibly in the high- size m = 128 is converted into a string
lighted area. of a = 3 symbols and size w = 8.

created: a = [−2.83, −1.43), b = [−1.43, −0.03) and c = [−0.03, 1.36]. Notice,

that we kept the zero mean and the standard deviation requirements. Finally,
the symbolic representation of S is Ŝ = abcbcccb.
To find the top K-Motifs, the brute force algorithm calculates the distance
of each subsequence Sq to every other subsequence. Our proposal reduces the
number of distance calculations by selecting only candidates Si that may be
a match – the trivial matches are discarded on the next stage. To select the
candidates we index all subsequences (already represented as a string) in a trie.
For example, consider the subsequences S1 , S2 , S3 and S4 , w = 4 and a = 4,
as shown in Fig. 5. After processed in the first stage, they become Sˆ1 = aabd,
Sˆ2 = babd, Sˆ3 = abda and Sˆ4 = dcca respectively. Figure 6(a) shows how the trie
is built.

2 2
1.5 1.5
1 1
0.5 0.5
0 0
-0.5 -0.5
-1 -1
-1.5 -1.5
^ ^ ^ ^
S1 S2 S3 S4 S 1 S2 S 3 S 4
-2 -2
0 2 4 6 8 10 12 14 0 2 4 6 8 10 12 14
Time Time
(a) Original time series. (b) PAA representantion.

Fig. 5. Symbolic conversion process of the subsequences S1 , S2 , S3 and S4 to the strings

Sˆ1 = aabd, Sˆ2 = babd, Sˆ3 = abda and Sˆ4 = dcca.

An exact search on the trie would return candidates faster, but some can-
didates Si that are a match could be discarded. If we search for candidates of
Sˆ1 , although Sˆ2 is probably a match, it would not be selected. To solve this
problem, we modified the exact search on the trie to a range-like search. On the
 Discovering Frequent Patterns on Agrometeorological Data with TrieMotif 97

exact search, when the algorithm is processing the j th element of Sˆq (Sˆq [j]),
it only visits the path of the trie where N odesymbol = Sˆq [j]. In our proposed
approach, we associate numerical values to the symbols. For example, A is equal
to 1, B is equal to 2, and so on. Therefore, we can take advantage of closer
values to compute the similarity when comparing the symbols. On our modified
search, the algorithm also visits paths where |Sˆq [j] − N odesymbol | ≤ δ. That is,
if Sˆq [j] = b and δ = 1, then the algorithm visits the paths of a, b and c (one
up or one down). As backtracking all possible paths might be computationally
expensive, we exploit pruning of unwanted paths by indexing the elements of
the strings in a non-sequential order. This approach is interesting, because time
series measures over a short period tend to have similar values. For example,
if in a given time there is a b symbol, probably the next symbol might be a, b
or c (even in large alphabets). Taking that into account, we interleave elements
from the beginning and the end of the string, i.e., the first symbol, then the last
symbol, then the second and so on.
Figure 6(a) shows how the modified search behave when searching for candi-
dates to Sˆ1 . On the first level, Sˆ1 [1] = A and the algorithm will visit the paths
of the symbols a and b. On the next level, Sˆ1 [4] = D and it will visit the paths
of c and d. This process is repeated until it reaches a leaf node. In this example,
it will return the candidates Sˆ1 and Sˆ2 , but Sˆ3 and Sˆ4 will not be checked. Fig-
ures 6(a) and 6(b) also show that changing the order of the elements can reduce
the backtracking. If the subsequences were indexed using the sequential order,
our modified search would also visit the path of Sˆ3 for at least one more level
(marked in blue), while changing the order, this path is never visited.

(a) Non-sequential order indexing. (b) Sequential order indexing.

Fig. 6. By selecting candidates using our modified search, the algorithm backtracks
only on nodes of the strings Sˆ1 and Sˆ2 paths. Creating the trie index using normal
order increases the backtracking, since the algorithm visits part of the string Sˆ3 path
(Color figure online).

On the last stage, after obtaining a list of candidates for each Sq , we calculate
the distance between Sq and Si on the original time series domain, discarding
trivial matches of Sq . We also make sure that the K-Motifs satisfy Definition 5.
Algorithm 1 shows how to use the TrieMotif algorithm to locate the top K-
Motifs. First, all subsequences Si are converted into a symbolic representation Ŝi
and every Ŝi is indexed into a trie index. Through the Trie index, we can reduce
98 D.Y.T. Chino et al.

the number of non-trivial match calculations by generating a set C of the possible

candidates for every Ŝi . Then, we check on the original time series domain if the
subsequences Cj in C satisfies D(Si , Cj ) ≤ R and it is not a trivial match. Since
it is not possible to know the top K most frequent motifs before computing every
subsequence, we store the motif on a list (ListOf M otif ). As the last step of the
algorithm, it is needed to check if the top K-Motifs satisfies Definition 5. To
do so, we sort ListOf M otif by the decreasing number of non-trivial matches.
Thus, the most frequent motifs appear at the list head. Thereafter we iterate
through the sorted ListOf M otif and whenever F {i} satisfies Definition 5, we
insert F {i} in the result set T opKM otif , otherwise F {i} is discarded.
To improve the efficiency of the method, we also calculate every distance
using the “early abandonment” approach. Note that, although we presented a
method to symbolic represent time series subsequences, it is possible to use others
symbolic representations, such as the SAX algorithm. In those cases, provided
that the distance function is the Euclidean one, it is possible to take advantage
of the low complexity of the M IN DIST and discard calculations on the original
time series whenever M IN DIST (Sˆq , Ŝi ) > R, since the M IN DIST is a lower
bound to the Euclidean distance.

4 Synthetic Data Analysis

To validate our method, we performed tests on a synthetic time series generated
by random walk of size m = 1, 000. We also embedded a motif of size 100 in
four different positions of the time series. Figure 7(a) shows the planted motif
and its variants. The motifs were planted on positions 32, 287, 568 and 875, as
shown in Fig. 7(b). We ran the TrieMotif using the bin discretization method
with n = 100, R = 2.5, w = 16, a = 4 and δ = 1. As expected, the TrieMotif
was able to successfully find the motif on the planted positions.
Knowing that our method is able to find motifs, we made a series of exper-
iments varying the length of the time series to evaluate the method efficiency.
We compared our method with the brute force algorithm and with a Random
Projection method, since it used extensively and is the basis of others algorithms

80
2
60

1 40

0 20

0
-1
-20
Original Motif
-2 Variation 1 -40
Variation 2 Time Series
Variation 3 Motif
-3 -60
0 10 20 30 40 50 60 70 80 90 100 0 100 200 300 400 500 600 700 800 900 1000
Time Time
(a) Planted motif and its variants. (b) A random walk time series with the implanted motif
marked in blue.

Fig. 7. Planted motifs into a longer dataset for evaluation tests.

 Discovering Frequent Patterns on Agrometeorological Data with TrieMotif 99

Algorithm 1. K-TrieMotif Algorithm.

Input: T , n, R, w, a, δ, K
Output: List of the top K-Motif
1: ListOf M otif ← ∅
2: for all Subsequence Si with size n of T do
3: Ŝi ← ConvertIntoSymbol(Si , w, a)
4: Insert Ŝi in the Trie index
5: end for
6: for all Subsequence Ŝi of T do
7: C ← GetCandidatesFromTrie(Ŝi , δ)
8: M otif S ← ∅
9: for all Cj ∈ C do
10: if N onT rivialM atch(Si , Cj , R) then
11: Add Cj to M otif S
12: end if
13: end for
14: Insert M otif S in the ListOf M otif
15: end for
16: Sort ListOf M otif by size in decreasing order
17: T opKM otif s ← ∅
18: k ← 0
19: for all (F {i} ∈ ListOf M otif ) and (k < K) do
20: if (Distance(F {i} , F {l} ) ≤ 2R), ∀F {l} ∈ T opKM otif then
21: Discard F {i}
22: else
23: Insert F {i} in T opKM otif
24: k++
25: end if
26: end for
27: return T opKM otif

in the literature. We searched for motifs of size n = 100 and used the Euclidean
distance. For both Random Projection and TrieMotif we used the same parame-
ters for the discretization, a = 4 and w = 16. We ran the Random Projection
with 100 iterations. For the TrieMotif, we used both bin and SAX representa-
tions with δ = 1. We varied the time series length from 1, 000 to 100, 000. Each
set of parameters were tested using 5 different seeds for the random walk. All
tests were performed 5 times, totalizing 25 executions. The wall clock time and
memory usage measurements were taken from these 25 runs of the algorithm.
The presented values correspond to the average of the 25 executions. As it is
too time consuming, the brute-force algorithm was not executed for time series
with lengths above 40, 000. The experiments were performed in an HP server
with 2 Intel Xeon 5600 quad-core processors with 96 GB of main memory, under
CentOS Linux 6.2. All methods (TrieMotif, Random Projection and Brute-force)
were implemented using the C++ programming language. The time spent com-
parison is shown in Fig. 8(a) and the memory usage is shown on Fig. 8(b).
100 D.Y.T. Chino et al.

3500 40
TrieMotif (SAX) TrieMotif
TrieMotif (Bin) 35 Random Projection
3000 Random Projection Brute-force

Memory Usage (GB)

Brute-force 30
Time (seconds)

2500
25
2000
20
1500
15
1000
10
500 5
0 0
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
Time Series Length (x1000) Time Series Length (x1000)
(a) Time efficiency comparison. (b) Memory usage comparison.

Fig. 8. Comparison of the TrieMotif with brute force and random projection.

30
Non−Sequential
Number of Nodes Visited (x1,000)

Sequential
25

20

15

10

0
0 10 20 30 40 50 60 70 80 90
Query Size

Fig. 9. Comparison of the number of visited nodes on the trie index using sequential
and non-sequential order.

As expected, both Random Projection and TrieMotif performed better than

the brute-force approach. For time series of length below 10, 000, both Random
Projection and the TrieMotif had similar performance. However, as the time
series length grows, the TrieMotif presented a increasingly better performance.
This result is due to the fact that TrieMotif selects only the candidates that
are probable matches for a motif, reducing the number of unnecessary distance
calculations. TrieMotif also presents better performance using SAX because the
interval of values is different for each symbol. The symbols corresponding to val-
ues near 0 are smaller and therefore, less candidates are selected. The TrieMotif
also consumes less memory than the Random Projection, since the Random
Projection algorithm needs at least m2 memory for the collision matrix. From
Fig. 8(b), we can see that the memory requirements of our method is significantly
less demanding than Random Projection, requiring 10 times less memory than
the state of the art (the Random Projection approach).
We also checked the differences between indexing the subsequences in sequen-
tial and non-sequential order. For this test we counted the number of nodes vis-
ited for each query on the trie index varying the query size. We executed the
queries for every subsequence of the random walk time series. The presented
 Discovering Frequent Patterns on Agrometeorological Data with TrieMotif 101

values correspond to the average these executions. Figure 9 shows that the
TrieMotif algorithm visited less nodes when using non-sequential order indexing.

5 Real Data Analysis

We also performed experiments using data from real applications. For this eval-
uation, we extracted time series from remote sensing images with two different
spatial resolution. The spatial resolution of a remote sensing image specifies
the size of the area that the pixel represents on the earth surface. On the first
experiment, we extracted data from AVHRR/NOAA1 images, a low spatial res-
olution image with a spatial resolution of 1 kilometer, i.e., each pixel represents
an area of 1km x 1km on the ground. The AVHRR/NOAA images correspond to
monthly measures of the Normalized Difference Vegetation Index (NDVI), which
indicates the soil vegetative vigor represented in the pixels of the images and is
strongly correlated with biomass. We used images of the São Paulo state, Brazil
(Fig. 10), corresponding to the period between April 2001 and March 2010. From
these images, we extracted 174,034 time series of size 108. Each time series corre-
sponds to a geographical point in São Paulo state, excluding the coastal region.

Fig. 10. State of São Paulo, Brazil.

In order to find the motifs in the time series database, we submitted all the
time series to the TrieMotif algorithm, generating a single index. The São Paulo
state is one of the greatest producers of sugarcane in Brazil. In this work, we

1
Advanced Very High Resolution Radiometer/National Oceanic and Atmospheric
Administration.
102 D.Y.T. Chino et al.

0.6 0.6 0.8 0.6

0.6
0.4 0.4 0.4
0.4
NDVI

NDVI

NDVI

NDVI
0.2
0.2 0.2 0.2
0

0 0 -0.2 0
Mar May Jul Sep Nov Jan Oct Dec Feb Apr Jun Aug Aug Oct Dec Feb Apr Jun May Jul Sep Nov Jan Mar

(a) 1-Motif (b) 2-Motif (c) 3-Motif (d) 4-Motif

Fig. 11. Top 4-Motif for the São Paulo state area.

consider a complete sugarcane cycle of approximately one year, each starting

in April and ending in March. We looked for the top 4-Motifs of size n = 12
(annual measures). We ran the TrieMotif algorithm using bin discretization with
R = 1.5, w = 4, a = 4 and δ = 1. The experiments were performed using the
same computational infrastructure of the synthetic data.

Fig. 12. The São Paulo state was divided into 5 regions of interest. Region 1 (purple)
corresponds to water/cities, Region 2 and 3 (blue and green respectively) to a mixture of
crops and grassland, Region 4 (yellow) is sugarcane crop and Region 5 (red) corresponds
to forest. The black lines represent the political counties of the state. The south of the
state is a forest nature federal reserve (Color figure online).

Figure 11 shows the top 4-Motifs for the São Paulo state. The plots are not
z-normalized due to the experts restrictions for analyzing them. As expected,
the two most frequent patterns (Figs. 11(a) and 11(b)) have a behavior similar
to a sugarcane cycle, with high values of NDVI on April and low values on
October. This behavior is due to the fact that sugarcane harvesting begins in
April, when the NDVI slowly starts to decrease. On October, the harvest finishes
and the soil remains exposed, when the NDVI has the lowest value. From October
to March, the sugarcane grows and the NDVI increases again. This pattern
 Discovering Frequent Patterns on Agrometeorological Data with TrieMotif 103

Region Color # of Series Represents

1 Purple 2,804 Water and Cities
2 Blue 55,815 Grassland and
3 Green 50,785 Perennial Crops
4 Yellow 48,301 Sugarcane crops
5 Red 16,329 Forest

Fig. 13. Number of time series in each region of interest.

was found on almost every area of the São Paulo state and according to the
experts was not a coherent result. To overcome this problem, we divided the
São Paulo state into 5 regions of interest as shown in Fig. 12. The regions were
obtained through a clustering algorithm and the results were analyzed by experts
in the agrometeorology and remote sensing fields. According to the experts, the
region 1 found by the algorithm corresponds to water and urban areas, regions
2 and 3 correspond to a mixture of crops (excluding sugarcane) and grassland,
region 4 corresponds to sugarcane crops and region 5 to forest. Figure 13 shows a
summary of each region of interest and the number of time series in each region.

0.8 0.8 0.8 0.8

0.6 0.6 0.6 0.6
0.4 0.4 0.4 0.4
NDVI

NDVI

NDVI

NDVI
0.2 0.2 0.2 0.2
0 0 0 0
-0.2 -0.2 -0.2 -0.2
-0.4 -0.4 -0.4 -0.4
-0.6 -0.6 -0.6 -0.6
Mar May Jul Sep Nov Jan Apr Jun Aug Oct Dec Feb Apr Jun Aug Oct Dec Feb Dec Feb Apr Jun Aug Oct

(a) 1-Motif (b) 2-Motif (c) 3-Motif (d) 4-Motif

Fig. 14. Top 4-Motif for region of interest 1.

0.8 0.8 0.8 0.8

0.6 0.6 0.6 0.6

NDVI

NDVI

NDVI

NDVI

0.4 0.4 0.4 0.4

0.2 0.2 0.2 0.2

0 0 0 0
Jan Mar May Jul Sep Nov Sep Nov Jan Mar May Jul Jul Sep Nov Jan Mar May May Jul Sep Nov Jan Mar

(a) 1-Motif (b) 2-Motif (c) 3-Motif (d) 4-Motif

Fig. 15. Top 4-Motif for region of interest 2.

Figures 14, 15, 16, 17 and 18 show the top 4-Motif for each region of interest.
Once again, the plots are not z-normalized due to the experts restrictions for ana-
lyzing them. The 1-Motif found on region 1 (Fig. 14(a)) has a pattern with low
values and variation, which experts say correspond to urban regions. The other
Motifs found in region 1 (Figs. 14(b), (c) and (d)) have a pattern that corresponds
to water regions. Figures 15 and 16 show patterns with a higher value of NDVI
104 D.Y.T. Chino et al.

0.8 0.8 0.8 0.8

0.6 0.6 0.6 0.6

NDVI

NDVI

NDVI

NDVI
0.4 0.4 0.4 0.4

0.2 0.2 0.2 0.2

0 0 0 0
Dec Feb Apr Jun Aug Oct Jun Aug Oct Dec Feb Apr Aug Oct Dec Feb Apr Jun Apr Jun Aug Oct Dec Feb

(a) 1-Motif (b) 2-Motif (c) 3-Motif (d) 4-Motif

Fig. 16. Top 4-Motif for region of interest 3.

0.8 0.8 0.8 0.8

0.6 0.6 0.6 0.6

NDVI

NDVI

NDVI

NDVI
0.4 0.4 0.4 0.4

0.2 0.2 0.2 0.2

0 0 0 0
Dec Feb Apr Jun Aug Oct Apr Jun Aug Oct Dec Feb Aug Oct Dec Feb Apr Jun Jun Aug Oct Dec Feb Apr

(a) 1-Motif (b) 2-Motif (c) 3-Motif (d) 4-Motif

Fig. 17. Top 4-Motif for region of interest 4.

0.8 0.8 0.8 0.8

0.6 0.6 0.6 0.6

NDVI

NDVI

NDVI

NDVI
0.4 0.4 0.4 0.4

0.2 0.2 0.2 0.2

0 0 0 0
Jun Aug Oct Dec Feb Apr Jan Mar May Jul Sep Nov Oct Dec Feb Apr Jun Aug Jan Mar May Jul Sep Nov

(a) 1-Motif (b) 2-Motif (c) 3-Motif (d) 4-Motif

Fig. 18. Top 4-Motif for region of interest 5.

and with a high variation, which experts attribute to agricultural areas. The top
3-Motifs found on region 4 (see Figs. 17(a), (b) and (c)) is clearly recognized by
agrometeorologists as a sugarcane cycle. Figure 18 shows a pattern with high val-
ues of NDVI and low variations, corresponding to the expected forest behavior,
where the NDVI follows the local temperature variation.
The results obtained by the TrieMotif algorithm on the most prominent
regions confirmed the correctness of the algorithm. However, the most interesting
results correspond to patterns that were not expected by the experts. According
to them, for some regions, the 4-Motifs found do not resemble the previously
known patterns. The 4-Motifs found on both regions 2 and 3 (Figs. 15(d) and
16(d) respectively) indicates the presence of sugarcane and the 4-Motif found on
region 4 (Fig. 17(d)) does not resemble a sugarcane cycle. This result indicates
the need for a further inspection in the areas where these patterns occur.
On the second experiment, we used data extracted from MODIS2 images.
Each pixel in a MODIS image represents an area of 250 m x 250m, thus a higher
spatial resolution. MODIS images correspond to biweekly NDVI measures. We
used images of the São Paulo state, Brazil, corresponding to the period between
2
Moderate-Resolution Imaging Spectroradiometer.
 Discovering Frequent Patterns on Agrometeorological Data with TrieMotif 105

April 2004 and March 2005. Since MODIS images have a higher spatial resolution
than AVHRR/NOAA images, they allow a better analysis of certain regions. We
focused our analysis only on sugarcane crops regions and we extracted 40,133
time series of size 24. We ran the TrieMotif algorithm using bin discretization
with R = 1.5, w = 4, a = 4 and δ = 1. This time, we looked for the top 8-
Motifs of size n = 24, annual measures for the MODIS images time series. The
experiments were performed using the same computational infrastructure of the
synthetic data.

1 1 1 1

0.8 0.8 0.8 0.8

0.6 0.6 0.6 0.6

NDVI

NDVI

NDVI

NDVI
0.4 0.4 0.4 0.4

0.2 0.2 0.2 0.2

0 0 0 0
Apr Jun Aug Oct Dec Feb Apr Jun Aug Oct Dec Feb Apr Jun Aug Oct Dec Feb Apr Jun Aug Oct Dec Feb

(a) 1-Motif (b) 2-Motif (c) 3-Motif (d) 4-Motif

1 1 1 1

0.8 0.8 0.8 0.8

0.6 0.6 0.6 0.6

NDVI

NDVI

NDVI

NDVI
0.4 0.4 0.4 0.4

0.2 0.2 0.2 0.2

0 0 0 0
Apr Jun Aug Oct Dec Feb Apr Jun Aug Oct Dec Feb Apr Jun Aug Oct Dec Feb Apr Jun Aug Oct Dec Feb

(e) 5-Motif (f) 6-Motif (g) 7-Motif (h) 8-Motif

Fig. 19. Top 4-Motif for region of interest 1.

The results are shown on Fig. 19. As expected, the most frequent motif was
a sugarcane cycle pattern (Fig. 19(a)) with high values of NDVI on April and
low values on October. This time, the experts were also able to identify regions
with different harvest periods (Figs. 19(b), (d) and (e)). And once again, they
were able to spot regions with unexpected frequent patterns. Figure 19(f) shows
a region where there was a problem with the sugarcane crop on March and
Figs. 19(c), (g) and (h) show regions that might not be sugarcane regions.

6 Conclusions

Finding patterns in time series is highly relevant for applications where both the
antecedent and the consequent events are recorded as time series. This is the
case of climate and agrometeorological data, where it is known that the next
state of the atmosphere depends in large amount of its previous state. Today,
a large network of sensing devices, such as satellites recording both earth and
solar activities, as well as ground-based whether monitoring stations keep track
of climate evolution. However, the large amount of data and their diversity makes
it hard to discover complex patterns able to support more robust analyzes. In
this scenario, is very important to have powerful and fast algorithms to help
analyzing that data.
106 D.Y.T. Chino et al.

In this paper we presented the TrieMotif, a technique that provides, in an

integrated framework, an automated technique to extract frequent patterns in
time series as K-Motifs. It reduces the resolution of the data, speeding up the
sub-sequence comparison operations, indexes them in a trie structure and adopts
heuristics commonly employed by the meteorologists to prune from the similarity
search operations those branches that do not represent interesting answers. In
this way, our technique is able to select only candidate subsequences that have
high probability to match a motif, thus reducing the number of comparisons
performed.
Experiments performed over data from both synthetic and real applications
showed that our technique indeed perform in average 3 times faster them the
state of the art approach (Random Projection), including the best methods pre-
viously available. It also requires less memory, and the experiments revealed that
it requires up to 10 times less memory than the competitor methods. For a qual-
itative analysis, we presented the results to experts in the field (meteorologists),
whom confirmed that the results are indeed correct and useful for their day-to-
day activities to process climate data. For future works, we intend to explore
data from larger regions, as the whole Brazil and South America. We also intend
to explore data from different sensors in order to evaluate improvements that
may be needed on sugarcane crop regions.

Acknowledgements. The authors are grateful for the financial support granted by
FAPESP, CNPq, CAPES, SticAmsud and Embrapa Agricultural Informatics, Cepa-
gri/Unicamp and Agritempo for data.

References
1. Catalano, J., Armstrong, T., Oates, T.: Discovering patterns in real-valued time
series. In: Fürnkranz, J., Scheffer, T., Spiliopoulou, M. (eds.) PKDD 2006. LNCS
(LNAI), vol. 4213, pp. 462–469. Springer, Heidelberg (2006)
2. Chiu, B., Keogh, E., Lonardi, S.: Probabilistic discovery of time series motifs. In:
Proceedings of the Ninth ACM SIGKDD International Conference on Knowledge
Discovery and Data Mining, KDD 2003, pp. 493–498, New York, NY, USA. ACM
(2003)
3. Faloutsos, C., Ranganathan, M., Manolopoulos, Y., Manolopoulos, Y.: Fast subse-
quence matching in time-series databases. In: Proceedings of the ACM SIGMOD
International Conference on Management of Data, pp. 419–429. Minneapolis, USA
(1994)
4. Goldin, D.Q., Kanellakis, P.C., Kanellakis, P.C.: On similarity queries for time-
series data: Constraint specification and implementation. In: Proceedings of the 1st
International Conference on Principles and Practice of Constraint Programming,
pp. 137–153. Cassis, France (1995)
5. Keogh, E., Chakrabarti, K., Pazzani, M., Mehrotra, S.: Dimensionality reduction
for fast similarity search in large time series databases. Knowl. Inf. Syst. 3, 263–286
(2001)
6. Keogh, E., Kasetty, S.: On the need for time series data mining benchmarks: a
survey and empirical demonstration. Data Min. Knowl. Disc. 7, 349–371 (2003).
Springer
 Discovering Frequent Patterns on Agrometeorological Data with TrieMotif 107

7. Keogh, E., Lin, J., Lee, S.-H., Herle, H.: Finding the most unusual time series
subsequence: algorithms and applications. Knowl. Inf. Syst. 11(1), 1–27 (2007)
8. Li, Y., Lin, J.: Approximate variable-length time series motif discovery using gram-
mar inference. In: Proceedings of the Tenth International Workshop on Multimedia
Data Mining, MDMKDD 2010, pp. 10:1–10:9, New York, NY, USA. ACM (2010)
9. Li, Y., Lin, J., Oates, T.: Visualizing variable-length time series motifs. In: SDM,
pp. 895–906. SIAM / Omnipress (2012)
10. Lin, J., Keogh, E., Lonardi, S., Chiu, B.: A symbolic representation of time series,
with implications for streaming algorithms. In: Proceedings of the 8th ACM SIG-
MOD Workshop on Research Issues in Data Mining and Knowledge Discovery,
DMKD 2003, pp. 2–11, New York, NY, USA. ACM (2003)
11. Lin, J., Keogh, E., Patel, P., Lonardi, S.: Finding motifs in time series. In: The
2nd Workshop on Temporal Data Mining, at the 8th ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining, International Conference
on Knowledge Discovery and Data Mining, Edmonton, Alberta, Canada. ACM
(2002)
12. Lin, J., Keogh, E.J., Wei, L., Lonardi, S.: Experiencing sax: a novel symbolic
representation of time series. Data Min. Knowl. Disc. 15, 107–144 (2007)
13. Mohammad, Y., Nishida, T.: Constrained motif discovery in time series. New
Gener. Comput. 27(4), 319–346 (2009)
14. Rouse, J.W., Haas, R.H., Schell, J.A., Deering, D.W.: Monitoring vegetation sys-
tems in the great plains with ERTS. In: Proceedings of the Third ERTS Sympo-
sium, Washington, DC, USA, pp. 309–317 (1973)
15. Udechukwu, A., Barker, K., Alhajj, R.: Discovering all frequent trends in time
series. In: Proceedings of the winter international synposium on Information and
communication technologies, WISICT 2004, pp. 1–6. Trinity College Dublin (2004)
16. Wang, L., Chng, E.S., Li, H.: A tree-construction search approach for multivariate
time series motifs discovery. Pattern Recogn. Lett. 31(9), 869–875 (2010)
17. Yankov, D., Keogh, E., Medina, J., Chiu, B., Zordan, V.: Detecting time series
motifs under uniform scaling. In: Proceedings of the 13th ACM SIGKDD Inter-
national Conference on Knowledge Discovery and Data Mining, KDD 2007,
pp. 844–853. ACM, New York, NY, USA (2007)
 An Approach to the Discovery of Accurate
and Expressive Fix-Time Prediction Models

Francesco Folino(B) , Massimo Guarascio, and Luigi Pontieri

ICAR-CNR, National Research Council of Italy,

via P. Bucci 41C, 87036 Rende, CS, Italy
{ffolino,guarascio,pontieri}@icar.cnr.it

Abstract. Predicting the fix time (i.e. the time needed to eventually

solve a case) is a key task in an issue tracking system, which attracted
the attention of data-mining researchers in recent years. Traditional
approaches only try to forecast the overall fix time of a case when it
is reported, without updating this preliminary estimate as long as the
case evolves. Clearly, the actions performed on a case can help refine
the prediction of its (remaining) fix time, by using Process Mining tech-
niques, but typical issue tracking systems lack task-oriented descriptions
of the resolution process, and store fine-grain records, just registering case
attributes’ updates. Moreover, no general approach has been proposed
in the literature that fully supports the definition of high-quality derived
data, which were yet proven capable to improve prediction accuracy con-
siderably. A new fix-time prediction framework is presented here, along
with an associated system, both based on the combination of two kinds
of capabilities: (i) a series of modular and flexible data-transformation
mechanisms, for producing an enhanced process-oriented log view, and
(ii) several induction techniques, for extracting a prediction model from
such a view. Preliminary results, performed on the logs of two real issue
tracking scenarios, confirm the validity and practical usefulness of our
proposal.

Keywords: Data mining · Prediction · Business process analysis · Bug

tracking

1 Introduction
Issue tracking systems (a.k.a. “trouble/incident ticket” systems) are widely used
in real collaboration environments, in order to manage, maintain and help resolve
various issues arising within an organization or a community.
An important problem in such a context concerns the prediction of the fix
time, i.e. the time needed to solve a case. Several data-mining works [1,9,10]
recently approached this problem in the specific field of bug tracking systems, a
popular sub-class of issue tracking systems, devoted to support the notification
and fixing of bugs affecting some software artifact, within a software develop-
ment/maintenance community.
c Springer International Publishing Switzerland 2015
J. Cordeiro et al. (Eds.): ICEIS 2014, LNBIP 227, pp. 108–128, 2015.
DOI: 10.1007/978-3-319-22348-3 7
 An Approach to the Discovery of Accurate and Expressive Fix-Time 109

The approaches proposed so far mainly try to induce discrete (i.e. classification-
oriented) or continuous (i.e. regression-oriented) predictors from historical bug
logs, by resorting to classical propositional prediction methods, and simply regards
each bug case as a tuple, encoding all information available when the bug was ini-
tially reported, and labelled with a discrete or numerical (target) fix-time value.
Therefore, all the data recorded along the life of a bug/issue case (e.g., changes
made to properties like its priority, criticality, status, or assignee) are completely
disregarded, despite they could turn useful for refining, at run-time, the prediction
of (remaining) fix times.
The discovery of predictive (state-aware) process models [4,5,13] is a young
line of research in the field of Process Mining. Unfortunately, the approaches
developed in that field assume that log records be mapped to well-specified
process tasks, which are, instead, hardly defined in real bug/issue systems, where
the logs only store the sequence of changes made to a case’s attributes. In fact,
despite many systems support the design of workflows, these are rarely used in
real applications. Moreover, different repositories tend to exhibit heterogeneous
data schemes, even if developed over the same platform (e.g., Bugzilla).
In this work, we right attempt to overcome the limitations of current solu-
tions by proposing a novel approach to the discovery of (state-aware) fix-time
prediction models, out of low-level bug/issue logs, which can fully exploit all
information stored in the form of both case attributes and attribute modifica-
tion records. In order to reach a suitable abstraction level over case histories, a
modular set of parametric data-transformation methods is defined, which allow
to convert each case into a process trace (with update records abstracted into
high-level activities), and to possibly enrich these traces with derived/aggregated
data. In particular, the analyst is offered the possibility to effectively exploit tex-
tual descriptions, by summarizing them into a lower-dimensional concept space,
with the help of SVD-based techniques. A high-quality process-oriented view of
log data is so obtained, which can be analyzed with predictive process mining
methods, to eventually induce a predictive process model, capable of supporting
fix-time forecasts, at run-time, for future cases. The approach has been imple-
mented into a system, provided with an integrated and extensible set of data-
transformation and predictive learning tools, along with a series of functionalities
for browsing, evaluating and analysing the discovered process models.
The remainder of the paper is organized as follows. After discussing rele-
vant related works in Sect. 2, we introduce some basic concepts and notation in
Sect. 3. Several core event manipulation methods and trace manipulation ones
are presented in Sects. 4 and 5, respectively. Section 6 illustrates the whole app-
roach to the discovery of predictive fix-time models, while Sect. 7 its current
implementation. We finally discuss a series of tests in Sect. 8, and draw a few
concluding remarks in Sect. 9.
110 F. Folino et al.

2 Related Work
As mentioned above, previous fix-time prediction works focus on the case of
bug tracking systems, and resort to classical learning methods, conceived for
analyzing propositional data labelled with a discrete or numerical target.
Random-forest classifiers and linear regressors were trained in [9] and in [1],
respectively, in order to predict bug lifetimes, using different bug attributes
as input variables. Different standard classification algorithms were employed
instead in [10] to the same purpose. Decision trees were also exploited in [7], in
order to estimate how promptly a new bug report will receive attention. Stan-
dard linear regression was also used in [8] to predict whether a bug report will
be triaged within a given amount of time.
As mentioned above, none of these approaches explored the possibly of imp-
roving such a preliminary estimate subsequently, as long as a case undergoes
different treatments and modifications. The only (partial) exception is the work
in [10], where some information gathered after the creation of a bug is used
as well, but just for the special case of unconfirmed bugs, and till they were
definitely accepted. Conversely, we want to exploit the rich amount of log data
stored for bug/issue cases (across their entire life), in order to build a history-
aware prediction model, capable of providing accurate run-time forecasts for the
remaining fix time of new (unfinished) cases.
Predicting processing times is the goal of an emerging research stream in
the field of Process Mining, which specifically addresses the induction of state-
aware performance model out of historical log traces. In particular, the discovery
of an annotated finite-state model (AFSM) was proposed in [13], where the
states correspond to abstract representations of log traces, and store processing-
time estimates. This learning approach was combined in [4,5] with a predictive
clustering scheme, where the initial data values of each log trace are used as
descriptive features for the clustering, and its associated processing times as
target features. By reusing existing induction methods, each discovered cluster
is then equipped with a distinct prediction model – precisely, an AFSM in [4],
and classical regression models in [5].
Unfortunately, these Process Mining techniques rely on a process-oriented
representation of system logs, where each event refers to a well-specified task;
conversely, common bug tracking systems just register bug attribute updates,
with no link to resolution tasks. In order to overcome this limitation, we want
to help the analyst extract high-level activities out of bug/issue histories, by
providing her/him with a set of data transformation methods, tailored to fine-
grain attribute-update records, like those stored in typical bug/issue tracking
systems.
The effectiveness of derived data in improving fix-time predictions was pointed
out in [2], where a few summary statistics and derived properties were computed
for certain Bugzilla repositories, in a pre-processing phase. We attempt to gen-
eralize such an approach, by devising an extensible set of data transformation
and data aggregation/abstraction mechanisms, allowing to extract and evaluate
such derived features from a generic bug/issue resolution trace.
 An Approach to the Discovery of Accurate and Expressive Fix-Time 111

3 Basic Concepts and Notation

We next illustrate the structure and main characteristics of issue/bug resolution
logs, which are the source for extracting a process-oriented fix-time prediction
model.
For the sake of concreteness, let us focus on the structure of bug reposito-
ries developed with Bugzilla (http://www.bugzilla.org), a general-purpose bug-
tracking platform, devoted to support people in various bug-related tasks – e.g.,
keep track of bugs, communicate with colleagues, submit/review patches, and
manage quality assurance (QA). We pinpoint that this choice does not limit
the generality of our approach, since similar strategies take place in most real
bug/issue -tracking applications.
Usually bug/issue resolution tasks are carried out in an unstructured fashion,
without being enforced by a prescriptive process model. Such applications mainly
looks like a repository, where an extensible set of attributes are kept for a case
(concerning the resolution of a bug/issue), and possibly updated along the entire
life of that case.
For example, in the Bugzilla repository of project Eclipse (also used in our
tests), the main attributes associated with each bug b are: the status and
resolution of b; who entered b into the system (reporter); the last solver b was
assigned to (assignee); the component and product affected by b; b’s severity’s
and priority’s levels; a milestone; the list of users to be kept informed on b’s
progress (CC); the lists of other bugs depending on b (dependsOn), and of related
documents (seeAlso).
Some attributes (e.g., reporter) are static, whereas others (e.g., assignee,
status, resolution) may change while a case evolves. In particular, the status
of a bug b may take the following values: unconfirmed (i.e. b was reported by an
external user, and it must be confirmed by a project member), new (i.e. b was
opened/confirmed by a project member), assigned, resolved (i.e. a fix was made
to b, but it needs to be validated), verified (i.e. a QA manager has validated
the fix), reopened (if the last fix was insufficient), and closed. For a resolved bug
b, the resolution field may take one of these values: fixed, duplicate (i.e. b is
a duplicate of another bug), works-for-me (i.e. b has been judged unfounded),
invalid, won’t-fix.
In any Bugzilla repository, the whole history of a bug is stored as a list of
update records, all sharing the same structure, which consists five predefined
fields (in addition to a bug identifier): who (the person who made the update),
when (a timestamp for the record), what (the attribute modified), removed (the
former value of that attribute) and added (the newly assigned value).
Figure 1 reports, as an example, all the update records of a single Ecplise’s bug.

Traces and Associated Data. The contents of a bug/issue tracking repository

can be viewed as a set of traces, each storing the sequence of events recorded
during a case. As explained above, each of these events concerns a modification
to a case attribute, and takes the form of the records in Fig. 1.
112 F. Folino et al.

Fig. 1. Activity log for a single Bugzilla’s bug (whose ID is omitted for brevity).
Row groups gather “simultaneous” update records (sharing the same timestamp and
executor).

Let E be the universe of all possible events, and T be the universe of all
possible traces defined over them. For any event e ∈ E, let who(e), when(e),
what(e), removed(e), and added(e) be the executor, the timestamp, the attribute
modified, the former value and new value stored in e, respectively. In the case
of Bugzilla repositories, each of these dimensions just corresponds to a label.
This does not happen, however, in other issue tracking systems, where an event
record can encode a change to multiple attributes, and gather different kinds
of information on the organizational resource involved (e.g., her/his role, team,
department, etc.). In such a case, the representation of log events can be gener-
alized, by letting who(e), when(e), what(e), removed(e), and added(e) take the
form of a tuple (over some suitable attribute space).
For each trace τ ∈ T , let len(τ ) be the number of events in τ , τ [i] be the i-th
event of τ , for any i = 1 .. len(τ ), and τ (i] ∈ T be the prefix trace gathering the
first i events in τ .
Prefix traces have the same form as fully unfolded ones, but only represent
partial case histories. Indeed, the prefix traces of a case depicts the evolution of
the case across its whole life. For example, the activity log of Fig. 1 (which just
stores the history of one bug) will be represented as a trace τex consisting of 12
events, one for each of the update records (i.e. rows of the table) in the figure;
in particular, for the first event, it is who(τex [1]) = svihovec, what(τex [1]) = CC,
added(τex [1])) = {margolis, svihovec}.
As mentioned above, typical tracking systems store several attributes for each
case (e.g., reporter, priority, etc.), which may take different values during its
life. Let F1 , . . . , Fn be all of the attributes defined for a bug. Then, for any (either
partial of completed) trace τ , let data(τ ) be a tuple storing the updated values
of these attributes associated with τ (i.e. the values taken by the corresponding
case after the last event of τ ), and data(τ )[Fi ] be the value taken by Fi (for
i = 1..n).
Finally, a log L is a finite subset of T , while the prefix set of L, denoted by
P(L), is the set of all possible prefix traces that can be extracted from L.
 An Approach to the Discovery of Accurate and Expressive Fix-Time 113

Fix-time Measurements and Models. Let μ̂F : T → R be an unknown function

assigning a fix-time value to any (sub-)trace. The value of μ̂F is known over
all P(L)’s traces, for any given log L; indeed, for any trace τ and prefix τ (i],
it is μ̂F (τ (i]) = when(τ [len(τ )]) − when(τ [i]). For example, for the trace τex (2]
encoding the first 2 events in Fig. 1, it is μ̂F (τex (2]) = when(τ [12])−when(τ [2]) =
197 days.
A Fix-time Prediction Model (FTPM) is a model for estimating the remaining
fix time of a case, based on its current trace. Learning such a model, which
approximates the unknown function μˆF , is an induction task, where log L plays
as the training set, and the value μ̂F (τ ) of the target measure is known for each
(prefix) trace τ ∈ P(L).

4 Event Manipulation Operators

In the discovery of an F T P M model we want to consider both case histories (i.e.
all sequences of update records) and case properties (e.g., the affected product,
severity level, reporter). To this end, we will regard some of the actions performed
on a case as a clue for the activities of an unknown (bug/issue resolution) process,
in order to eventually exploit Process Mining techniques. In fact, simply defining
such activities as all possible changes to the status of a case would lead to
discarding relevant events, such as the (re-)assignment of the case to a solver,
or the update of key properties (like its severity, criticality, or category). On the
other hand, we do not either want to look at all attribute updates as resolution
tasks, since many of them are hardly linked to fix times, and they may even
produce a noise-like effect on the discovery of fix-time predictors.

4.1 Activity-Oriented Event Abstraction

An event abstraction function α is a function mapping each event e ∈ E to an
abstract representation α(e), which captures relevant facets of the action per-
formed. In current process mining approaches, log events are usually abstracted
into their associated tasks, possibly combined with other properties of them
(e.g., their executors), under the assumption that the events correspond to the
execution of work-items, according to a workflow-oriented view of the process
analyzed.
In our framework, such a function α is intended to turn each issue/bug -
tracking event into a high-level bug/issue-resolution activity, by mapping the
former to a label that captures well its meaning. Since only attribute-update
events are tracked, the analyst is allowed to define this function in terms of their
fields (i.e. who, when, what, added, and removed).
The default instantiation of α, denoted by α and conceived for Bugzilla’s
records, is defined as follows (with + denoting string concatenation):


what(e)+ “:=”+added(e), if what(e) ∈ {status, resolution}
α(e) = (1)
“Δ”+what(e), otherwise
114 F. Folino et al.

This particular definition of α focuses on what attribute was modified, while

abstracting any other event’s field (namely, who, when, removed, and added); as
an exception, the assigned value is included in the abstract representation when
the update concerns the status or resolution, as such information, character-
izing the state of a bug, can improve fix-time predictions. For example, for the
first two events of trace τex (gathering all the records in Fig. 1), it is α(τex [1]) =
ΔCC, and α(τex [2]) = ΔTargetMilestone, while the activity label of the last
event (τex [12]) is status:=closed.
Different event abstraction functions can be defined by the analyst, in order to
focus on other facets of bug activities, or to change the level of detail, depending
on the specific bug attributes (and associated domains) available in the applica-
tion scenario at hand. For instance, with regard to the scenario of Sect. 3, one may
refine the representation of severity-level changes by defining two distinct activ-
ity labels for them, say ΔSeverity-Eclipse and ΔSeverity-N otEclipse, based
on the presence of substring “eclipse” in the e-mail address of the person who
made the change.

4.2 Macro-Events’ Restructuring

In real bug tracking environments, multiple fields of a case are often modified
in a single access session, and the corresponding activity records are all stored
with the same timestamp, in a rather arbitrary order. For example, in our tests,
we encountered many cases where the closure of the bug (i.e. an event of type
status:=closed ) preceded a “contemporaneous” change to the assignee (or a
message dispatch).
Regarding each set of contemporaneous events as one macro-event, the ana-
lyst can define three kinds of data-manipulation rules, in order to rearrange them
based on their fields: (i) a predominance rule, assigning different relevance levels
to simultaneous events (with the ultimate aim of purging off less relevant ones);
(ii) a set of merging rules, indicating when two or more contemporaneous events
(with the same predominance level) must be merged together, and which activity
label must be assigned to the resulting aggregated event; (iii) a set of sort rules,
specifying an ordering relation over (non-purged and non-merged) simultaneous
events.
Any combination of the above kinds of rules will be collectively regarded,
hereinafter, as a macro-event criterion. The default instantiation of this crite-
rion, tailored to Bugzilla’s bugs, is summarized in Table 1, where each event is
given a “predominance” level, only based on what attribute was updated in the
event. Such levels acts as a sort of priority score in the selection of events (the
lower the level, the greater the priority): an event is eventually kept only if there
is no simultaneous event with a lower level than it. In particular, events involv-
ing a change to the status or resolution hide assignee/priority/severity
updates, which, in their turn, hide changes to the milestone or CC.
A merging rule is defined in Table 1 only for 1-level simultaneous events,
which states that, whenever the status and resolution of a bug are modified
contemporarily, the respective events must be merged into a single macro-event,
 An Approach to the Discovery of Accurate and Expressive Fix-Time 115

which is labelled with the concatenation of their associated activity labels. For
example, this implies that the ninth and tenth events in Fig. 1 will be merged
together, and labelled with the string “state:=resolved + resolution:=fixed ”.
Also the default sort rule (shown in the table as an ordering relation <) only
depends on the what field, and states that events involving attribute milestone
must precede those concerning CC, and that priority (resp., severity) updates
must precede severity (resp., assignee) ones. In this way, e.g., the first two
events in Fig. 1 will be switched with one another.

Table 1. Default macro-event criterion: predominance, merging and sort rules over
simultaneous events. No merging rules for levels 2 and 3, and no sort rules for level-1
events are defined.

Predominance levels Merging rules Sort rules

(lev.) (bug attributes) (macro-event activity label) (ordering relation)
1 status, resolution α( , , status, , )+ “+” —
+ α( , , resolution, , )
2 priority, severity, — priority<severity<assignee
assignee
3 milestone, CC — milestone<CC

Example 1. Let us apply all default log abstraction operators introduced above
(i.e. the event abstraction function in Eq. 1 and the macro-event criterion of
Table 1) to the bug trace τex encoding the events in Fig. 1. For the sake of concise-
ness, we only consider events involving the attributes in Table 1. The resulting

trace τex consists of 8 events, which are associated with the following activ-
ity labels, respectively: l1 =“Δmilestone”, l2 =“ΔCC”, l3 = “Δassignee”, l4 =
“ΔCC”, l5 = “ΔCC”, l6 = “Δassignee”, l7 = “status:=resolved +resolution:=

fixed ”, l8 = “status := closed ”, where li denotes the activity label of τex [i], for
i ∈ {1, . . . , 8}. The respective timestamps (at 1-hour granularity) of these events
are: t1 = t2 =(2012-06-20 10EDT ), t3 =(2012-06-20 22EDT ), t4 =(2012-06-21
11EDT ), t5 =(2012-06-25 05EDT ), t6 =(2012-07-02 10EDT ), t7 =(2012-07-02
15EDT ), t8 =(2013-01-03 11EST ). 

5 Trace Manipulation Operators

By applying the event manipulation operators introduced in the previous section,
each bug history can be turned into a “process trace”, i.e. a sequence of resolution
tasks. In order to profitably apply process-oriented prediction techniques, such
a trace must be converted into some abstract state-oriented form, capturing
somewhat the stage reached by the corresponding bug case along the resolution
process. On the other hand, the accuracy of fix-time prediction can be empowered
by taking advantage of context information stored for each case, in the form of
native or derived trace attributes. These two issues are discussed in details next,
in two separate subsections.
116 F. Folino et al.

5.1 State-Oriented Trace Abstraction

For each trace τ , a collection of relevant prefixes (i.e. sub-traces) rp(τ ) is selected,
in order to extract an abstract representation for the states traversed by the
associated bug, during its life. Two strategies can be adopted to this end, named
event-oriented and block-oriented. In the former strategy all possible τ ’s prefixes
are considered, i.e. rp(τ ) = {τ (i] | i = 1 . . . len(τ )}, whereas in the latter only
prefixes ending with the last event of a “macro-activity” are selected, i.e. rp(τ ) =
{ τ (i] | 1 ≤ i ≤ len(τ ) and when(τ [j]) > when(τ [i]) ∀j ∈ {i + 1, . . . , len(τ )} }.
Independently of the selection strategy, each trace τ  in rp(τ ) is turned into
a tuple stateα (τ  ), whose attributes are all the abstract activities produced by a
given event abstraction function α (e.g., that in Eq. 1). The value taken by each
of these activities, say a, is denoted by state(τ  )α [a] and computed as follows:

stateα (τ  )[a]= SUM({ δ(τ  [i]) | α(τ  [i]) = a for i = 1..len(τ  )}) (2)

where δ is a function assigning an integer weight to each event, based on its

properties; by default it is (i) δ(e) = |added(e)|, if e is not an aggregation of
multiple simultaneous events (i.e. it corresponds to one raw update record) and
e involves a multivalued attribute (like CC, seeAlso), or (ii) δ(e) = 1 otherwise.
Any prefix trace τ  is hence encoded by an integer vector in the space of the
abstract activities extracted by α, where each component accounts for all the
occurrences, in τ  , of the corresponding activity. Such a vector captures the state
of a bug (at any step of its evolution) through a summarized view of its history.

Example 2. Let us consider the trace τex shown in Example 1. The unfolding of
 
this trace gives rise to 8 distinct prefix sub-traces, denoted by τex (1], τex (2], . . .,

τex (8]. Five abstract activities occur in these traces:
a1 =“Δmilestone”,
a2 =“ΔCC”,
a3 =“Δassignee”,
a4 =“status:=resolved +resolution:=fixed ”,
a5 =“status :=closed ”.
As concerns trace abstractions, all components of stateα (τex 
(1]) (i.e. the tuple
encoding the state reached after the first step) are 0 but that associated with a1
and a2 , which are stateα (τex

(1])[a1 ] = 1, and stateα (τex

(1])[a2 ] = 2. If using the
event-oriented strategy, the above traces will generate 8 state tuples:
stateα (τex

(1])=1, 2, 0, 0, 0, stateα (τex

(2])=1, 2, 1, 0, 0,
α  α 
state (τex (3])=1, 4, 1, 0, 0, state (τex (4])=1, 5, 1, 0, 0,
stateα (τex

(5])=1, 5, 2, 0, 0, stateα (τex

(6])=1, 5, 2, 1, 0,
α 
state (τex (7])=1, 5, 2, 1, 1. 
Such a state-oriented representation of a log L will be eventually exploited
to induce a fix-time prediction model (i.e. a F T P M ) for L, as explained in the
next section.
 An Approach to the Discovery of Accurate and Expressive Fix-Time 117

5.2 Insertion of Additional Trace Attributes

In order to possibly improve the accuracy of discovered prediction models, we

want to give the analyst the possibility to inject additional attributes, accounting
for the context where resolution tasks took place.
More specifically, our framework is meant to supports different kinds of oper-
ators for inserting additional data into bug traces, or for transforming a given
(raw or derived) bugs/events attribute, and making it more expressive: (i) indica-
tors’ derivation operators, (ii) attribute enrichment operators, and (iii) attribute
summarization operators.

Indicators’ Derivation. The first kind of operators is essentially meant to com-

pute some statistics over some suitable trace collection. In fact, the insertion of
such additional data was already considered in previous bug analysis works [8,9].
We defined five standard indicator derivation operators, which allow for auto-
matically inserting the following fields into each given trace τ , respectively:
(i) a workload measure indicating the overall number of cases currently opened
in the system globally, and several variants of it, computed over specific groups
of cases (e.g., all bugs regarding the same product version, component or OS as
τ ); (ii) an analogous collection of counters for the cases fixed in the past year
(globally, and for the version/component/OS associated with τ ); (iii) a “rep-
utation” coefficient, computed for the reporter of τ as in [2]; (iv) the average
fix time for various groups of related cases (e.g., all bugs concerning the same
project or reporter as τ ) and closed in the past year; (v) several seasonality
dimensions (such as, week-day, month, and year) derived from the date of the
last event in τ .

Attribute Enrichment. This class of operators is meant to extend the values of an

attribute with correlated information, extracted from the same repository. As a
standard attribute enrichment mechanism, we consider the possibility of replac-
ing users’ identifiers (possibly appearing, e.g., in who fields of bug history records,
or in certain bug attributes) with their respective e-mail addresses — actually,
no further information on people is available in many real bug repositories. To
implement such an operator, a greedy matching procedure was developed, which
founds on comparing any user ID with all the email addresses appearing in var-
ious attributes of the bugs.

Clustering-based Attribute Summarization. These operators aim at reducing the

dimensionality of an attribute by partitioning its domain into classes. In partic-
ular, in order to get an effective level of abstraction over a trace attribute, say F ,
we considered the problem of finding a partition of F ’s domain exhibiting high
correlation with fix-time measurements, based on a given aggregation hierarchy
H over F . Let dom(F ) be the domain of F , and h ∈ N be the height (i.e. the
number of levels) of H. Let also H(i) be the i-th level of H (for i = 0, l . . . , h),
and values(H, i) be the values forming that aggregation level. Each level of H
corresponds to a partition of dom(F ), with a distinguished aggregation degree;
118 F. Folino et al.

in particular, level 0 (H’s leaves) just encodes the raw values in dom(F ) without
any aggregation, and level h just consists of one node (H’s root). Such a hierar-
chy may be already available for the attribute (this is the case of email addresses
and software products, which follow an implicit meronomical and taxonomical
scheme, respectively), or it can be computed automatically through clustering
methods. Let CH be a cut of the given hierarchy H, and values(CH ) be the
associated nodes, i.e. the set of aggregated F ’s values identified by CH . Any
such a cut can be associated with a loss measure, denoted by Loss(CH ), w.r.t.
the target metrics μF , as follows:
 2
τ ∈P(L) [μ̂F (τ [len(τ )]) − μ̃F (τ, CH )]
Loss(CH ) =
|P(L)|
where μ̃F (τ, CH ) is average performance value over all trace prefixes owning
a value of F that was assigned to the same CH ’s node as τ . An optimal cut
of H is one minimizing this loss measure, among all those satisfying a user-
defined property. A greedy procedure has been implemented in our framework
for computing such a cut, based on a maximal constraint on cut’s size (i.e.
number of nodes in the cut).

Text-oriented Attribute Summarization. In many real application scenarios, bug/

issue handling systems store free-text contents for each case, ranging from notes
to messages. In particular, in the case of Bugzilla’s bug repositories, a summary
field is kept for each bug case, providing a short description of the problem
occurred. Let us assume that the textual contents of any bug τ are stored into
a vector text(τ ), encoding a bag-of-word representation for the concatenation of
all τ ’s free-text fields, based on classic vector-space model and TF-IDF weights.
Clearly, such a textual description can hold important information on the nature
and characteristics of a bug, and it can help improve the estimation of its fix
time, as pointed out in [6]. However, as the vocabulary used in such descriptions
may be rather large and sparse, extending the attributes of each trace with a
mere bag-of-word view of its text contents will hardly improve the accuracy of
fix-time predictions; indeed, such an overly detailed representation might even
confuse the learner and lead to overfitting and imprecise models.
In order to face this issue, we introduced a further kind of text-oriented
attribute-summarization operators, meant to grasp a concise high-level repre-
sentation of text contents. The standard implementation of this operator class
relies on computing a low-rank approximation of the term-by-bug matrix, based
on Singular Value Decomposition (SVD) techniques [12]. The mapping of cases
to the underlying concept space (i.e., the U matrix in the U × Σ × V T factor-
ization of the term-by-bug matrix) is used to build a low-dimensional (and more
abstract) representation of text contents. Such summarized “latent-semantics”
vectors will be used to complement the attributes of each trace, so obtaining a
richer description of the corresponding bug/issue case.
 An Approach to the Discovery of Accurate and Expressive Fix-Time 119

6 F T P M Discovery Algorithm
This section is meant to offer a summarized view of our whole approach to the
discovery of a Fix-time Prediction Model (F T P M ), based on a given set of raw
bug records.
The approach is illustrated in Fig. 2 as a meta-algorithm, named FTPM
Discovery, encoding the main steps of our (process-oriented) data-transformation
methodology, and the eventual application of a predictive induction method to the
transformed log.
The algorithm takes as input a bug/issue tracking repository, storing a col-
lection of records (like those at the beginning of Sect. 3), and some parameters
regarding the application of data-manipulation operators. In order to apply the
abstraction operators introduced in Sect. 4, these data are first turned into a set
of traces (i.e. a log).
Based on a given filtering criterion Φ, function filterEvents possibly removes
uninteresting events (e.g., outliers or noisy data), which may confuse the learner.
Function handleMacroEvents allows us to apply a given macro-event crite-
rion Γ (like that described in Table 1) to rearrange each group of simultaneous
log events according to the associated predominance, reordering and/or merging
rules.
Steps 4 and 5 are meant to possibly associate each trace τ with additional
“derived” data, to complement the original contents of data(τ ) with context
information.

Fig. 2. Meta-algorithm FTPM Discovery.

120 F. Folino et al.

Basically, function deriveTraceAttributes is devoted to insert new trace

attributes, with the help of indicators’ derivation operators and of attribute
enrichment operators like those defined in the previous section.
Conversely, function refineTraceAttributes is meant to transform a num-
ber of (raw or derived) cases’/events’ attributes, and eventually them into more
expressive ones, by way of attribute-summarization operators.
Steps 6–10 are simply meant to extract a set RS of relevant (sub-)traces out of
P(L), based on the chosen event selection strategy S. RS is then used by function
mineFTPM as a training set, in order to eventually induce an F T P M . To this
end, as explained in Sect. 4, each trace τ ∈ RS is converted into a tuple labelled
with the fix-time measurement μF (τ ), and encoding both the representation
of τ ’s state (w.r.t. the given event abstraction function α), and its associated
(augmented) data tuple data(τ ). More precisely, data(τ ) and stateα (τ ) are used
as descriptive/input attributes, while regarding the actual remaining-time value
μF (τ  ) as the target of prediction.
At this point, various learning methods (including those mentioned in Sect. 2)
can be used to induce a regression or classification model. Different solutions
for carrying out this task have been implemented in our prototype system, as
described next.

7 The Implemented System

A prototype system was developed to fully implement the approach above, and
support the discovery and analysis of F T P M s. The system allows the analyst
to flexibly and modularly apply all the data-processing operators presented in
this work, in an interactive and iterative way, as well as to define and store (in a
reusable form) new variants of them, according to a template-based paradigm.
The conceptual architecture of the system is shown in Fig. 3. System’s func-
tionalities are organized in three layers, which concern respectively: the storage,
handling and manipulation of log data (Data Management and Processing), the
implementation of core data mining techniques (Data Mining), and the handling
and evaluation the prediction and clustering models found with these techniques
(Models’ Management).
The Log Repository Gateway module supports the importation of historical
data stored in a remote bug/issue tracking repository. Its current implementation
allows for connecting to the web interface of any Bugzilla repository, and con-
verting the histories of selected bugs into bug traces. The imported log, possibly
cleaned according to suitable filtering rules, is stored in the Event Log archive.
The Event Manipulation and Trace Manipulation modules provide various
data-transformation mechanisms (possibly already stored in the form of Event/
Trace Transformation Rules), in order to produce enhanced views of the log,
which are more suitable for fix time prediction. More specifically, the first mod-
ule is responsible for manipulating groups of simultaneous events according to a
desired macro-event criterion (sub-module Macro-events’ manipulation), as dis-
cussed in Sect. 4, and for extracting high-level activity labels from event prop-
 An Approach to the Discovery of Accurate and Expressive Fix-Time 121

Fig. 3. Logical architecture of the implemented system.

erties (sub-module Event abstraction). On the other hand, the Trace Manip-
ulation module allows to build a state-oriented abstraction for each selected
trace (sub-module State-oriented encoding), while possibly enriching each trace
with additional (derived and/or summarized) attributes (sub-module Attribute
derivation), as explained in Sect. 5.2. In particular, one can extract high-level
trace attributes capturing the semantics of text case fields, based on an SVD
analysis of the term-by-case matrix. This task leverages the Text Mining module
(from the Data Mining layer), implementing core text processing functionalities
(including tokenization, stemming, TF-IDF encoding), and SVD calculations.
Each “refined” log view, obtained with of the above described functionalities.
can be used as input for module Predictive Models’ Mining. To this end, the
abstracted traces are delivered to either the Regression or Classification sub-
module, depending on which learning task was chosen by the user. Preliminary
to the induction of a prediction model, the traces can be possibly partitioned
into groups by one of the predictive clustering methods implemented by the
Predictive Clustering module, which will also produce a set of decision rules for
discriminating among the discovered clusters. In this case, each cluster will be
eventually equipped with a distinct fix-time predictor.
All clustering and predictive models discovered can be explored and stored for
subsequent analyses. The Models’ Management layer also supports the evalua-
tion of prediction models via several accuracy metrics (e.g., rmse, mae, precision,
recall ). Moreover, the Aggregated views module can compute aggregate statis-
tics and graphics, concerning the distribution of any trace attribute in any trace
cluster (like that in Fig. 4).

Regression Algorithms. Several alternative learning methods are currently imple-

mented in our system, which support the induction of a F T P M , from a propo-
sitional training set like that described in the previous section. These methods,
122 F. Folino et al.

ranging from classical regression methods to state-aware Process Mining ones

are listed next:

– IBK, a lazy (case-based) naı̈ve regression method available in Weka [12], used
with k = 1, Euclidean distance, and nominal attributes’s binarization;
– RepTree, implementing the homonymous regression-tree method [12], while
using variance reduction and 4-fold reduced-error pruning;
– CATP, implementing the approach in [4], which first partitions the traces into
clusters, via a multi-target predictive clustering strategy, and then equips each
cluster with an “A-FSM” predictor, as in [13] (using no history horizon and
the bag-oriented trace abstraction);
– AATP-IBK and AATP-RepTree [5], which still adopt a multi-target predictive
clustering scheme for partitioning fully-unfolded traces, and regressors IBK
and RepTree, respectively, as the base learner for each discovered cluster.

Classification Algorithms. Like in previous bug analysis works, the analyst can
also induce a classification model for the prediction of (discrete) fix times, after
defining a set of a-priori classes in terms of fix-time ranges (possibly with the help
of automated binning tools). To this end, a number of existing classifier-induction
algorithms can be exploited, including popular decision-tree learning algorithm
J48 (with 3-fold reduced error pruning), and the Random Forest Algorithm [3]
(with size 10).

8 Experimental Evaluation

We performed a series of tests with our prototype system against two real
datasets: a log from project Eclipse’s Bugzilla repository, and a collection of
traces from a problem management system BPI Challenge [11]. In all tests,
FTPM Discovery was used to find different numeric fix-time predictors, by vary-
ing the underlying regression algorithm, combined with the event-based prefix
selection strategy (i.e. S = EVENT).
The accuracy of the discovered models was evaluated with the standard error
metrics root mean squared error (rmse) and mean absolute error (mae) — both
computed via 10-fold cross validation, and normalized by the average fix time
(i.e. 59 and 179 days for bug cases and problem management cases, respectively),
for ease of interpretation.

8.1 Experiments on Bug Logs of Project Eclipse

A sample of 3906 bug records (gathered from January 2012 to March 2013) was
turned into a set of bug trace like those described in Sect. 3. The length of full
bug traces ranged from 2 to 27, while bug fix time ranged from one day (i.e.,
a bug is opened and closed in the same day) to 420 days, with an average of
about 59 days.
 An Approach to the Discovery of Accurate and Expressive Fix-Time 123

Log Manipulation Steps and Enhanced Log Views. In order to put this log into
a more suitable for prediction, we applied some of the log manipulation mecha-
nisms defined in our framework. Five views of the original log, named L0 , . . . , L4 ,
were produced by incrementally applying the basic event abstraction function
α in Eq. 1 and all the data-processing functions appearing in algorithm FTPM
Discovery (cf. Figure 2).

– The basic “abstracted” view L0 was obtained by only applying the event
abstraction function α to the selected bug traces;
– A “cleaned” view L1 , consisting of 2283 traces, was produced making L0 undergo
a specific instantiation of function filterEvents, removing the following data:
(i) bugs never fixed, (ii) “trivial” bug cases (i.e. all bugs opened and closed
in the same day), and (iii) trace attributes (e.g., version, whiteboard, and
milestone) featuring many missing values, and bug/event fields (e.g. summary)
containing long texts.
– View L2 was produced by restructuring the simultaneous events in L1 with the
default implementation of function hanldeMacroEvents (based on the rules
of Table 1).
– View L3 was obtained by applying all the derivation mechanisms described in
Sect. 5. As a result, each trace is equipped with the new attributes: week-day,
month, and year; indicators workload, workload version, workload compo-
nent, workload OS (i.e. the number of bugs open in the system in general,
an for the same product version, component and operating system, resp.);
reporter avgFT, project avgFT (average fix-time for all past years’ bugs
with the same reporter or project, resp.); coefficient reporter reputation [2].
– View L4 was obtained by the built-in implementation of abstractTrace-
Attributes, applied to L3 . In particular, all people identifiers in the reporter
bug attribute were replaced with a number of reporters’ groups representing
different organizational units (namely, {oracle, ibm.us, ibm.no us, vmware,
other}), and extracted semi-automatically from e-mail addresses. A simi-
lar approach was used to produce a binary abstraction (namely {eclipse,
not eclipse}) of attribute assignee, and an aggregate representation of both
product and component. Each bug was finally equipped with a series of
attributes encoding a low-dimensional representation of its textual contents,
computed with the help of the SVD-based technique described before.

Prediction Results. Table 2 reports the normalized rmse and mae errors obtained,
with different regression methods available in our system, on the five log views
described above. Two different learning setting were considered to this end: using
bug history information (originally registered in terms of attribute-update events),
and neglecting it. Note that the latter setting is intended to provide the reader with
a sort of baseline, mimicking the approach followed by previous fix-time prediction
works.
In general, it is easily seen that results obtained in the “no bug history”
setting — where only the initial data of reported bugs are used as input variables
124 F. Folino et al.

Table 2. Regression results on Eclipse bugs. Rows correspond to different F T P M

induction methods, tested in two learning settings: without and with bug history events.
Columns L0 , . . . , L4 correspond to different views of original data, obtained with our
log manipulation operators.

Predictors rmse mae

Setting Methods L0 L1 L2 L3 L4 L0 L1 L2 L3 L4
No Bug History IBK 1.051 1.051 1.050 1.092 1.093 0.569 0.569 0.561 0.583 0.584
(Baseline)
RepTree 0.973 0.973 0.970 0.966 0.925 0.562 0.562 0.552 0.547 0.546
Avg (no history) 0.973 0.973 0.970 0.966 0.925 0.562 0.562 0.552 0.547 0.546
History-aware CATP 0.967 0.873 0.880 0.737 0.640 0.510 0.467 0.440 0.380 0.316
IBK 0.983 0.823 0.807 0.793 0.803 0.430 0.360 0.360 0.347 0.345
AATP-IBK 1.003 1.007 0.827 0.800 0.710 0.437 0.473 0.367 0.353 0.305
RepTree 1.013 0.883 0.907 0.910 0.773 0.533 0.473 0.473 0.477 0.351
AATP-RepTree 0.970 0.930 0.887 0.783 0.657 0.510 0.530 0.437 0.390 0.306
Avg (history- 0.987 0.903 0.862 0.8059 0.717 0.484 0.461 0.415 0.389 0.325
aware)

for the prediction — are rather poor, if compared to the average ones obtained in
the “history-aware” setting. Indeed, the errors measured in the former setting are
rather high, no matter which inductive method (i.e. IBK or RepTree) is used, and
which pre-processing operations are applied to the original logs. Interestingly,
this result substantiates our claim that exploiting bug activity information helps
improve the precision of fix-time forecasts.
On the other hand, in the second setting, both rmse and mae errors tend to
decrease when using more refined log views. In particular, substantial reductions
were obtained with the progressive introduction of macro-event manipulations
(view L2 ), and of derived and abstracted data (views L3 and L4 , respectively).
Very good results are obtained (in the history-aware setting) when using
some kind of predictive clustering method, be it single-target (RepTree) or multi-
target (CATP, AATP-RepTree and AATP-IBK). However, trace-centered clustering
approaches (namely, CATP, AATP-RepTree, AATP-IBK achieve better results than
RepTree, which considers all possible trace prefixes for the clustering. The benefit
of using a clustering procedure is quite evident in the case of IBK, which generally
gets worse achievements than any other approach, presumably due to its inability
to fully exploit derived data. Indeed, still focusing on the history-aware setting,
we note that IBK gets more precise when embedded in the predictive clustering
scheme of AATP-IBK.

8.2 Experiments on the Problem Management Log

The analyzed log comes from a problem handling system (named VINST ),
offered by Volvo IT Belgium as a benchmark dataset [11] for 2013 BPI Challenge.
The main attributes of each problem resolution case p are: p’s impact (i.e.,
medium, low, high); the product affected by p; the support team currently
solving p, and the associated functional division and organization line;
 An Approach to the Discovery of Accurate and Expressive Fix-Time 125

Table 3. Regression results on the VINST problem management system’s logs. Rows
and columns still correspond to different F T P M induction methods and different log
views, respectively.

Predictors rmse mae

Setting Methods L0 L1 L0 L1
No Bug History (Baseline) IBK 0.924 0.818 0.499 0.363
RepTree 0.937 0.539 0.523 0.245
Avg (no history) 0.931 0.679 0.511 0.304
History-aware CATP 0.930 0.563 0.542 0.278
IBK 0.916 0.802 0.466 0.357
AATP-IBK 0.914 0.613 0.465 0.242
RepTree 0.888 0.508 0.470 0.235
AATP-RepTree 0.887 0.467 0.466 0.205
Avg (history-aware) 0.907 0.591 0.482 0.265

the country where the support team is located (organization country); the
resource currently operating on p and its nationality (resource country); the
status and substatus of p. More specifically, the status of a problem may
take the following values: accepted, queued, completed, and closed. Attribute
substatus gives finer grain information on the resolution stage reached by a
problem, and can be: assigned, awaiting assignment, cancelled, in progress, wait
(i.e. information from third parts is needed to correctly diagnose the issue),
unmatched, and closed. With respect to the formalization of Sect. 3, for any
event e, what(e) takes the form of a pair status, substatus (i.e. both case
attributes are modified contemporaneously), while added(e) (resp., removed (e))
stores the new (resp., old) values associated with these attributes.

Log Manipulation Steps and Enhanced Log Views. A sample of 1487 issue records
(from January 2006 to May 2012) was turned into a set of traces, with lengths
ranging from 3 to 37, and fix times ranging from 1 to 2255 days (179 days in the
average).
By applying only a subset of our log manipulation operators, we generated
two (slightly) refined views of the original log, named L0 and L1 , as follows:

– View L0 was obtained by only applying, to the selected problem traces, a
basic event abstraction function, just amounting to extracting a combination
of the values assigned to attributes status and substatus;
– View L1 was produced by equipping each trace with several attributes derived
from its starting time: the global workload indicator (i.e. the number of prob-
lems open at that time), and the starting week-day, month and year. Each
trace was also enriched with an attribute firstOrg, indicating which organi-
zation line was initially chosen to work on it. Neither filtering nor macro-events
restructuring was employed to generate this view.
126 F. Folino et al.

Prediction Results. Table 3 shows the normalized rmse and mae errors obtained,
with different regression methods, on the two log views described above, still
considering two learning settings: with and without information on the past
history of a case.
Similarly to the bug tracking scenario, a neat decrease in both error mea-
sures is achieved when working with the more refined log view L1 (w.r.t. L0 ),
even though this latter was obtained by just using very simple log manipula-
tion operators. Moreover, the performances of baseline predictors seem to get
improved when exploiting information on past case history (“history-aware” set-
ting) and/or some kind of predictive clustering method (compare AATP-RepTree
and AATP-IBK to RepTree and IBK, respectively).

Fig. 4. Aggregated statistics for some trace clusters found (with AATP-IBK) on the
Eclipse log.

Table 4. Clusters found on the bug scenario (left) and the problem management
scenario (right).

id condition id condition
Reporter avgFT > 137.4 days ∧ workload > 600 product ∈ {PROD659, PROD476, PROD289, . . . }
c1
∧ Component ∈ {Data, LuaTools, Graphiti, MQTT, ∧ year > 2011 ∧ workload > 437
c2
documentation, Editor, General, Define, Photran, product ∈ {PROD378, PROD660, PROD473, . . . }
c6
Scout, Report, runtime, RDT, Website} ∧ year > 2011
54.46 days < Reporter avgFT ≤ 117.9 days product ∈ {PROD276, PROD277, PROD372, . . . }
c11 ∧ Component ∈ {IDE, Deployment, EDT, spdy, c8 ∧ firstOrg ∈ {Line C, Line B, Line V2, Line G4}
JavaGen, documentation, runtime} ∧ year > 2011 ∧ workload > 219
16.57 days< Reporter measure ≤ 54.47 days product ∈ {PROD659, PROD793, PROD597,
∧ Component ∈ {Search, Extra, Model, . . . } PROD96, PROD671, PROD231, PROD637,
∧ Product ∈ {Data, JDT, EclipseLink, STEM, c12 PROD209, PROD13, PROD812, PROD813,
c25
Jubula, Platform, Linux, Orbit, WTP, OSEE, MDT, PROD153} ∧ 2010 < year ≤ 2011
PTP, Virgo, JSDT, Orion} ∧ workload > 587
c28 year ≤ 2010 ∧ workload ≤ 11
 An Approach to the Discovery of Accurate and Expressive Fix-Time 127

8.3 Qualitative Results: Discovered Clusters

and Aggregate Statistics

Table 4 shows the classification rules associated with some of the (about 30)
clusters discovered, on each of the application scenarios, when using our app-
roach with AATP-IBK as base regressor. Beside supporting accurate cluster-based
predictions, these rules offer a readable and interesting description of differ-
ent context-dependent execution variants discovered for both issue resolution
processes. In particular, in the case of Ecplise’s bugs, most of the discovered
variants mainly depend on the reporter-based average fix time, on the compo-
nent/product affected, and on the workload indicator. Conversely, the variants
discovered for the problem management scenario are mainly linked to the prod-
uct, the workload and the year when the problem was reported.
Figure 4 shows a chart for the distribution of certain reporters’ groups (com-
puted over attribute reported with one of ours attribute summarization proce-
dures) in some of the clusters discovered (namely, c2 , c11 , c24 , c25 , c28 ).

9 Conclusions

We proposed a framework and an associated system for predicting issue/bug res-

olution times, which fully exploit case attributes’ change logs. A rich collection of
flexible data-transformation methods allows the analyst to produce high-quality
process-oriented logs, and eventually reuse regression/classification to discover a
state-based prediction model. Encouraging results obtained on real logs empiri-
cally proved the validity of our approach, and the interestingness of discovered
models.
As to future work, we plan to extend our approach as to deal with a wider
variety of target measures, ranging from Quality of Services measures to process-
centric Key Performance Indicators. We will also investigate the application of
our methods to the logs of other kinds of data-centric and lowly-structured
collaboration environments.

References
1. Anbalagan, P., Vouk, M.: On predicting the time taken to correct bug reports
in open source projects. In Proceedings of International Conference on Software
Maintenance (ICSM 2009), pp. 523–526 (2009)
2. Bhattacharya, P., Neamtiu, I.: Bug-fix time prediction models: can we do better?
In: Proceedings of 8th International Conference on Mining Software Repositories
(MSR 2011), pp. 207–210 (2011)
3. Breiman, L.: Random forests. Mach. Learn. 45(1), 5–32 (2001)
4. Folino, F., Guarascio, M., Pontieri, L.: Discovering context-aware models for pre-
dicting business process performances. In: Proceedings of 20th Intlernational Con-
ference on Cooperative Information Systems (CoopIS 2012), pp. 287–304 (2012)
128 F. Folino et al.

5. Folino, F., Guarascio, M., Pontieri, L.: A data-adaptive trace abstraction approach
to the prediction of business process performances. In: Proceedings of 15th Inter-
national Conference on Enterprise Information Systems (ICEIS 2013), pp. 56–65
(2013)
6. Folino, F., Guarascio, M., Pontieri, L.: Discovering high-level performance models
for ticket resolution processes. In: Proceedings of 21st International Conference on
Cooperative Information Systems (CoopIS 2013), pp. 275–282 (2013)
7. Giger, E., Pinzger, M., Gall, H.: Predicting the fix time of bugs. In: Proceedings of
2nd International Workshop on Recommendation Systems for Software Engineer-
ing (RSSE 2010), pp. 52–56 (2010)
8. Hooimeijer, P., Weimer, W.: Modeling bug report quality. In: Proceedings of 22nd
IEEE/ACM International Conference on Automated Software Engineering (ASE
2007), pp. 34–43 (2007)
9. Marks, L., Zou, Y., Hassan, A.E.: Studying the fix-time for bugs in large open
source projects. In: Proceedings of 7th Intlernational Conference on Predictive
Models in Software Engineering (Promise 2011), pp. 11:1–11:8 (2011)
10. Panjer, L.: Predicting eclipse bug lifetimes. In: Proceedings of 4th International
Workshop on Mining Software Repositories (MSR 2007), p. 29 (2007)
11. Steeman, W.B.P.I.: Challenge : closed problems. Ghent University (2013). doi:10.
4121/uuid:c2c3b154-ab26-4b31-a0e8-8f2350ddac11
12. Tan, P.-N., Steinbach, M., Kumar, V.: Introduction to Data Mining. Addison-
Wesley Longman, Boston (2005)
13. van der Aalst, W.M.P., Schonenberg, M.H., Song, M.: Time prediction based on
process mining. Inf. Syst. 36(2), 450–475 (2011)
Artificial Intelligence and Decision
Support Systems
 Solving the Unrelated Parallel Machine
Scheduling Problem with Setup Times
by Efficient Algorithms Based
on Iterated Local Search

Matheus N. Haddad1(B) , Luciano P. Cota1 ,

Marcone J.F. Souza1 , and Nelson Maculan2
1
Departamento de Computação, Universidade Federal de Ouro Preto,
Ouro Preto, MG 35.400-000, Brazil
{mathaddad,lucianoufop}@gmail.com, [email protected]
2
Programa de Engenharia de Sistemas e Computação,
Universidade Federal Do Rio de Janeiro, Rio de Janeiro, RJ 21941-972, Brazil
[email protected]

Abstract. The Unrelated Parallel Machine Scheduling Problem with

Setup Times (UPMSPST) is a problem that belongs to the N P-Hard
class and it is frequently found in many practical situations, like in textile
and chemical industries. The objective in UPMSPST is to schedule jobs
in machines in order to achieve the maximum completion time, known
as makespan. In an attempt to solve this problem, it is proposed two
algorithms: the AIV and the HIVP. Both algorithms are based on Iter-
ated Local Search (ILS) and Variable Neighborhood Descent (VND). The
difference between AIV and HIVP is that the first one generates a greedy
initial solution, while the second applies a partially greedy procedure to
construct the initial solution and it includes the Path Relinking (PR)
technique. Neighborhoods based on swaps and multiple insertions are
investigated in the developed algorithms. AIV and HIVP were tested on
benchmark test problems from literature and statistical analysis of the
computational results showed the superiority of them, outperforming the
previously best known solutions for UPMSPST.

Keywords: Unrelated parallel machine scheduling · Iterated local

search · Random Variable Neighborhood Descent · Makespan

1 Introduction
This paper deals with the Unrelated Parallel Machine Scheduling Problem with
Setup Times (UPMSPST), which can be formally defined as follows. Let N =
{1, ..., n} be a set of jobs and let M = {1, ..., m} be a set of unrelated machines.
The UPMSPST consists of scheduling n jobs on m machines, satisfying the
following characteristics: (i) Each job j ∈ N must be processed exactly once by
only one machine k ∈ M . (ii) Each job j ∈ N has a processing time pjk which
c Springer International Publishing Switzerland 2015
J. Cordeiro et al. (Eds.): ICEIS 2014, LNBIP 227, pp. 131–148, 2015.
DOI: 10.1007/978-3-319-22348-3 8
132 M.N. Haddad et al.

depends on the machine k ∈ M where it will be allocated. (iii) There are setup
times Sijk , between jobs, where k represents the machine on which jobs i and
j are processed, in this order. (iv) There is a setup time to process the first
job, represented by S0jk . The objective is to minimize the maximum completion
time of the schedule, the so-called makespan or also denoted by Cmax . Because
of such characteristics, UPMSPST is defined as RM | Sijk | Cmax [1]. In this
representation, RM represents the unrelated machines, Sijk the setup times and
Cmax the makespan.
Figure 1 illustrates a schedule for a test problem composed by two machines
and seven jobs. In Table 1 are presented the processing times of these jobs in
both machines. The setup times of these jobs in these machines are showed in
Table 2.

Table 1. Processing times in machines M1 and M2.

M1 M2
1 20 4
2 25 21
3 28 14
4 17 32
5 43 38
6 9 23
7 58 52

Table 2. Setup times in machines M1 and M2.

M1 1 2 3 4 5 6 7 M2 1 2 3 4 5 6 7
1 0 1 8 1 3 9 6 1 0 4 6 5 10 3 2
2 4 0 7 3 7 8 4 2 1 0 6 2 7 7 5
3 7 3 0 2 3 5 3 3 2 6 0 6 8 1 4
4 3 8 3 0 5 2 2 4 5 7 1 0 12 10 6
5 8 3 7 9 0 5 7 5 7 9 5 7 0 4 8
6 8 8 1 2 2 0 9 6 9 3 5 4 9 0 3
7 1 4 5 2 3 5 0 7 3 2 6 1 5 6 0

Fig. 1. An example of a possible schedule.

 Solving the Unrelated Parallel Machine Scheduling Problem 133

It can be observed that in machine M1 the jobs 2, 1 and 7 are allocated in

this order. In machine M2 the schedule of the jobs 5, 4, 6 and 3, in this order,
is also perceived by this figure. The cross-hatched areas of the figure represent
the setup times between jobs and the numbered areas the processing times. On
the line below the schedule there is the timeline, in which the times 120 and 130
represent the completion times of each machine.
As the job 6 is allocated to machine M2 its processing time p62 will
be 23. Its predecessor and its successor are the jobs 4 and 3, respectively.
So, in this example, are computed the times S462 = 5 and S632 = 5. Thus,
it can be calculated the completion time of machine M1 as S021 + p21 + S211 +
p11 + S171 + p71 = 120. Equivalently it is also calculated the completion time of
machine M2 as S052 + p52 + S542 + p42 + S462 + p62 + S632 + p32 = 130. After
the calculation of the completion times of machines M1 and M2, it can be con-
cluded that the machine M2 is the bottleneck machine. In other words, M2 is
the machine that has the highest completion time, the makespan.
The UPMSPST appears in many practical situations, like in textile and chem-
ical industries [2]. On the other hand, the UPMSPST is in N P-Hard class, as
it is a generalization of the Parallel Machine Scheduling Problem with Identical
Machines and without Setup Times [3,4]. The theoretical and practical impor-
tance instigate the study of the UPMSPST. Under these circumstances, finding
the optimal solution for UPMSPST using exact methods can be computation-
ally infeasible for large-sized problems. Thus, metaheuristics and local search
heuristics are usually developed to find good near optimal solutions.
In order to find these near optimal solutions for the UPMSPST, this paper
proposes two algorithms based on Iterated Local Search – ILS [5] and Variable
Neighborhood Descent – VND [6]. The first algorithm is called AIV, it starts
from an initial solution constructed on a greedy way by the Adaptive Shortest
Processing Time – ASPT rule. Then, this initial solution is refined by ILS,
using as local search the Random VND procedure. In this procedure, here called
RVND, there is no fixed sequence of neighborhoods, because they are sorted on
each application of the local search. In [7] the authors showed the effectiveness
of RVND over the conventional VND. The second algorithm named HIVP is
an endeavour to upgrade AIV, constructing the solution by a partially greedy
procedure and including the Path Relinking – PR technique.
Both algorithms were tested using benchmark instances from [8] and the
computational results showed that they are able to produce better solutions
than the ones found in literature, with lower variability and setting new upper
bounds for the majority of instances.
The rest of this paper is structured as follows. Firstly, works that inspired
the development of this paper are described. Then, the methodology used for the
deployment of this paper is presented. The computational results are shown on
sequence. Finally, this paper is concluded and possible proposals to be explored
are described.
134 M.N. Haddad et al.

2 Literature Review
In literature are found several works that seek to address the UPMSPST and
similar problems. These approaches were inspirations for the development of this
paper.
In [9] propose the development of seven heuristics with the objective of min-
imizing the weighted mean completion time. In [10], a problem with common
due dates is addressed and four heuristics are implemented for minimizing the
total weighted tardiness. In [11] aim to minimize the total weighted tardiness,
considering dynamic releases of jobs and dynamic availability of machines and
they used four dispatching rules in order to generate initial solutions and a Tabu
Search as the basis for the development of six search algorithms. This problem
is also addressed in [2], where a Branch-and-Price algorithm is developed.
More recent references are found when dealing with the UPMSPST. In [12]
created a Three-phase Partitioning Heuristic, called PH. In [13] it is proposed a
Metaheuristic for Randomized Priority Search (Meta-RaPS). In [14] bet in Tabu
Search for solving the UPMSPST. In [15] implement the Ant Colony Optimiza-
tion (ACO), considering its application to problems wherein the ratio of jobs
to machines is large. In [16] it is implemented a Restricted Simulated Anneal-
ing (RSA), which aims to reduce the computational effort by only performing
movements that the algorithm consider effective. In [17] is defined and proved a
set of proprieties for the UPMSPST and also implemented an Genetic Algorithm
and a Simulated Annealing using these proprieties. A hybridization that joins the
Multistart algorithm, the VND and a mathematical programming model is made
in [18]. In [19] solve the UPMSPST using Genetic Algorithms, with two sets of
parameters, yielding two algorithms, GA1 and GA2. In [19], the authors created
and provided test problems for the UPMSPST [8]. Also in [8] are presented the
best known solutions to the UPMSPST so far.

3 Methodology

3.1 Representation and Evaluation of a Solution

To represent a solution s it is used a vector v of m positions, with each position

representing a machine. In addition, each machine is associated with a list of
jobs allocated to it. The order of the jobs in this list represents the sequence of
execution.
The evaluation of a solution s is done by calculating the processing time of
the machine that will be the last to conclude the execution of its jobs. In other
words, the evaluation value is the makespan.

3.2 Proposed Algorithms

In this section two algorithms, AIV and HIVP, are proposed for solving UPM-
SPST. The details of operation of these algorithms are described below.
 Solving the Unrelated Parallel Machine Scheduling Problem 135

The AIV Algorithm. The first proposed algorithm, named AIV, combines the
heuristic procedures Iterated Local Search (ILS) and Random Variable Neighbor-
hood Descent (RVND). The main structure of AIV is based on ILS, using the
RVND procedure to perform the local searches.
The pseudo-code of AIV is presented in Algorithm 1.
The Algorithm 1 has only two input parameters: (1) timesLevel, which rep-
resents the number of times in each level of perturbation; (2) executionT ime,
the time in milliseconds that limits the execution of the algorithm.
First of all, AIV begins initializing the variable that controls the time limit,
currentT ime (line 1). Next, it initializes three empty solutions: the current solu-
tion s, the modified solution s and the solution that will store the best solution
found bestSol (line 2).
In line 3 a new solution is created based on the Adaptive Shortest Processing
Time (ASPT) rule (see Subsect. 3.3). Then, this new solution passes through
local searches at line 4, using the RVND module (see Subsect. 3.6).
136 M.N. Haddad et al.

In the next step, the current best known solution, bestSol, is updated (line 5)
and the level of perturbations is set to 1 (line 6).
After all these steps, the execution time is recalculated in line 7.
The iterative process of ILS is situated in lines 8 to 32 and it finishes when
the time limit is exceeded. A copy of the current solution to the modified solution
is made in line 9.
In lines 10 and 11 the variable that controls the number of times in each level
of perturbation (times) is initialized, as well as the variable that limits the max-
imum number of perturbations (maxP erturb). The following loop is responsible
to control the number of times in each level of perturbation (lines 12–27).
The next loop, lines 15 to 18, executes the perturbations (line 17) in the
modified solution (see Subsect. 3.8). The number of times this loop is executed
depends on the level of perturbation. With the perturbations accomplished,
 Solving the Unrelated Parallel Machine Scheduling Problem 137

the new solution obtained is evaluated and the RVND procedure is applied in this
new solution until a local optimum is reached, in relation to all neighborhoods
adopted in RVND.
In lines (20–24) it is verified if the changes made in the current solution were
good enough to continue the search from it. When the time is up, in bestSol will
be stored the best solution found by AIV.
Each module of AIV will be detailed afterwards.

The HIVP Algorithm. The second proposed algorithm, HIVP, is an attempt

to improve the AIV algorithm, differing only by two procedures: the construction
of the initial solution and the incorporation of an intensification/diversification
technique, the Path Relinking – PR [20]. The Algorithm 2 presents the pseudo-
code of HIVP.
Since HIVP is an upgrade of AIV, only the different lines will be
described next. The elite set is initialized at line 3. At line 4 the solution s
receives a solution generated by the partially greedy procedure CP GHBSS (see
Subsect. 3.4). Following, s passes through local searches using the RVND proce-
dure (see Subsect. 3.6) and bestSol receives the resulting solution from RVND.
This solution is inserted in the elite set (line 7). In the iterative loop of HIVP,
after the application of the RVND in the modified solution (line 21) the elite
set is updated and a random real number between 0 and 1 is generated (lines
22 and 23). If this number is less than or equal to 0.05 and the elite set is full
(with 5 solutions) it is performed an intensification/diversification of the search
using the PR technique - lines 23 to 27 (see Subsect. 3.7). The modified solution
and a random solution from the elite set are the input parameters of PR. If the
modified solution is considered an improvement (line 28), then the elite set is
updated at line 32.

3.3 ASPT: The Adaptive Shortest Processing Time Rule

The Adaptive Shortest Processing Time (ASPT) rule is an extension of the Short-
est Processing Time rule [21].
In ASPT, firstly, it is created a set N = {1, ..., n} containing all jobs and a
set M = {1, ..., m} that contains all machines.
From the set N , the jobs are classified according to an evaluation function
gk . This function is responsible to obtain the completion time of the machine k.
Given a Candidate List (CL) of jobs, it is evaluated, based on the gk function,
the insertion of each of these jobs in all positions of all machines. The aim is to
obtain in which position of what machine that the candidate job will produce
the lowest completion time, that is, the gmin .
If the machine with the lowest completion time has not allocated any job
yet, its new completion time will be the sum of the processing time of the job
to be inserted with the initial setup time for such job.
If this machine has some job, its new completion time will be the previous
completion time plus the processing time of the job to be inserted and the setup
times involved, if it has sequenced jobs before or after.
138 M.N. Haddad et al.

This allocation process ends when all jobs are assigned to some machine,
thus producing a feasible solution, s. This solution is returned by the heuristic.
The algorithm is said to be adaptive because the choice of a job to be inserted
depends on the preexisting allocation.

3.4 CP GHBSS : The Partially Greedy Procedure

The CP GHBSS procedure is a partially greedy constructive method which also
uses the evaluation function gk to classify the jobs. The same Candidate List
(CL) created in ASPT is generated here. Notwithstanding, instead of deter-
ministically choosing the job that produces the gmin , the CP GHBSS procedure
chooses a candidate in the CL based on the Heuristic-Biased Stochastic Sampling
(HBSS) [22]. The jobs in the CL are chosen according to a probability arising
from a bias function.
In Table 3 are presented the following bias functions: Random, Logarithmic,
Polynomial of degree 2, Polynomial of degree 3, Polynomial of degree 4 and
Exponential. The first column shows the index of the candidates. In each cell of
the table there is the probability of the candidate i be chosen using the respective
bias function. For example, if the bias function selected is the Exponential, the
fifth element of the CL will have 1.2 % chance of being chosen.
The bias function chosen in this work was the Exponential, this choice
was made based on empirical tests. With the bias function selected, it can be
explained how the CP GHBSS procedure constructs a solution.
Initially, as in ASPT, all the jobs and machines are stored in the sets N and
M , respectively. Then, all jobs are inserted in the CL.
While there are jobs in the CL: firstly, jobs are classified according to the
gk function (see Subsect. 3.3), analyzing their inclusion at the last position of a
machine and creating a list named Rank. Rank is composed by combinations
of (j, k), with j ∈ LC and k ∈ M , ordered according to the value calculated by
the gk function.
A probability according to the Exponential bias function values, as presented
in Table 3, is associated to each pair in Rank. Secondly, the Wheel Selection
method is performed, in order to select which job will be inserted in the partial

Table 3. Table containing bias functions. Source: [22].

Rand Log Linear P oly 2 P oly 3 P oly 4 Exp

1 0.033 0.109 0.250 0.620 0.832 0.924 0.632
2 0.033 0.069 0.125 0.155 0.104 0.058 0.233
3 0.033 0.055 0.083 0.069 0.031 0.011 0.086
4 0.033 0.047 0.063 0.039 0.013 0.004 0.031
5 0.033 0.042 0.050 0.025 0.007 0.001 0.012
6-30 0.033 0.678 0.429 0.092 0.013 0.002 0.006
 Solving the Unrelated Parallel Machine Scheduling Problem 139

Fig. 2. Example of a multiple insertion movement.

solution. Then, the selected job is removed from LC and all pairs containing this
job are removed from Rank. This loop is executed until all jobs are inserted in
the solution, generating a feasible solution.

3.5 Neighborhood Structures

Three neighborhood structures are used to explore the solution space. These
structures are based on swap and multiple insertion movements of the jobs.

Multiple Insertion. The Multiple Insertion (MI) movement consists in real-

locating a job from a position and inserting it on any position of the entire
schedule, including the machine that it was previously allocated. The MI move-
ment belongs to the N M I (.) neighborhood. To exemplify this movement, Fig. 2
illustrates the transfer of the job 4 from the second position of machine M 2 to
the second position of machine M 1.

Swap in the Same Machine. The Swap in the Same Machine (SSM) move-
ment, as the name suggests, is done by swapping the positions of two jobs
presented in the same machine. The SSM movement belongs to the N SSM (.)
neighborhood. Figure 3 shows the swap of jobs 5 and 6 in machine M 2.

Swap Between Different Machines. The Swap Between Different Machines

(SDM), which belongs to the N SDM (.) neighborhood, consists in swapping two
jobs that are allocated in different machines. This movement can be better exem-
plified by Fig. 4, where the swap of the jobs 7 (M 1) and 5 (M 2) is perceived.

3.6 RVND: The Random Variable Neighborhood

Descent Procedure

The Random Variable Neighborhood Descent – RVND procedure [7,23] is a vari-

ant of the VND procedure [6].
140 M.N. Haddad et al.

Fig. 3. Example of a swap in the same machine movement.

Fig. 4. Example of a swap between different machines movement.

Each neighborhood of the set {N M I , N SSM , N SDM } described in Sect. 3.5

defines one local search. Unlike VND, the RVND explores the solution space
using these three neighborhoods in a random order each time the procedure
is triggered. The RVND is finished when it is found on a local optimum with
relation to the three considered neighborhoods.
Following are described the local searches procedures used in RVND.
F I 1M I : Local Search with Multiple Insertion. The first local search uses
multiple insertions movements (N M I (.) neighborhood) with the First Improve-
ment strategy. In this search, each job of each machine is inserted in all positions
of all machines.
The selection of the jobs to be removed respects the allocation order in the
machines. That is, initially, the first job is selected to be removed, then the second
job until all jobs from a machine are chosen. The machines that will have their
jobs removed are selected based on their completion times. The search starts
with machines with higher completion times to machines with lower completion
times.
By contrast, the insertions are made from machines with lower completion
times to machines with higher completion times. The jobs are inserted starting
from the first position and stopping at the last position.
The movement is accepted if the completion times of the machines involved
are reduced. If the completion time of a machine is reduced and the completion
 Solving the Unrelated Parallel Machine Scheduling Problem 141

time of another machine is added, the movement is also accepted. However,

in this case, it is only accepted if the value of reduced time is greater than the
value of time increased.
It is noteworthy that even in the absence of improvement in the value of
makespan, the movement can be accepted. Upon such acceptance of a movement,
the search is restarted and only ends when it is found a local optimum, that is,
when there is no movement that can be accepted in the neighborhood of multiple
insertion.
F I SM D : Local Search with Swaps Between Different Machines. The
second local search makes swap movements between different machines, exploring
the N SM D (.) neighborhood. For each pair of existing machines every possible
swap of jobs between them are analyzed.
Exchanges are made from machines that have higher completion times to
machines with lower completion times. The acceptance criteria are the same as
those applied in the first local search. If there are reductions in completion times
on two machines involved, then the movement is accepted. If the reduced value
of the completion time of a machine is larger than the completion time plus
another machine, the movement is also accepted. Once a movement is accepted,
the search stops.
BI SSM : Local Search with Swaps on the Same Machine. The third
local search examines the N SSM (.) neighborhood and uses the strategy Best
Improvement.
The machines are ordered from the machine that has the highest value of
completion time to the machine that has the lowest value of completion time.
For each machine, starting from the first, all possible swaps involving pairs
of jobs are investigated. The best movement is accepted if the completion time
of the machine is reduced and, in this case, the local search is repeated from this
solution; otherwise, the next machine is analyzed.
This local search only ends when no improvements is found in 30 % of the
machines.

3.7 Path Relinking

The Path Relinking – PR [20] technique makes a balance between intensification
and diversification of the search. Its objective is to explore existing paths between
high quality solutions. These high quality solutions are stored in an elite set.
In order to a solution s enter in this elite set, one of the following conditions
must be satisfied: (i) be better than the best solution of the elite set, in terms of
the makespan value; (ii) be better (lower makespan) than the worst solution of
the elite set and differentiate itself from all solutions of the elite set at least by
10 %. The diversity criterion, when comparing two solutions, is the percentage
of jobs allocated in different positions.
The Backward Path Relinking (BkPR) strategy is used. According [20] BkPR
usually outperforms forward path relinking. Thus, a path is constructed from a
base solution to a guide solution, being the best solution as the base solution and
142 M.N. Haddad et al.

the worst solution as the guide solution. In this work, this strategy is applied
over the following solutions: (1) a solution chosen randomly in the elite set;
(2) the solution returned by the RVND local searches.
The attribute chosen for building the path is the position of a job. Initially,
the jobs positions of the guide solution are inserted in a list named Φ. In each iter-
ation is analyzed the insertion, in the base solution, of an attribute that belongs
to the guide solution. Following, the other copy of this job is removed from the
base solution. Moreover, if the machine that receives this job has another differ-
ent job at the same position, then this job is relocated to another position that
has not set its attribute from the guide solution yet.
With all attributes from guide solution analyzed, it is included to base solu-
tion the attribute which produces the lower cost. This cost is given by the sum
of all completion times of all machines in base solution. After the insertion of
2
an attribute, to this new base solution is applied the F IM I local search, defined
next. It is important to highlight that once an attribute is inserted in base solu-
tion, this attribute can not be changed.
Following, the selected attribute is removed from Φ. These steps are repeated
until Φ is empty. At the end of the algorithm the base solution will have the
same scheduling as the guide solution and the best solution found during this
procedure is returned.
F I 2M I : Local Search with Multiple Insertion. As the local search F IM 1
I,
2 MI
the local search F IM I also explores the N (.) neighborhood with the First
Improvement strategy. But it differs by two characteristics: (i) the only accep-
tance criterion is the improvement of the makespan; (ii) when an improvement
occurs, the method is stopped and the new solution is returned.

3.8 Perturbations

A perturbation is characterized by applying an insertion movement in a local

optimum, but this movement differs when inserting the job in another machine.
The job will be inserted into its best position, that is, in the position that
will produce the lowest completion time. Doing so, sub parts of the problem are
optimized after each perturbation. The machines and the job involved are chosen
randomly.
In both AIV and HIVP, the number of perturbations applied to a solution is
controlled by the level of perturbation. A level l of perturbation consists in the
application of l + 1 insertion movements. The maximum level allowed for the
perturbations is set to 3.
If timeslevel perturbed solutions are generated without an improvement in
the current solution the perturbation level is increased. If an improvement of
the current solution is found, the level of perturbation is set to its lowest level
(l = 1).
 Solving the Unrelated Parallel Machine Scheduling Problem 143

3.9 Efficient Evaluation of the Objective Function

The evaluation of an entire solution after every movement, multiple insertion

or swap, demands a large computational effort. Aiming to avoid this situation,
it was created a procedure that evaluates only the processing and setup times
involved in the movements. In this way, in order to obtain the new completion
time of each machine it is necessary few additions and subtractions.
Taking the example of the multiple insertion movement illustrated in Fig. 2,
the new completion time of machine M2 is obtained by subtracting from its
previous value the processing time of job 4 p42 and also subtracting the setup
times involved, S542 and S462 . The setup time S562 also needs to be added to the
completion time of machine M2. In machine M1, the processing time of job 4
p41 and the setup times S241 and S411 are included in the new completion time.
Then, the new completion time of machine M1 is M 1 = 120−4+3+17+3 = 139
and the new completion time of machine M2 is M 2 = 130 − 7 − 32 − 5 + 4 = 90.
Although the given example is for the multiple insertion movement, it is trivial
to apply the same procedure for a swap movement.

4 Computational Results
Using a set of 360 test problems from [8] the computational tests were performed.
This set of test problems involves combinations of 50, 100 and 150 jobs with 10,
15 and 20 machines. There are 40 instances for each combination of jobs and
machines. The best known solutions for each of these test problems are also
provided in [8].
AIV was developed in C++ language and HIVP was developed in Java lan-
guage. All experiments were executed in a computer with Intel Core i5 3.0 GHz
processor, 8 GB of RAM memory and in Ubuntu 12.04 operational system.
The parameters used in both AIV and HIVP are: (i) the number of iterations
on each level of perturbation: timeslevel = 15; (ii) the stop criterion: T imemax ,
which is the maximum time of execution, in milliseconds, obtained by Eq. 1. In
this equation, m represents the number of machines, n the number of jobs and
t is a parameter that was tested with three values for each instance: 10, 30 and
50. It is observed that the stop criterion, with these values of t, was the same
adopted in [19].
T imemax = n × (m/2) × t ms (1)
With the objective to verify the variability of final solutions produced by the
algorithms it was used the metric given by Eq. 2. This metric is used to compare
algorithms. For each algorithm Alg applied to a test problem i is calculated the
Relative Percentage Deviation RP Di of the solution found f¯iAlg in relation to
the best known solution fi∗ .
In this paper, the algorithms AIV and HIVP were executed 30 times, for
each instance and for each value of t, calculating the Average Relative Percent-
age Deviation RP Diavg of the RP Di values found. In [19] the algorithms were
144 M.N. Haddad et al.

executed 5 times for each instance and for each value of t.

f¯iAlg − fi∗
RP Di = (2)
fi∗

Table 4. Average Relative Percentage Deviation of the algorithms AIV, HIVP and
GA2 with t = 10/30/50.

Set of Instances AIVa HIVPa GA2b

50 x 10 3.69/1.83/1.30 2.27/0.25/-0.44 7.79/6.92/6.49
50 x 15 1.52/-0.77/-1.33 -0.6/-2.78/-3.47 12.25/8.92/9.20
50 x 20 5.26/2.01/1.65 -2.14/-4.06/-4.68 11.08/8.04/9.57
100 x 10 5.06/2.93/2.00 4.45/2.04/1.23 15.72/6.76/5.54
100 x 15 1.80/-0.40/-1.29 2.17/-0.64/-1.78 22.15/8.36/7.32
100 x 20 0.52/-1.64/-2.89 0.58/-2.44/-3.92 22.02/9.79/8.59
150 x 10 3.77/1.99/1.07 4.21/2.15/0.98 18.40/5.75/5.28
150 x 15 1.83/-0.24/-1.04 3.37/0.42/-0.44 24.89/8.09/6.80
150 x 20 -1.04/-3.10/-4.00 1.22/-2.19/-3.39 22.63/9.53/7.40
avg
RP D 2.49/0.29/-0.50 1.72/-0.8/-1.77 17.44/8.02/7.35
a
Executed on Intel Core i5 3.0 GHz, 8 GB of RAM, 30 runs for each instance
b
Executed on Intel Core 2 Duo 2.4 GHz, 2 GB of RAM, 5 runs for each
instance

In Table 4 are presented, for each set of instances, the RP Diavg values
obtained for each value of t = 10, 30, 50 by AIV and HIVP, and also it con-
tains the RP Diavg values obtained by GA2, a genetic algorithm developed in
[19]. To our knowledge, the results reported in the literature for this set of test
problems are only presented in [19] and the best algorithm tested by the authors
was GA2.
There are three values of RP Diavg separated by a ’/’ for each set of instances
in the table. Each separation represents test results with different values of t,
10/30/50. If a negative value is found, it means that the reached result outper-
formed the best value found in [19] on their experiments.
The best values of RP Davg are highlighted in bold. It is remarkable that
HIVP is the algorithm that found the best results. Not only it improved the
majority of best known solutions, but also it won in 63 % of the sets of instances.
The algorithm AIV found the best results in 37 % of the remainder sets of
instances.
A table with all results found by both algorithms and also the previous best
known values for the UPMSPST can be found in http://www.decom.ufop.br/
prof/marcone/projects/upmsp/Experiments UPMSPST AIV HIVP.ods
The box plot, Fig. 5, contains all RP Davg values for each algorithm. It is
notable that 100 % of the RP D values encountered by both AIV and HIVP
 Solving the Unrelated Parallel Machine Scheduling Problem 145

25
20
15
RPD_avg
10
5
0
−5

AIV GA2 HIVP

Algorithms

Fig. 5. Box plot showing the RP Davg of the algorithms.

Table 5. Results from Tukey HSD test.

Algorithms diff lwr upr p

GA2-AIV 10.077407 7.495481 12.659333 0.0000000
HIVP-AIV -1.041481 -3.623408 1.540445 0.6018344
HIVP-GA2 -11.118889 -13.700815 -8.536963 0.0000000

outperformed the ones obtained by GA2 algorithm. By the way, it is observed

that 75 % of solutions found by the developed algorithms are near the best known
solutions. Besides, more than 50 % of solutions found by the HIVP algorithm are
better than the best known so far. With AIV this percentage drops to 25 %.
The results were submitted to the Shapiro-Wilk test [24] to verify if the
sample satisfies the normality test, so that it can be decided which test to use
for analyzing statistical differences between all algorithms. The Shapiro-Wilk
returned, with significance level of 5 %, W = 0.9261 and p = 0.2692. As p =
0.2692 > 0.05 then it can be concluded with 95 % of confidence level that the
sample are taken from a normal distribution.
Thus, in order to verify if exist statistical differences between the RP D val-
ues, it was applied an analysis of variance (ANOVA) [25]. This analysis returned,
with 95 % of confidence level and threshold = 0.05, that p = 2 × 10−16 . As
p < threshold, it is possible to ensure that exist statistical differences between
the RP D values.
A Tukey HSD test, with 95 % of confidence level and threshold = 0.05, was
used for checking where are these differences. Table 5 contains the differences in
the average values of RP D (diff), the lower end point (lwr), the upper end point
(upr) and the p-value (p) for each pair of algorithms.
The p-value shows that when comparing AIV to GA2 there is a statistical
difference between them, because it was less than the threshold. The same con-
clusion can be achieved when comparing HIVP to GA2. However, when AIV is
146 M.N. Haddad et al.

95% family−wise confidence level

GA2−AIV
HIVP−AIV
HIVP−GA2

−10 −5 0 5 10
Differences in mean levels of algoritmos

Fig. 6. Graphical results from Tukey HSD test.

compared to HIVP they are not statistically different from each other, since the
p-value was greater than the threshold.
By plotting the results from the Tukey HSD test (Fig. 6) it is more noticeable
that HIVP and AIV are statistically different from GA2, as their graphs do not
pass through zero. Comparing algorithms HIVP and AIV it can be perceived that
they are not statistically different from each other, because the graph passes
through zero. Thus, with a statistical basis it can be concluded, within the
considered instances, that AIV and HIVP are the best algorithms on obtaining
solutions for UPMSPST.

5 Conclusions

This paper studied the Unrelated Parallel Machine Scheduling Problem with
Setup Times (UPMSPST), aiming to the minimization of the maximum com-
pletion time of the schedule, the makespan.
In order to solve the UPMSPST it was proposed two algorithms based on
Iterated Local Search (ILS) and Variable Neighborhood Descent (VND). The first
algorithm was named AIV and it implements the Adaptive Shortest Processing
Time (ASPT) rule to create an initial solution. The Random Variable Neigh-
borhood Descent (RVND) procedure was used to perform the local searches,
randomly exploring the solution space with insertions and swap movements.
A perturbation in AIV is an application of an insertion movement. The second
algorithm, called HIVP, is an attempt to upgrade AIV, constructing the solution
using a partially greedy procedure and incorporating the Path Relinking (PR)
technique in order to intensify and diversify the search.
The two developed algorithms were applied to instances taken from literature
and the results were compared the genetic algorithm GA2, developed in [19].
Statistical analysis of the computational results showed that AIV and HIVP are
 Solving the Unrelated Parallel Machine Scheduling Problem 147

able to produce 100 % of better solutions than GA2. HIVP and AIV were also
able to generate new upper bounds for these test problems. Although HIVP
seems to be better than AIV, statistically this was not proved. Nevertheless, it
can be concluded that both AIV and HIVP are two efficient algorithms when
dealing with the UPMSPST.
For future works it is proposed the application of both algorithms on the
entire set of test problems available in [8]. An improvement that will be studied
is an incorporation of a Mixed Integer Programming (MIP) model to AIV or
HIVP for solving related sub problems.

Acknowledgements. The authors thank the Brazilian agencies FAPEMIG and

CNPq, and the Universidade Federal de Ouro Preto (UFOP) for the financial sup-
port on the development of this work.

References
1. Graham, R., Lawler, E., Lenstra, J., Kan, A.: Optimization and approximation in
deterministic sequencing and scheduling: a survey. Ann. Discret. Math. 5, 287–326
(1979)
2. Lopes, M.J.P., de Carvalho, J.M.: A branch-and-price algorithm for scheduling
parallel machines with sequence dependent setup times. Eur. J. Oper. Res. 176,
1508–1527 (2007)
3. Karp, R.M.: Reducibility among combinatorial problems. Complex. Comput. Com-
put. 40, 85–103 (1972)
4. Garey, M., Johnson, D.: Computers and Intractability: A Guide to the Theory of
Np-Completeness., vol. 174. WH Freeman & Co., San Francisco (1979)
5. Lourenço, H.R., Martin, O., Stützle, T.: Iterated local search. In: Glover, F.,
Kochenberger, G. (eds.) Handbook of Metaheuristics. International Series in Oper-
ations Research and Management Science, vol. 57, pp. 321–353. Kluwer Academic
Publishers, Norwell (2003)
6. Mladenovic, N., Hansen, P.: Variable neighborhood search. Comput. Oper. Res.
24, 1097–1100 (1997)
7. Souza, M., Coelho, I., Ribas, S., Santos, H., Merschmann, L.: A hybrid heuristic
algorithm for the open-pit-mining operational planning problem. Eur. J. Oper.
Res. 207, 1041–1051 (2010)
8. de Optimización Aplicada, S.: A web site that includes benchmark prob-
lem data sets and solutions for scheduling problems (2011). http://soa.iti.es/
problem-instances
9. Weng, M.X., Lu, J., Ren, H.: Unrelated parallel machine scheduling with setup
consideration and a total weighted completion time objective. Int. J. Prod. Econ.
70, 215–226 (2001)
10. Kim, D.W., Na, D.G., Chen, F.F.: Unrelated parallel machine scheduling with
setup times and a total weighted tardiness objective. Robot. Comput. Integr.
Manuf. 19, 173–181 (2003)
11. Logendran, R., McDonell, B., Smucker, B.: Scheduling unrelated parallel machines
with sequence-dependent setups. Comput. Oper. Res. 34, 3420–3438 (2007)
12. Al-Salem, A.: Scheduling to minimize makespan on unrelated parallel machines
with sequence dependent setup times. Eng. J. Univ. Qatar 17, 177–187 (2004)
148 M.N. Haddad et al.

13. Rabadi, G., Moraga, R.J., Al-Salem, A.: Heuristics for the unrelated parallel
machine scheduling problem with setup times. J. Intell. Manuf. 17, 85–97 (2006)
14. Helal, M., Rabadi, G., Al-Salem, A.: A tabu search algorithm to minimize the
makespan for the unrelated parallel machines scheduling problem with setup times.
Int. J. Oper. Res. 3, 182–192 (2006)
15. Arnaout, J., Rabadi, G., Musa, R.: A two-stage ant colony optimization algorithm
to minimize the makespan on unrelated parallel machines with sequence-dependent
setup times. J. Intell. Manuf. 21, 693–701 (2010)
16. Ying, K.C., Lee, Z.J., Lin, S.W.: Makespan minimisation for scheduling unrelated
parallel machines with setup times. J. Intell. Manuf. 23, 1795–1803 (2012)
17. Chang, P., Chen, S.: Integrating dominance properties with genetic algorithms for
parallel machine scheduling problems with setup times. Appl. Soft Comput. 11,
1263–1274 (2011)
18. Fleszar, K., Charalambous, C., Hindi, K.: A variable neighborhood descent heuris-
tic for the problem of makespan minimisation on unrelated parallel machines with
setup times. J. Intell. Manuf. 23, 1949–1958 (2011). doi:10.1007/s10845-011-0522-8
19. Vallada, E., Ruiz, R.: A genetic algorithm for the unrelated parallel machine
scheduling problem with sequence dependent setup times. Eur. J. Oper. Res. 211,
612–622 (2011)
20. Resende, M.G.C., Ribeiro, C.C.: GRASP: greedy randomized adaptive search pro-
cedures. In: Burke, E.K., Kendall, G. (eds.) Search Methodologies, 2nd edn, pp.
287–312. Springer, New York (2013)
21. Baker, K.R.: Introduction to Sequencing and Scheduling. John Wiley & Sons,
New York (1974)
22. Bresina, J.L.: Heusistic-biased stochastic sampling. In: Proceedings of the Thir-
teenth National Conference on Artificial intelligence, vol. 1, pp. 271–278 (1996)
23. Subramanian, A., Drummond, L., Bentes, C., Ochi, L., Farias, R.: A parallel heuris-
tic for the vehicle routing problem with simultaneous pickup and delivery. Comput.
Oper. Res. 37, 1899–1911 (2010)
24. Shapiro, S.S., Wilk, M.B.: An analysis of variance test for normality (complete
samples). Biometrika 52, 591–611 (1965)
25. Montgomery, D.: Design and Analysis of Experiments, 5th edn. John Wiley &
Sons, New York (2007)
 Using Data Mining for Prediction
of Hospital Length of Stay: An Application
of the CRISP-DM Methodology

Nuno Caetano1 , Paulo Cortez2(B) , and Raul M.S. Laureano3

1
HFAR - Hospital Das Foras Armadas, Azinhaga Ulmeiros, 1620-060
Lisboa, Portugal
[email protected]
2
ALGORITMI Research Centre, Department of Information Systems,
University of Minho, 4800-058 Guimarães, Portugal
[email protected]
3
Business Research Unit (UNIDE-IUL), Instituto Universitário de Lisboa
(ISCTE-IUL), Av. Das Forças Armadas, 1629-026 Lisboa, Portugal
[email protected]

Abstract. Hospitals are nowadays collecting vast amounts of data

related with patient records. All this data hold valuable knowledge that
can be used to improve hospital decision making. Data mining tech-
niques aim precisely at the extraction of useful knowledge from raw
data. This work describes an implementation of a medical data mining
project approach based on the CRISP-DM methodology. Recent real-
world data, from 2000 to 2013, were collected from a Portuguese hospi-
tal and related with inpatient hospitalization. The goal was to predict
generic hospital Length Of Stay based on indicators that are commonly
available at the hospitalization process (e.g., gender, age, episode type,
medical specialty). At the data preparation stage, the data were cleaned
and variables were selected and transformed, leading to 14 inputs. Next,
at the modeling stage, a regression approach was adopted, where six
learning methods were compared: Average Prediction, Multiple Regres-
sion, Decision Tree, Artificial Neural Network ensemble, Support Vector
Machine and Random Forest. The best learning model was obtained by
the Random Forest method, which presents a high quality coefficient of
determination value (0.81). This model was then opened by using a sensi-
tivity analysis procedure that revealed three influential input attributes:
the hospital episode type, the physical service where the patient is hos-
pitalized and the associated medical specialty. Such extracted knowledge
confirmed that the obtained predictive model is credible and with poten-
tial value for supporting decisions of hospital managers.

Keywords: Medical data mining · Hospitalization process · Length of

stay · CRISP-DM · Regression · Random forest

c Springer International Publishing Switzerland 2015

J. Cordeiro et al. (Eds.): ICEIS 2014, LNBIP 227, pp. 149–166, 2015.
DOI: 10.1007/978-3-319-22348-3 9
150 N. Caetano et al.

1 Introduction
In recent decades, hospitals have been collecting large amounts of data into their
clinical information systems. All this data hold valuable knowledge and therefore
there is an increasing potential of the use of Data Mining (DM) [1], to facilitate
the extraction of useful knowledge and support clinical decision making, in what
is known as medical data mining [2]. There are several successful medical data
mining applications, such as the prediction of mortality [3] and degree of organ
failure [4] at Intensive Care Units, and the segmentation of tissue from magnetic
resonance imaging [5], among others.
This work focuses on the prediction of the Length Of Stay (LOS), defined
in terms of the inpatient days, which are computed by subtracting the day of
admission from the day of discharge. Extreme LOS values are known as prolonged
LOS and are responsible for a major share in the hospitalization total days and
costs. The use of data-driven models for predicting LOS is of value for hospital
management. For example, with an accurate estimate of the patients LOS, the
hospital can better plan the management of available beds, leading to a more
efficient use of resources by providing a higher average occupancy and less waste
of hospital resources [6,7].
DM aims at the extraction of useful knowledge from raw data [1]. With the
growth of the field of DM, several DM methodologies were proposed to systematize
the discovery of knowledge from data, including the tool neutral and popular Cross-
Industry Standard Process for Data Mining (CRISP-DM) [8], which is adopted
in this work. The methodology is composed of six stages: business understanding,
data understanding, data preparation, modeling, evaluation and implementation.
This study describes the adopted DM approach under the first five stages
of CRISP-DM, given that implementation is left for future work. The main
goal was to predict generic LOS (for all hospital services) under a regression
approach using past patterns existing in the hospitalization process, based on a
DM techniques. The data is related with a Portuguese hospital, based on recent
data collected from the hospitalization process between 2000 and 2013, including
a total of 26462 records from 15253 patients. At the preprocessing stage, the
data were cleaned and attributes were selected, leading to 14 inputs and the
LOS target. During the modeling stage, six regression methods were tested and
compared: Average Prediction (AP), Multiple Regression (MR), Decision Tree
(DT) and state-of-the-art regression methods [9], including an Artificial Neural
Network (ANN) ensemble, Support Vector Machine (SVM) and Random Forest
(RF). The predictive models were compared using a cross-validation procedure
with three popular regression metrics: coefficient of determination (R2 ), Mean
Absolute Error (MAE) and Root Mean Squared Error (RMSE). Moreover, the
best predictive model (RF) was opened using a sensitivity analysis procedure
[10] that allows ranking the input attributes and also measuring the average
effect of a particular input in the predictive response.
This work is organized as follows. Firstly, the relevant related work is pre-
sented (Sect. 2). Then, the adopted DM approach is detailed in terms of the
CRISP-DM methodology first five phases (Sect. 3). Finally, closing conclusions
are drawn (Sect. 4).
 Using Data Mining for Prediction of Hospital Length of Stay 151

2 Related Work
Nowadays, hospital managers are pressured to accomplish several goals, such
as providing better health care, increasing the number of available beds for
new admissions and reduce surgical waiting lists. Under this context, LOS is
used worldwide as a highly relevant measure to analyze the hospital resources
consumption and to monitor hospital performance [7]. Given the importance
of LOS, a large number of studies have adopted a data-driven approach for
modeling LOS. In the next few paragraphs, we present some examples of related
studies.
In 1998, Merom et al. [11] estimated the rate of inappropriate hospital days
(failure of established criteria for admission) and the identification of the vari-
ables associated with this impropriety. During such study, 1369 patients from
24 hospitals were analyzed under a multiple regression model. Several attributes
were used in their analysis: occupation, group age, gender, inappropriate days,
government, another hospital entity, another diagnosis, origin, admission diag-
nosis and period of stay.
In 2007, Abelha et al. [12] evaluated LOS of patients submitted to a non-
cardiac surgery and admitted to a surgical Intensive Care Unit (ICU) between
October 2004 and July 2005. The attributes used to categorize patients were:
age, gender, body mass index, physical status, type and magnitude of surgery,
type and duration of anesthesia, temperature on admission, LOS in the ICU
and in hospital mortality in the ICU and hospital. A simple linear regression
model was adopted and from the results it was found that the average LOS was
4.22±8.76 days.
In 2010, Oliveira et al. [13], proposed to evaluate factors associated with
higher mortality and prolonged LOS in ICUs. The study included 401 patients
consecutively admitted to the ICU, within a six-month period. The collected
attributes were: gender, age, diagnosis, personal history, APACHE II score,
mechanical ventilation days, endotracheal reintubation, tracheostomy, LOS in
the ICU, ICU discharge or death. In terms of results, the average LOS in the
ICU was 8.2±10.8 days. The study concluded that factors such as APACHE II,
reintubation and tracheostomy were associated with higher mortality and pro-
longed LOS in the ICU.
Also in 2010, Karla et al. [14] studied the temporal trends of the workflow
in the internal medicine service of an University Hospital. The data analyzed
were obtained in that service for three different time periods spanning through
13 years. The most relevant data features data were: date of admission, date
of departure or death, gender, age, residence code, financial entity and primary
diagnosis. Their results have confirmed several changes in LOS behavior through
time (e.g., the number of admissions in the internal medicine service statistically
increased from 1991 to 2004).
More recently, Freitas et al. [15] analyzed in 2012 LOS outliers based on inpa-
tient episodes of Portuguese public hospitals belonging to the national health sys-
tem, with data collected between 2000 and 2009. The variables used for analysis
152 N. Caetano et al.

were: age, distance from residence to hospital, year of discharge, comorbidities,

A-DRG complexity, readmission, admission and DRG type, discharge status and
hospital type. In the analysis they used logistic regression models to examine the
association of each variable with the time of admissions outliers, and model with
all variables to calculate the adjusted odds ratios and respective confidence inter-
vals (95 %). In terms of results, nine million inpatient episodes were analyzed, of
which 3.9 % were considered high LOS outliers. They concluded that age, type
of admission and hospital type attributes were significantly associated with high
LOS outliers.
In the same year (2012), Azari et al. [6] explored a classification approach to
predict LOS. The main attributes of their analysis were: specialty services, days
elapsed since the first act of the year, primary condition group (generalized code
for the principal diagnosis) and Charlson index (diagnostic code) and LOS. The
LOS was divided in three different classification groups: one to two days, greater
than two and less than seven days, and longer or equal to seven days. The study
concluded that the performance of classification techniques could be improved
by incorporating a clustering step during the training stage.
Also in 2012, Castillo [7] developed a statistical model to predict the LOS in
Mexican public hospitals. The following attributes were used: age, gender, occu-
pation, education level, previous visits, origin, surgical first diagnosis, diagnosis,
surgical procedure, number of surgical procedures and ward. The best predictive
model was given by a probabilistic model based on a cluster analysis.
Finally, Sheikh-Nia in 2012 [16] used a sequential ensemble of classification
algorithms to predict LOS of patients in the next year, based on the patient
previous medical history. The main attributes considered were: age at first claim,
gender, provider, year, medical specialty, number of days from the first record,
primary condition group, Charlson index and LOS. The results showed that all
of the independent classifiers exceeded the baseline by a factor of 1.78 for the
ANN, 1.20 for K-Nearest Neighbor and 1.17 for DT.
Instead of predicting LOS in specialized medical services, as in UCI [12,13]
or internal medicine [14], in this work we predict generic LOS, for all hospi-
tal services (e.g., internal medicine, general surgery, pneumology), which is a
more challenging task. Also, as a case study, we only analyze data from one
hospital. Nevertheless, we approach a much larger dataset (with 26462 records
collected from 2000 to 2013), when compared with the datasets used by some of
the mentioned works (e.g., Merom et al. [11] included data from 1369 patients
and Oliveira et al. [13] analyzed only 401 records). In addition, the attributes
that we adopt (described in Sect. 3) were defined by a hospital expert’s medical
panel and are commonly available at the hospitalization process. Most of the
proposed attributes (e.g., gender, age, episode type, medical specialty) were also
adopted by the literature (as shown in Sect. 3.3). Moreover, in contrast with
several literature works [6,7,16], we do not perform a classification task, which
requires defining a priori which are the interesting LOS class intervals. Instead,
we adopt the more informative pure regression approach, which predicts the
actual number of LOS days and not classes.
 Using Data Mining for Prediction of Hospital Length of Stay 153

3 CRISP-DM Methodology
In this section, we describe the main procedures and decisions performed when
following the first five phases of the CRISP-DM methodology for LOS predic-
tion of a Portuguese hospital, namely: business understanding (Sect. 3.1), data
understanding (Sect. 3.2), data preparation (Sect. 3.3), modeling (Sect. 3.4) and
evaluation (Sect. 3.5).

3.1 Business Understanding

The prediction of LOS is inserted within the wider problem of hospital admission
scheduling, where there is a pressure to increase the availability of beds for new
patients. In this particular Hospital, most patients come from the emergency
department and from the region of Lisbon. The goal was set in terms of predicting
LOS using regression models, thus favoring predictions that are closer to the
target values. As a baseline business objective (to determine if there is success),
we defined a coefficient of determination with a minimum value of R2 = 0.6,
which often corresponds to a reasonable regression.
In terms of software, we adopted open source tools, using structured query
language (SQL) to extract data from the hospital database and the R tool for the
data analysis (http://www.r-project-org). In particular, we adopt the rminer
package [17], for applying the DM regression models (i.e., AP, MR, DT, ANN,
SVM and RF) and sensitive analysis methods.

3.2 Data Understanding

The data was collected between October 2000 and March 2013. During this
period, a total of 26462 inpatient episodes were stored, related with 15253
patients and associated with the distinct hospital medical specialties.
The selection of relevant data attributes for LOS prediction was performed
by an expert medical panel. The panel was composed with 9 physicians from
different medical specialties (e.g., internal medicine, general surgery, gynecol-
ogy). The panel presented a total of 28 attributes that were considered related
with LOS and that were analyzed in the data preparation phase (Table 1). The
first seven rows of Table 1 are related with the patient’s characteristics while
the remaining rows are related with the inpatient clinical process. The descrip-
tion column of the table contains in brackets the attribute type (date, nominal,
ordinal or numeric), as found in the original hospital database.

3.3 Data Preparation

In this phase, a substantial effort was performed using a semi-automated app-
roach to preprocess the data. In particular, the R tool was adopted to perform
an exploratory data analysis (e.g., histograms and box plots) and preprocess the
original dataset. The processing involved the operations of cleaning, discarding
redundant attributes, handling missing values and attribute transformations.
154 N. Caetano et al.

Table 1. List of attributes related with LOS prediction (attributes used by the regres-
sion models are in bold.

Name Description (attribute type)

Patient Characteristics:
Sex Patient gender (nominal)
Date of Birth Date of birth (date)
Age Age at the time of admission (numeric)
Country Residence country (nominal)
Residence Place of residence (nominal)
Education Educational attainment (ordinal)
Marital Status Marital status (nominal)
Inpatient clinical process:
Initial Diagnosis Initial diagnosis description (ordinal)
Episode Type Patient type of episode (nominal)
Inpatient Service Physical inpatient service (nominal)
Medical Specialty Patient medical specialty (nominal)
Origin Episode Type Origin episode type of hospitalization (nominal)
Admission Request Date Date for hospitalization admission request (date)
Admission Date Hospital admission date (date)
Admission Year Hospital admission year (ordinal)
Admission Month Hospital admission month (ordinal)
Admission Day Hospital admission day of week (ordinal)
Admission Hour Hospital admission hour (date)
Main Procedure Main procedure description (nominal)
Main Diagnosis Main diagnosis description (ordinal)
Physician ID Identification of the physician responsible for the
internment (nominal)
Discharge Destination Patient destination after hospital discharge (nominal)
Discharge Date Hospital discharge date (date)
Discharge Hour Hospital discharge hour (date)
GDH Homogeneous group diagnosis code (numeric)
Treatment Clinic codification for procedures, treatments and
diseases (ordinal)
GCD Great diagnostic category (ordinal)
Previous Admissions Number of previous patient admissions (numeric)
Target attribute:
LOS Length Of Stay (numeric)
 Using Data Mining for Prediction of Hospital Length of Stay 155

During the exploratory data analysis step, a few outliers were first detected
and then confirmed by the Physicians. The respective records were cleaned: one
LOS with 2294, an age of 207 and 29 entries related with a virtual medical
specialty, used only for testing the functionalities of the hospital database. After
cleaning, the database contained 26431 records.
Then, fourteen attributes from Table 1 were discarded in the variable selec-
tion analysis step: Date of Birth (reason: reflected in Age); Country (99 %
patients were from Portugal); Residence (30 % of missing values, very large num-
ber of nominal levels); Admission Request Date (48 % of missing values, reflected
in Admission Date); Admission Date (reflected in Admission Month, Day, Hour
and LOS); admission year (not considered relevant); Physician ID (19 % of miss-
ing values and large number of 156 nominal levels); Initial Diagnosis (63 % of
missing values); and attributes not known at the patient’s hospital admission
process (i.e., GDH, GCD, Treatment, Discharge Destination, Date and Hour).
The remaining 14 attributes (bold in Table 1) were used as input variables of
the regression models (Sect. 3.4). As shown in Table 2, all input attributes pro-
posed in this study (except for Marital Status) were also used in previous works,
which is a clear indication that the selected attributes (bold in Table 1) can have
a potential predictive LOS value. In particular, there are three input variables
(gender, age and main diagnosis) that were used in five or more studies.
Next, missing values were replaced by using the hotdeck method [18], which
substitutes a missing value by the value found in the most similar case. In particu-
lar, the rminer package uses a 1-nearest neighbor applied over all attributes with
full values to find the closest example [17]. The following attributes were affected
by this operation: Education (11771 missing values), Marital Status (10046 val-
ues), Main Procedure (19407 values) and Main Diagnosis (19268 values).

Table 2. List of input attributes proposed in this work and that were also used in the
literature.

Attribute Name Previous LOS studies that adopted this attribute

Sex [7, 11–14, 16]
Age [7, 12–16]
Education [7]
Episode Type [7, 15]
Inpatient Service [7]
Medical Specialty [6, 16]
Origin Episode Type [11]
Admission Month [14]
Admission Day [14]
Admission Hour [14]
Main Procedure [7, 12]
Main Diagnosis [6, 7, 11, 13, 14, 16]
Previous Admissions [7]
156 N. Caetano et al.

Finally, several attributes were transformed, to facilitate the modeling stage.

To reduce skewness and improve symmetry of the underlying variable distrib-
ution, the logarithm transform y=ln(x +1) was applied to the Previous Admis-
sions and LOS variables. This is a popular transformation that often improves
regression results for right-skewed variables [19]. Also, the Admission Hour vari-
able was standardized to include only 24 levels. Moreover, the values of nominal
attributes with a large number of levels were recoded/standardized to reduce the
number of levels: Education (transformed from 14 to 6 levels), Main Procedure
(from hundreds of values to 16 levels) and Main Diagnosis (from hundreds to
19 levels). Finally, using medical knowledge, we transformed the Age numeric
attribute into 5 ordinal classes: A - lower than 15 years; B - between 15 and 44;
C - between 45 and 64; D - between 65 and 84; and E - equal or higher than 85.

3.4 Modeling
Due to its importance, in the last decades, several methods have been proposed
for regression, such as DT, ANN, SVM and RF [9]. In this phase, we tested six
regression methods, as implemented in the rminer package [17]: AP, MR, DT,
ANN, SVM and RF.
The AP is a naive model that consists in predicting the same average LOS (y,
as found in the training set) and is used as baseline method for the comparison.
The DT is a branching structure that represents a set of rules, distinguishing
values in a hierarchical form.
The MR is a classical statistical model defined by the equation:
I

ŷ = β0 + βi xi (1)
i=1

where β0 , . . . , βi are the set of parameters to be adjusted, usually by applying

an ordinary least squares (OLS) algorithm.
ANN is based in the popular multilayer perceptron, with one hidden layer
of H hidden nodes and logistic activation functions, while the output node uses
the linear function. Since ANN training is not optimal, the final solution is
dependent of the choice of starting weights. To solve this issue, rminer first
trains Nr different networks and then uses an ensemble of these networks such
that the final output is set in terms of the average of the distinct Nr individual
predictions.
The SVM model performs a nonlinear transformation to the input space by
adopting the popular Gaussian kernel. SVM regression is achieved under the
commonly used -insensitive loss function. Under this setup, the SVM perfor-
mance is affected by three parameters: γ – Gaussian kernel parameter;  and C
– a trade-off between fitting the errors and the flatness of the mapping. Finally,
RF is an ensemble of T unpruned DT, where each tree is based on a random
feature selection with up to m features from bootstrap training samples. The
RF predictions are built by averaging the outputs of T trees. RF is a substantial
modification of bagging (fit of several models to bootstrap samples of training
 Using Data Mining for Prediction of Hospital Length of Stay 157

data) and on many problems RF performance is similar to boosting, while being

more simpler to train and tune [9].
The rminer package full implementation details can be found in [17]. Under
this package, before fitting the MR, ANN and SVM models, the input data is
first standardized to a zero mean and one standard deviation [9]. Except for the
hyperparameters of the most complex methods (ANN, SVM and RF), rminer
adopts the default parameters of the learning algorithms, such as: MR and ANN
– BFGS algorithm, as implemented in nnet package; DT - CART algorithm,
as implemented in the rpart package; SVM - sequential minimal optimization
algorithm, as implemented in the kernlab package; and RF - Breiman’s random
forest algorithm, as implemented in the randomForest package.
In this work, we set Nr = 3 for the ANN ensemble. Also, heuristics were
adopted to set two of the
 three SVM hyperparameters [17]: C = 3 (for standard-
ized data) and  = 3σy log (N )/N , where σy denotes the standard deviation of
the predictions given by a 3-nearest neighbor and N is the dataset size. For RF,
we adopted the default T = 500 value. For the most complex methods, rminer
uses grid search to select the best hyperparameter values: H for ANN, γ for
SVM and m for RF. In this work, the grid method searches ten values for each
hyperparameter (H ∈{0,1,...,9}; γ ∈ {2−15 , 2−13 , ..., 23 }; and m ∈ {1, 2, ..., 10}).
During the grid search, the absolute error is measured over a validation set
(with 33 % of the training data). The configuration that corresponds to the low-
est validation error is selected. Finally, the selected model is retrained with all
training data.
The method used for estimating the predictive performance of a model was
a 5-fold cross-validation, which divides the data into 5 partitions of equal size.
In each 5-fold iteration, a given subset is used as test set (to measure predictive
capability) and the remaining data is used for training (to fit the model). To
assure statistical robustness, 20 runs of this 5-fold procedure were applied to all
methods. For demonstration purposes, we present here a portion of the R/rminer
code used to test the RF model:

library(rminer) # load the library

# read the data:
d=read.table("data.csv",header=T,sep=",")
# execute 20 runs of 5-fold using RF:
M=mining(LOS~.,data=d,Runs=20, method=c("kfold",5),
model="randomforest", search="heuristic10")
# save the results into a file:
savemining(M,"rf.results")

3.5 Evaluation

To evaluate the predictions, three regression metrics were selected, the coefficient
of determination (R2 ), Mean Absolute Error (MAE) and Root Mean Squared
Error (RMSE), which can be computed as [20]:
158 N. Caetano et al.
N 2
i=1 (yi −ŷi )
R2 = 1 −N 2
i −y i )
i=1 (y
N
M AE = 1/N
 × i=1 |yi − ŷi | (2)
N
i=1 (yi − ŷi ) /N
RM SE = 2

where N denotes the number of predictions (test set size), yi is the target value
for example i, y i is the average of the target values in the test set and ŷi is the
predicted value for example i.
R2 is a popular regression metric that is scale independent, the higher the
better, with the ideal model presenting a value of 1.0. The lower the RMSE and
MAE values, the better the predictions. When compared with MAE, RMSE is
more sensitive to extreme errors. The Regression Error Characteristic (REC)
curve is useful to compare several regression methods in a single graph [21]. The
REC curve plots the error tolerance on the x-axis versus the percentage of points
predicted within the tolerance on the y-axis.
Table 3 presents the regression predictive results, in terms of the average of
the 20 runs of the 5-fold cross-validation evaluation scheme. From Table 3, it is
clear that the best results were obtained by the RF model, which outperforms
other DM models for all three error metrics. A pairwise t-student statistical
test, with a 95 % confidence level, was applied, confirming that the differences
are significant (i.e., p-value<0.05) when comparing RF with other methods. We
emphasize that a very good R2 value was achieved (0.813), much higher than
the minimum success value of 0.6 set in Sect. 3.1.

Table 3. Predictive results (average of 20 runs, as measured over test data; best values
in bold).

Metrics
Method R2 MAE RMSE
AP 0.000 0.861 1.085
MR 0.641 0.446 0.650
DT 0.622 0.415 0.667
ANN 0.736 0.340 0.558
SVM 0.745 0.296 0.547
RF 0.813 0.224 0.469
 


statistically significant under a
pairwise comparison with other
methods.

The REC analysis, shown in Fig. 1, also confirms the RF as the best predictive
model, presenting always a higher accuracy (y-axis) for any admitted absolute
tolerance value (x-axis). For instance, for a tolerance of 0.5 (at the logarithm
transform scale), the RF correctly predicts 85.4 % of the test set examples. The
REC results are further complemented in Table 4, which compares the accuracy
 Using Data Mining for Prediction of Hospital Length of Stay 159

1.0
0.8
0.6
Accuracy
0.4

RF
MR
DT
0.2

ANN
AP
SVM
0.0

0.0 0.5 1.0 1.5 2.0 2.5 3.0

Absolute deviation

Fig. 1. REC curves for all tested models.

Table 4. RF vs SVM accuracy for some absolute deviation values (average of 20 runs,
best values in bold).

Absolute Deviation SVM Accuracy RF Accuracy

0.0 0.0 % 15.6%
0.1 50.1 % 61.3%
0.2 63.3 % 70.9%
0.3 70.9 % 77.2%
0.4 76.3 % 81.8%
0.5 80.6 % 85.4%
0.6 84.0 % 88.2%
0.7 86.7 % 90.3%
0.8 89.0 % 91.9%
0.9 80.8 % 93.3%
1.0 92.3 % 94.4%

of the best two models (RF and SVM) for eleven absolute deviation values within
the range [0,1]. The table confirms the superiority of the RF model, which always
presents higher accuracy values, with a difference that ranges from 2.1 % points
(for a tolerance of 1.0) to 15.6 (for a tolerance of 0.0).
The quality of the predictions for the RF model can also be seen on Fig. 2,
which plots the observed (x-axis) versus the predicted values (y-axis). In the
160 N. Caetano et al.

4
B

3
Predicted
2
1

A
within 0.5 tolerance
other points
ideal method
0

0 1 2 3 4 5 6 7
Observed

Fig. 2. Observed versus predicted RF values.

plot, values within the 0.5 tolerance are shown with solid circles (85.4 % of the
examples), values outside the tolerance range are plotted with the + symbol
and the diagonal dashed line denotes the performance of the ideal prediction
method. It should be noted that the observed (target) values do not cover the
full space of LOS values, as shown in Fig. 2. This is an interesting property of
this problem domain that probably explains the improved performance of RF
when compared with other methods, since ensemble methods (such as RF) tend
to be useful when the sample data does not cover the tuple space properly. The
large diversity of learners (i.e., T =500 unpruned trees) can minimize this issue,
since each learner can specialize into a distinct region of the input space.
It should be noted that the presented predicted results were computed over
the logarithm transform scale (see Sect. 3.3). In Fig. 2 and within a 0.5 tolerance
(solid circles), the predictions are above the origin point (point A, x =0) and
below the right upper observed values (point B, x=4.2). This means that at
the normal scale (x , using the inverse of the logarithm transform), the RF
model error is capable of correctly predicting 85.4 % of the examples with a real
maximum error that ranges from 0.7 days (point A, x = 0) to 26.0 days (point
B, x = 65.7 days).
When compared with DT and MR, the ANN, SVM and RF data-driven
models are difficult to be interpreted by humans. Yet, sensitivity analysis and
visualization techniques can be used to open these complex models [10]. The
procedure works by analyzing the responses of a model when a given input
is varied though its domain. By analyzing the sensitivity response changes, it
 Using Data Mining for Prediction of Hospital Length of Stay 161

0.00 0.05 0.10 0.15 0.20 0.25 0.30

Episode.Type
Inpatient.Service
Medical.Specialty
Main.Procedure
Main.Diagnosis
Previous.Admissions
Education
Marital.Status
Admission.Hour
Age
Admission.Day
Origin.Episode.Type
Sex
Admission.Month

0.00 0.05 0.10 0.15 0.20 0.25 0.30

Fig. 3. Input importance bar plot for the RF model.

is possible to measure input relevance (higher changes denote a more relevant

input) and average impact of an input in the model. The former can be shown
using an input importance bar plot and the latter by plotting a Variable Effect
Characteristic (VEC) line curve or segments.
To extract explanatory knowledge from the RF model and open the black-box,
we applied the Data-Based Sensitivity Analysis (DSA) method, as implemented in
the Importance function of the rminer package. DSA has the advantage of being
a fast method that can measure the overall influence of a particular input, includ-
ing its iterations with other inputs (Cortez and Embrechts, 2013). The DSA algo-
rithm was executed over the RF model fit with all data. The obtained sensitivity
responses were first used to rank the RF inputs, according to their relevancy in
the predictive model (Fig. 3). Then, the average effects of the most relevant inputs
were analyzed using VEC line segments (Figs. 4, 5 and 6).
The input importance bar plot (Fig. 3) ranks the Episode Type (30.1 %
impact) as the most relevant attribute, followed by Inpatient Service (12.3 %) and
Medical Specialty (10.1 %). Overall, the bar plot shows a much greater influence
of the inpatient clinical process attributes (e.g., Episode Type, Medical Spe-
cialty, Previous Admissions) when compared with the patients’ characteristics
(e.g., Education, Sex). This is an interesting outcome for hospital managers. In
the next paragraphs, we detail the particular influence of the top three inputs
by analyzing their VEC line segments.
162 N. Caetano et al.

1.6
1.4
1.2
1.0
0.8
0.6
0.4

Ambulatory Internment

Fig. 4. VEC line segments, showing the average influence of the Episode Type
(x-axis) on the RF model output (y-axis).

Figure 4 shows the global influence of the most relevant input (Episode Type),
which is a nominal attribute with two classes. The VEC line segments clearly
confirm that the ambulatory type (scheduled admission, typically involving a
1 day LOS) is related with an average lower LOS (0.1 in the logarithm transform
scale, 0.1 days in the normal scale) when compared with the internment type
(1.58 in the logarithm scale, 3.9 days).
Next, we analyze the average influence of the Inpatient service (Fig. 5). The
greatest LOS is associated with five services: medicine, average LOS of 1.45,
corresponding to 3.3 days at the normal scale; orthopedics, average of 1.39, cor-
responding to 3.0 days; specialties, average of 1.37, corresponding to 2.9 days;
surgery, average of 1.36, corresponding to 2.9 days; and pneumology, average of
1.32, corresponding to 2.7 days.
Finally, we analyze the third most relevant attribute, the Medical Specialty
(Fig. 6). The internal medicine is related with the highest average LOS (1.64,
corresponding to 4.2 days). The second highest average LOS (1.50, corresponding
to 3.5 days) is related with orthopedics. Two Medical Specialty values are ranked
third in terms of their average effect on LOS: general surgery and urology, both
related with an average LOS of 1.40, corresponding to 3.1 days.
These results were shown to hospital specialists and a positive feedback was
obtained, confirming meaningful and interesting effects between these attributes
and the average expected LOS. Moreover, we would like to stress that the top
four relevant attributes were also in agreement with several literature works
 Using Data Mining for Prediction of Hospital Length of Stay 163

Medicine
Orthopedics

1.4
Specialties
Surgery

Pneumology
1.3
1.2
1.1

OR Plastic Surgery
Plastic Surgery

Fig. 5. VEC line segments, showing the average influence of the Inpatient service (x-
axis) on the RF model output (y-axis).
1.65

Internal Medicine
1.60
1.55

Orthopedics
1.50
1.45

Surgery Urology
1.40

Plastic Surgery
1.35

Ophthalmology
1.30

Otorhinolaryngology

Fig. 6. VEC line segments, showing the average influence of the Medical specialty
(x-axis) on the RF model output (y-axis).
164 N. Caetano et al.

(Table 2). For instance, the Episode Type was proposed by [7,15], the Inpatient
Service was adopted by [7], the Medical Specialty was used in [6,16], and the
Main Procedure was approached in [7,12]. We also highlight that Education and
Marital Status are two of the proposed attributes that are scarcely adopted by
the literature. Yet, these attributes were ranked at 7th and 8th place, with a total
contribution of around 10 % of the input importance (Fig. 3), thus confirming
their added value for the LOS prediction model.

4 Conclusions

Due to advances in Information Technology, hospitals are collecting vast amounts

of data related with their clinical information systems. All this data can hold
valuable knowledge. The development of the Data Mining (DM) field has created
new exciting possibilities for extracting such clinical knowledge, in what is known
as medical data mining. This work describes an implementation of a medical data
mining project approach based on the CRISP-DM methodology. In particular,
a DM approach was applied to estimate the Length Of Stay (LOS) of patients
at their hospital admission process. We analyzed recent real-world data from a
Portuguese hospital, involving a large dataset that included 26462 records (from
15253 patients) and an initial set of 28 attributes (as defined by a medical panel).
The DM approach was guided by the popular CRISP-DM methodology,
under a regression approach. After the data preparation phase of CRISP-DM, a
cleaned dataset (without outliers and missing data) was achieved, with a total
of 26431 records, 14 input attributes and the LOS target. During the mod-
eling phase, six distinct regression models were explored: Average Prediction
(AP), Multiple Regression (MR), Decision Tree (DT), Artificial Neural Network
(ANN) ensemble, Support Vector Machine (SVM) and Random Forest (RF).
These models were compared and tested under a robust evaluation scheme that
used 20 runs of a 5-fold cross-validation. Finally, at the evaluation phase of
CRISP-DM, the obtained results were analyzed.
The best prediction performance was achieved by the RF model, which
presents a very good coefficient of determination value of R2 = 0.81 and that is
21 % points higher than the minimum threshold of R2 = 0.60 set in the business
understanding phase. A Regression Error Characteristic (REC) curve analysis
revealed that the RF model can correctly predict 85.4 % of the examples under a
tolerance deviation that ranges from 0.7 (for observed LOS of 0 days) to 26 days
(for observed LOS of 66 days). At the same evaluation phase of CRISP-DM,
sensitivity analysis and visualization techniques were used to extract explana-
tory knowledge from the best predictive model (RF). The sensitivity analysis
revealed a high impact of inpatient clinical process attributes, instead of the
patient’s characteristics. In effect, the top three influential input attributes were:
the hospital Episode Type, the Inpatient Service where the patient is hospital-
ized and the associated Medical Specialty. Moreover, the average influence of
each of these input attributes in the prediction model has been detailed by using
a Variable Effect Characteristic (VEC) analysis. Such analysis has confirmed
 Using Data Mining for Prediction of Hospital Length of Stay 165

that several input values associated with high LOS, such as: ‘internment” (for
Episode Type), “medicine” (for Inpatient Service) and “internal medicine” (for
Medical Specialty).
The obtained DM predictive and explanatory knowledge results were consid-
ered credible by the hospital specialists and are valuable for hospital managers.
By having access to better estimates of the LOS that is more likely to occur in the
future and which factors affect such estimates, hospital managers can make more
informed decisions. Such informed decisions can lead to a better planning of the
hospital resources, resulting in a better hospital management performance, with
an increase in the number of available beds for new admissions and reduction of
surgical waiting lists.
In the future, we intend to address the implementation phase of CRISP-
DM by testing the obtained data-driven model in a real-environment (e.g., by
designing a friendly interface to query the RF model). After some time, this
would allow us to obtain additional feedback from the hospital managers and
also enrich the datasets by gathering more examples. The proposed approach
has also the potential to predict well LOS using data from other hospitals, since
we address generic LOS and use 14 variables that are easily available at the
hospitalization process.

Acknowledgments. We wish to thank the physicians that participated in this study

for their valuable feedback. Also, we would like to thank the anonymous reviewers for
their helpful suggestions. The work of P. Cortez has been supported by FCT – Fundação
para a Ciência e Tecnologia within the Project Scope: PEst-OE/EEI/UI0319/2014.

References
1. Fayyad, U., Piatetsky-Shapiro, G., Smyth, P.: Advances in Knowledge Discovery
and Data Mining. MIT Press, Cambridge (1996)
2. Cios, K., Moore, G.: Uniqueness of Medical Data Mining. Artificial Intelligence in
Medicine 26(1–2), 1–24 (2002)
3. Silva, Á., Cortez, P., Santos, M.F., Gomes, L., Neves, J.: Mortality assessment in
intensive care units via adverse events using artificial neural networks. Artif. Intell.
Med. 36(3), 223–234 (2006)
4. Silva, Á., Cortez, P., Santos, M.F., Gomes, L., Neves, J.: Rating organ failure via
adverse events using data mining in the intensive care unit. Artif. Intell. Med.
43(3), 179–193 (2008)
5. Chiusano, G., Staglianò, A., Basso, C., Verri, A.: Unsupervised tissue segmentation
from dynamic contrast-enhanced magnetic resonance imaging. Artif. Intell. Med.
61(1), 53–61 (2014)
6. Azari, A., Janeja, V.P., Mohseni, A.: Predicting hospital length of stay (phlos): a
multi-tiered data mining approach. In: 2012 IEEE 12th International Conference
on Data Mining Workshops (ICDMW), pp. 17–24. IEEE (2012)
7. Castillo, M.G.: Modelling patient length of stay in public hospitals in Mexico. PhD
thesis, University of Southampton (2012)
8. Clifton, C., Thuraisingham, B.: Emerging standards for data mining. Comput.
Stan. Interfaces 23(3), 187–193 (2001)
166 N. Caetano et al.

9. Hastie, T., Tibshirani, R., Friedman, J.: The Elements of Statistical Learning: Data
Mining, Inference, and Prediction, 2nd edn. Springer, New york (2008)
10. Cortez, P., Embrechts, M.J.: Using sensitivity analysis and visualization techniques
to open black box data mining models. Information Sciences 225, 1–17 (2013)
11. Merom, D., Shohat, T., Harari, G., Oren, M., Green, M.S.: Factors associated with
inappropriate hospitalization days in internal medicine wards in israel: a cross-
national survey. Int. J. Qual. Health Care 10(2), 155–162 (1998)
12. Abelha, F., Maia, P., Landeiro, N., Neves, A., Barros, H.: Determinants of outcome
in patients admitted to a surgical intensive care unit. Arq. Med. 21(5–6), 135–143
(2007)
13. Oliveira, A., Dias, O., Mello, M., Arajo, S., Dragosavac, D., Nucci, A., Falcão, A.:
Fatores associados à maior mortalidade e tempo de internação prolongado em uma
unidade de terapia intensiva de adultos. Rev. Bras. de Terapia Intensiva 22(3),
250–256 (2010)
14. Kalra, A.D., Fisher, R.S., Axelrod, P.: Decreased length of stay and cumulative
hospitalized days despite increased patient admissions and readmissions in an area
of urban poverty. J. Gen. Intern. Med. 25(9), 930–935 (2010)
15. Freitas, A., Silva-Costa, T., Lopes, F., Garcia-Lema, I., Teixeira-Pinto, A., Brazdil,
P., Costa-Pereira, A.: Factors influencing hospital high length of stay outliers. BMC
Health Serv. Res. 12(265), 1–10 (2012)
16. Sheikh-Nia, S.: An Investigation of Standard and Ensemble Based Classification
Techniques for the Prediction of Hospitalization Duration. University of Guelph,
Ontario, Canada, Thesis for Master Science Degree (2012)
17. Cortez, P.: Data mining with neural networks and support vector machines using
the R/rminer Tool. In: Perner, P. (ed.) ICDM 2010. LNCS, vol. 6171, pp. 572–583.
Springer, Heidelberg (2010)
18. Brown, M., Kros, J.: Data mining and the impact of missing data. Ind. Manage.
Data Syst. 103(8), 611–621 (2003)
19. Menard, S.: Applied logistic regression analysis, vol. 106. Sage, Thousand Oaks
(2002)
20. Witten, I.H., Frank, E., Hall, M.: Data Mining: Practical Machine Learning Tools
and Techniques, 3rd edn. Morgan Kaufmann, San Franscico (2011)
21. Bi, J., Bennett, K.: Regression error characteristic curves. In: Fawcett, T., Mishra,
N. (eds.) Proceedings of 20th International Conference on Machine Learning
(ICML). AAAI Press, Washington DC, USA (2003)
 Applying Artificial Immune Systems for Deploying
Router Nodes in Wireless Networks
in the Presence of Obstacles

Pedro Henrique Gouvêa Coelho ✉ , Jorge Luís Machado do Amaral,

( )

José Franco Machado do Amaral, Luciane Fernanda de Arruda Barreira,

and Adriano Valladão de Barros
State University of Rio de Janeiro, FEN/DETEL,
Rua São Francisco Xavier, 524, Sala 5001E, Maracanã, Rio de Janeiro,
RJ 20550-900, Brazil
{phcoelho,jamaral,franco}@uerj.br,
{lucianebarreira,adriano_vbarros}@yahoo.com.br

Abstract. This work deals with the deploying of router nodes using artificial
immune systems techniques, particularly for industrial applications of wireless
sensor networks. Possible scenarios include configurations with sensors blocked
by obstacles. These nodes make possible the transmission of data from sensors
to the gateway in order to meet criteria especially those that lead to a low degree
of failure and reducing the number of retransmissions by the routers. These
criteria can be set individually or in groups, associated with weights. Router nodes
deploying is accomplished in two phases, the first uses immune networks concepts
and the second employs potential fields ideas for deploying the routers in such
way that critical sensors attract them while obstacles and other routers repel them.
A large number of case studies were considered from which some representative
ones were selected to illustrate the method, for different configurations in the
presence of obstacles.

Keywords: Artificial immune systems · Node positioning and wireless sensor

networks

1 Introduction

Data transmission in wireless technology has grown dramatically in society. The wire‐
less technology has taken over the world and the field of industrial automation is no
exception. Main advantages are reduced installation time of devices, no need of cabling
structure, cost saving projects, infrastructure savings, device configuration flexibility,
cost savings in installation, flexibility in changing the existing architectures, possibility
of installing sensors in hard-to-access locations and others.
Safety, reliability, availability, robustness and performance are of paramount impor‐
tance in the area of industrial automation. The network cannot be sensitive to interference
nor stop operation because of an equipment failure, nor can have high latency in data
transmission and ensure that information is not misplaced [1, 3].

© Springer International Publishing Switzerland 2015

J. Cordeiro et al. (Eds.): ICEIS 2014, LNBIP 227, pp. 167–183, 2015.
DOI: 10.1007/978-3-319-22348-3_10
168 P.H.G. Coelho et al.

In industrial automation environment, data transmission in a wireless network faces

the problem of interference generated by other electrical equipment, such as walkie-
talkies, other wireless communication networks and electrical equipment, moving obsta‐
cles (trucks, cranes, etc.) and fixed ones(buildings, pipelines, tanks, etc.). In an attempt
to minimize these effects, frequency scattering techniques and mesh or tree topologies
are used, in which a message can be transmitted from one node to another with the aid
of other nodes, which act as intermediate routers, directing messages to other nodes until
it reaches its final destination. This allows the network to get a longer range and to be
nearly fault tolerant, because if an intermediate node fails or cannot receive a message,
that message could be routed to another node. However, a mesh network also requires
careful placement of these intermediate nodes, since they are responsible for doing the
forwarding of the data generated by the sensor nodes in the network to the gateway
directly or indirectly, through hops. Those intermediate nodes are responsible for
meeting the criteria of safety, reliability and robustness of the network and are also of
paramount importance in the forwarding of data transmission. They could leave part or
all the network dead, if they display any fault [7]. Most solutions to the routers placement
solve this problem with optimization algorithms that minimize the number of inter‐
mediate router nodes to meet the criteria for coverage, network connectivity and
longevity of the network and data fidelity [8, 9].
This work proposes to solve this problem using Artificial Immune Networks,
based on the human immune system. The algorithms based on immune networks
have very desirable characteristics in the solution of this problem, among which we
can mention: scalability, self-organization, learning ability and continuous treat‐
ment of noisy data [10].
The present work is divided into four sections. Section 2 discusses briefly artifi‐
cial immune systems. Section 3 presents the application of artificial immune systems
to the node positioning problem and Sect. 4 discusses case studies results including
obstacles and conclusions.

2 Immune Systems Foundations

Artificial immune systems (AISs) are models based on natural immune systems
which protect the human body from a large number of pathogens or antigens [13].
Due to these characteristics the AISs are potentially suitable for solving problems
related to computer security and they inherit from natural immune systems the prop‐
erties of uniqueness, distributed sensing, learning and memory efficiency. In fact, the
immune system is unique to each individual. The detectors used by the immune
system are small, efficient and highly distributed and are not subject to centralized
control. Moreover, it is not necessary that every pathogen is fully detected, because
the immune system is more flexible, there is a compromise between the resources
used in the protection and breadth in coverage.
Anomaly detection is another important feature, since it allows the immune system
to detect and respond to pathogens (agents that cause diseases) for which they have not
previously been found. The immune system is able to learn the structures of pathogens
 Applying Artificial Immune Systems for Deploying Router Nodes 169

and remember these structures so that future responses to these agents are faster. In
summary, these features make the immune system scalable, robust and flexible. The
immune system uses distributed detection to distinguish the elements of the organism
itself, the self, and foreign to the body, the non-self. The detection of the non-self is a
difficult task because its number, of the order of 1016, is much superior to the number of
self-patterns, around 106, taking place in a highly distributed environment. It should be
also noted that all these actions occur while the living organism must continue in oper‐
ation and the available resources are scarce.
Cells that perform the detection or recognition of pathogens in the acquired or adap‐
tive immune system are called lymphocytes that recognize pathogens joining them. The
antigens are detected when a molecular bond is established between the pathogens and
the receptors present on the surface of lymphocytes. A given receiver will not be able
to join all antigens. A lymphocyte has approximately 100,000 receptors on its surface
which however have the same structure, and therefore can only join with structurally
related epitopes (the site on an antigen at which a specific antibody becomes attached).
Such epitopes define a similarity subset of epitopes which lymphocytes can detect.
The number of receivers that can join the pathogens defines the affinity of a lympho‐
cyte to a certain antigen. Lymphocytes can only be activated by an antigen if their
affinities exceed a certain threshold. As this threshold increases, the number of epitopes
types capable of activating a lymphocyte decreases, i.e., the similarity subset becomes
smaller. A receiver may be obtained by randomly recombining possible elements (from
the memory of the immune system), producing a large number of possible combinations
indicating a wide range in the structure of the receptors. Although it is possible to
generate approximately 1015 receptor types, the number present at a given instant of time
is much smaller, in the range of 108 to 1012 [12]. The detection is approximate, since it
is a difficult task to evolve structures that are complementary to receptor epitopes for
which the organism has never encountered before. If an exact complementarity was
needed, the chance of a random lymphocyte epitope join a random would be very small.
An important consequence of that approximate detection is that one single lymphocyte
is capable of detecting a subset of epitopes, which implies that a smaller number of
lymphocytes is required for protection against a wide variety of possible antigens.

2.1 Metaphors of the Immune System

The main algorithms that implement the artificial immune systems were developed from
metaphors of the immune system: the mechanism of negative selection, the theory of
immune network and the clonal selection principle.
The function of the negative selection mechanism is to provide tolerance to self-
cells, namely those belonging to the organism. Thus, the immune system gains the
ability to detect unknown antigens and not react to the body’s own cells. During the
generation of T-cells, which are cells produced in the bone marrow, receptors are
generated by a pseudo-random process of genetic arrangement. Later on, they
undergo a maturation mechanism in the thymus, called negative selection, in which
T cells that react to body proteins are destroyed. Thus, only cells that do not connect
170 P.H.G. Coelho et al.

to the body proteins can leave the thymus. The T cells, known as mature cells,
circulate in the body for immune functions and to protect it against antigens.
The theory of immune system network considers several important aspects like the
combination of antibodies with the antigens for the early elimination of the antigens.
Each antibody has its own antigenic determinant, called idiotope. In this context, Jerne
[14] proposed the Immune Network Theory to describe the activity of lymphocytes in
an alternative way. According to Jerne [14], the antibodies and lymphocytes do not act
alone, but the immune system keeps a network of B cells for antigen recognition. These
cells can stimulate and inhibit each other in various ways, leading to stabilization of the
network. Two B cells are connected if they share an affinity above a certain threshold
and the strength of this connection is directly proportional to the affinity they share.
The clonal selection principle describes the basic features of an immune response to
an antigenic stimulus, and ensures that only cells that recognize the antigen are selected
to proliferate. The daughter cells are copies or clones of their parents and are subject to
a process of mutation with high rates, called somatic hypermutation. In the clonal selec‐
tion the removal of daughter cells are performed, and these cells have receptors that
respond to the body’s own proteins as well as the most suitable mature cell proliferation,
i.e., those with a greater affinity to the antigens [5].

3 Router Nodes Positioning

Router Nodes positioning has been addressed in the literature by several researchers.
Cannons et al. [4] proposed an algorithm for positioning router nodes and determine
which router will relay the information from each sensor. Gersho and Gray [6]
proposed one to promote the reliability of wireless sensors communication network,
minimizing the average probability of sensor transmission error. Shi et al. [11]
proposed a positioning algorithm of multi-router nodes to minimize energy consump‐
tion for data transmission in mobile ad hoc network (MANET - Mobile Ad Hoc
Network). The problem was modeled as an optimization clustering problem. The
suggested algorithm to solve the problem uses heuristic methods based on the k-
means algorithm. Costa and Amaral [2] described an approach for router nodes place‐
ment based on genetic algorithm which minimizes the number of nodes required for
network routers, decreasing the amount critical nodes for all involved devices and the
number of hops of the transmitted messages.
The use of wireless sensor network in industrial automation is still a matter of concern
with respect to the data reliability and security by users. Thus, an appropriate node
positioning is of paramount importance for the wireless network to meet safety, relia‐
bility and efficiency criteria.
Positioning of nodes is a difficult task, because one should take into account all the
obstacles and interference present in an industrial environment. The gateway as well as the
sensors generally have a fixed position near the control room. But the placement of router
nodes, which are responsible for routing the data, generated by the sensors network to the
gateway directly or indirectly, is determined by the characteristics of the network.
 Applying Artificial Immune Systems for Deploying Router Nodes 171

The main characteristics of wireless sensor networks for industrial automation differ
from traditional ones by the following aspects: The maximum number of sensors in a
traditional wireless network is on the order of millions while automation wireless
networks is on the order of tens to hundreds; The network reliability and latency are
essential and fundamental factors for network wireless automation. To determine the
number of router nodes and define the position in the network, some important aspects
in industrial automation should be considered. It should be guaranteed:
(1) Redundant paths so that the system be node fault-tolerant;
(2) Full connectivity between nodes, both sensors and routers, so that each node of
the network can be connected to all the others exploring the collaborative role
of routers;
(3) Node energy efficiency such that no node is overwhelmed with many relaying
information from the sensors;
(4) Low-latency system for better efficiency in response time;
(5) Combined attributes for industrial processes to avoid accidents due to, for example,
high monitored process temperature;
(6) Self-organization ability, i.e. the ability of the network to reorganize the retrans‐
mission of data paths when a new sensor is added to the network or when a
sensor stops working due to lack of power or a problem in wireless communi‐
cation channel.
All these factors must be met, always taking into consideration the prime factor
security: the fault tolerance. In the end of the router nodes placement, the network of
wireless sensors applied to industrial automation should be robust, reliable, scalable
and self-organizing.
The positioning of router nodes in industrial wireless sensor networks is a complex
and critical task to the network operation. It is through the final position of routers that
one can determine how reliable, safe, affordable and robust the network is.
In the application of immune systems to router nodes positioning reported in this
work, B cells that make up the immune network will be composed by a set of sensor
nodes and a set of router nodes. The sensor nodes are located in places where the
plant instrumentation is required. These nodes have fixed coordinates, i.e. they
cannot be moved. For security to be guaranteed it is necessary to have redundant
paths between these nodes and the gateway. The set of router nodes will be added to
allow redundant paths. The position of these nodes will be changed during the
process of obtaining the final network. The stimulation of the B cells, corresponding
to the set of routers, is defined by the affinity degree among B cells in the training
of the network. In this work, the role of the antigen is viewed more broadly as the
entity that stimulates B cells. Thus, the function of the antigen takes into considera‐
tion possible missing paths to critical sensors, the number of times that a router is
used and its proximity to sensors. The modeling of B cells affinity is the weighted sum
of the three criteria that the positioning of each router will answer. The criteria are:
fault degree of each router, number of times each router is used depending on the
path and number of sensor nodes neighboring to each router.
172 P.H.G. Coelho et al.

3.1 The Proposed Router Nodes Algorithm

In the proposed algorithm, process dynamics can be divided into two processes: network
pruning and cloning, and node mutation of the network routers.
In the pruning process, np router nodes that during a certain time failed to become
useful to the network will be removed from it. The cloning process is responsible for
generating nc clones of router nodes that were over stimulated. The clones may suffer
mutations of two kinds:
(i) Hypermutation – for positioning new elements in the network which are inversely
proportional to the degree of stimulation of the router node selected;
(ii) Net Mutation – for positioning the new information into the network in order to
assure the new clones are neighbors of the selected clone [16].
After the inclusion of the new router nodes, a stop condition is performed. If the
condition is not met, all routers undergo an action of repulsive forces, generated by
obstacles and routers for other nodes, followed by attractive forces created by critical
sensor nodes. Those critical nodes are the ones that do not meet the minimum number
of paths necessary to reach the gateway. The actions of repelling potential fields have
the function of driving them away from obstacles, to allow direct line of sight for the
router network nodes to increase the reliability of transmission and also increase the
distance among the routers to increase network coverage. On the other hand, the attrac‐
tive potential fields attract routers to critical sensors, easing the formation of redundant
paths among sensors and the gateway. After the action of potential fields, from the new
positioning of routers, a new network is established and the procedure continues until
the stopping criterion is met.
The algorithm proposed in this work deals with a procedure based on artificial
immune networks, which solves the problem of positioning the router nodes so that
every sensor device is able to communicate with the gateway directly and or indirectly
by redundant paths.
Figure 1 shows the main modules of the algorithm. The first module is called immune
network, and the second, is called potential fields (i.e. positioning module) containing
elements used in positioning sensor networks using potential fields [15].
The immune network module performs an algorithm that can be described by the
following steps:
– Creation: Creation of an initial set of B cells to form a network.
– Evaluation: Determination of the B cells affinity to calculate their stimulation.
– Pruning: Performs the resource management and remove cells that are without
resources from the network.
– Selection: Selects the most stimulated B cells to be cloned.
– Cloning: Generates a set of clones from the most stimulated B cells.
– Mutation: Does the mutation of cloned cells.
In the stage of creation, an initial set of routers is randomly generated to initiate the
process of obtaining the network, and the user can specify how many routers to place it
initially.
 Applying Artificial Immune Systems for Deploying Router Nodes 173

Fig. 1. The proposed algorithm.

In the evaluation phase, a network which is represented by a graph is formed with

sensor nodes and router nodes. From this graph, values of several variables are obtained
that will be used to calculate the affinity. Examples of such variables are the number of
paths that exist between each sensor and the gateway, the number of times that a router
is used on the formed paths, etc.
It should be stressed that the affinity value is calculated for each router and comprises
three parts. The first part provides the degree of fault of each network router - this affinity
is the most important of all. It defines the value or importance each router has in the
network configuration. This is done as follows: a router is removed from the network,
and the number of paths that remain active for the sensors send information for the
gateway is evaluated. If the number of active paths remaining after the node removal is
small, the router node needs another nearby router to reduce their degree of fault.
Furthermore, if the node suffers battery discharge or hardware problems, other paths
to relay information should be guaranteed until the problem is solved.
The second part relates to the number of times that each router is used in paths that
relay the information from the sensors to the gateway. The greater the number of times
it is used, more important is that router.
The third part relates the number of sensor nodes neighboring to each router – one
can say that the more sensor nodes neighbors, the greater the likelihood that it will
become part of the way that the sensor needs to transmit your message to the gateway.
174 P.H.G. Coelho et al.

4 Case Studies and Conclusions

Case studies were simulated in a 1 × 1 square scenario. The cloning procedure consid‐
ered that only the router with higher affinity would be selected to produce three clones
in each generation. For each case study 10 experiments were carried out that demon‐
strated the algorithm’s ability to create at least two redundant paths to get the information
from any sensor to the gateway.
Two set of results will be presented. The first set does not consider obstacles which
are treated in the second set.

4.1 Case Studies with no Obstacles

Two configurations were considered to demonstrate the functionality of the proposed

algorithm. The configurations used in the simulation were motivated by oil and gas
refinery automation applications.
The first one called PosA consists of five network nodes, where node 1 is the gateway
and nodes 2, 3, 4 and 5 are fixed sensors. The gateway has direct line of sight with all
the network nodes as shown in Fig. 2.

PosA Configuration
1

0.9

0.8 2

0.7

0.6
Distance

0.5 5 1 3

0.4

0.3

0.2 4

0.1

0
0 0.2 0.4 0.6 0.8 1
Distance

Fig. 2. Sensors and gateway for PosA configuration. Node 1 is the gateway and nodes 2 to 5 are
the sensors.
 Applying Artificial Immune Systems for Deploying Router Nodes 175

The second one (PosB) considers a network with nine nodes, where node one is the
gateway and the others are fixed sensors. As in configuration PosA, PosB has direct line
of sight with all the network nodes as shown in Fig. 3. For both configurations it will be
considered that there is no connectivity among them, i.e. the distance between them will
be greater than their operating range.
For the case study simulations considered, the goal is to get any two paths for each
sensor to transmit the monitored sensor data to the gateway node. The operating range
for both cases is 0.2, i.e. for both configurations there is no connectivity between any
sensor and the gateway.
Table 1 describes the parameters used in the case study 1. After completion of ten
experiments, the best network configuration can be seen in Fig. 4, and the consolidations
of the tests are shown in Table 2.

PosB Configuration
1

0.9

0.8 9 2 3

0.7

0.6
Distance

0.5 8 1 4

0.4

0.3

0.2 7 6 5

0.1

0
0 0.2 0.4 0.6 0.8 1
Distance

Fig. 3. Sensors and gateway for PosB configuration. Node 1 is the gateway and nodes 2 to 9 are
the sensors.

Table 1. Case study 1. PosA configuration parameters.

Simulation Parameters Values Method
No. of generations 50 –
Initial no. of routers 10 –
Mutation operator as indicated in Fig. 1 – Hypermutation (Type 1)
Affinity – Fault degree
176 P.H.G. Coelho et al.

Table 2. Network performance for case study 1. PosA configuration.

Network Min. Average Max. Standard deviation
Number of nodes 13 13.7 15 0.67
Number of routers 8 8.7 10 0.67
Number of critical sensors 0 0 0 –
Number a router is used 1 2.1 3 0.57

Figure 4 also shows that one of the paths from sensor node 3 to the gateway shows
three jumps (3-7-8-1) i.e. the information had to be relayed by two routers to reach the
gateway node. Regarding the degree of fault, all eight routers have 20 % degree of fault
tolerance. This means that 80 % of the paths from the sensors to the gateway continue
to exist even after the removal of a node. With respect to the maximum number of routers
used in terms of paths, the router node 8 is used twice in the paths 3-8-1 and 3-7-8-1.
Consequently, this router will have a greater battery consumption than the others, which
could make it stop working and be disconnected from the network. But even if that
happens, there will still be a path (3-7-10-1) for node 3 to communicate to the
gateway.

Case Study 1
1

0.9

0.8 2

0.7 7
10
9
0.6
Distance

0.5 5 11 1 8 3

0.4 12
13

0.3 6

0.2 4

0.1

0
0 0.2 0.4 0.6 0.8 1
Distance

Fig. 4. Best router nodes positioning for case study 1 in PosA configuration. Node 1 is the
gateway, nodes 2 to 5 are the sensors, and nodes 6 to 13 are the router nodes.
 Applying Artificial Immune Systems for Deploying Router Nodes 177

Table 3. Case study 2. PosB configuration parameters.

Simulation Parameters Values Method

No. of generations 15 –
Initial no. of routers 3 –
Mutation Operator (as Net mutation (Type 2)
indicated in Fig. 1)
Affinity – Fault degree, number of times a router is used
and number of neighbor sensors

Case study 2 considers configuration PosB for the sensors and gateway. Table 3
shows the parameters used in the case study 2 simulations. The goal is still obtain at
least two paths for each sensor and gateway but now the affinity criteria consider fault
degree, number of times a router is used and number of neighbor sensors. After ten
experiments the best network configuration is shown in Fig. 5 and the network perform‐
ance is seen in Table 4. Table 4 indicates that even using a low number of initial routers
the algorithm was able to reach a positioning result meeting the goals and avoided again
critical nodes.
Figure 5 also shows that node 3 in the path 3-15-17-13-1 features four hops to the
gateway. That means that the information sent by these devices will be delayed when
received by the gateway node, since it will need to be relayed through three intermediate
nodes. Regarding to the degree of fault, the intermediate node 22 has 22 % degree of
fault, and all the other routers have an index less than 22 %. Thus if node 22 is lost for
device failure or end of battery, it results that information sent by sensor 5 will not reach
the gateway. Regarding to the maximum number of routers used in terms of paths, router
nodes 10 and 13 are used three times in the paths 3-15-17-13-1, 3-26-13-1, 4-13-1, 9
-18-12-10-1, 9-24-10-1 and 2-11-10-1.That means that these devices will have their
lifetime reduced because their high levels of retransmission. As far as the number of
sensors to neighboring routers is concerned, routers 15, 19, 22 and 26 can relay the data
sent by two sensors, and the sensors are also used with relays. This makes these sensors
and routers consume more power, and as a result, battery runs out sooner.

Table 4. Network performance for case study 2. PosB configuration.

Network Min. Average Max. Standard deviation
Number of nodes 23 25.3 28 1.49
Number of routers 14 16.3 19 1.49
Number of critical sensors 0 0 0 –
Number a router is used 3 3.7 5 0.67
178 P.H.G. Coelho et al.

Case Study 2
1

0.9

0.8 9 2 3
18 15
0.7 11
24 25 26
17
0.6
12
Distance

10 13
0.5 8 1 4
16
20
0.4 21
19
14 23
0.3
22
0.2 7 6 5

0.1

0
0 0.2 0.4 0.6 0.8 1
Distance

Fig. 5. Best router nodes positioning for case study 2 in PosB configuration. Node 1 is the
gateway, nodes 2 to 9 are the sensors, and nodes 9 to 26 are the router nodes.

4.2 Case Studies with Obstacles

Two configurations with obstacles were considered to illustrate the proposed router
nodes positioning algorithm in environments with obstacles.
The first configuration (PosC) comprises two circular obstacles with a radius of 0.1,
and five nodes, in which node 1 is the gateway and the others are sensor nodes. Initially,
the gateway has not direct line of sight with sensor nodes 3 and 5 and is not connected,
i.e. out of range, to any of the network nodes, as depicted in Fig. 6.
The second configuration (PosD) has eight obstacles: three circular ones have radius
of 0.05, another circular one has radius 0.15 and four rectangular obstacles with different
sizes. Besides, the gateway is node 1 and nodes 2 to 8 are the seven sensor nodes.
Initially, the gateway has not direct line of sight to any of the sensor nodes and is not
connected to any network node as it is out range to the other nodes. Moreover, sensor
nodes have also not a direct line of sight with each other and are also not connected as
they are out of range with each other too. Figure 7 shows the PosD configuration.
In this section, case studies 3 and 4 are considered using the configurations PosC
and PosD.
For case study 3, the network configuration is cross shaped, the operating range of
the network nodes is 0.2 and the positioning procedure lead to two disjoint paths for the
sensors send data to the gateway.
 Applying Artificial Immune Systems for Deploying Router Nodes 179

PosC Configuration
1

0.9

0.8 2

0.7

0.6
Distance

0.5 5 1 3

0.4

0.3

0.2 4

0.1

0
0 0.2 0.4 0.6 0.8 1
Distance

Fig. 6. Sensors and gateway for PosC configuration. Node 1 is the gateway and nodes 2 to 5 are
the sensors.

Case study 4 uses configuration PosD and considers the same operating range as in
the case study 3, 0.2, and now three disjoint paths are required.
Tables 5 and 7 show the used parameters for case studies 3 and 4 respectively.

Table 5. Case study 3. PosC configuration parameters.

Simulation Parameters Values Method
No. of generations 30 –
Initial no. of routers 10 –
Mutation Operator (as indicated Net mutation (Type 2)
in Fig. 1)
Affinity – Fault degree, number of times a router is used and
number of neighbor sensors

Table 6. Network performance for case study 3. PosC configuration.

Network Min. Average Max. Standard deviation
Number of nodes 19 19.9 22 0.99
Number of routers 14 14.9 17 0.99
Number of critical sensors 0 0 0 –
Number a router is used 2 2 2 0
180 P.H.G. Coelho et al.

PosD Configuration
1

0.9 2

0.8
1
0.7

0.6 4
Distance

0.5

0.4
5
0.3 6

0.2
7 8
0.1 3

0
0 0.2 0.4 0.6 0.8 1
Distance

Fig. 7. Sensors and gateway for PosD configuration. Node 1 is the gateway and nodes 2 to 8 are
the sensors.

Table 7. Case study 4. PosD configuration parameters.

Simulation Parameters Values Method
No. of generations 100 –
Initial no. of routers 10 –
Mutation Operator (as indicated Net mutation (Type 2)
in Fig. 1)
Affinity – Fault degree, number of times a router is used and
number of neighbor sensors

Figure 8 shows the best configuration obtained from the 10 experiments. Table 6
shows the network performance for case study 3.
It can be seen in Fig. 8 that the sensor nodes 3 and 5 in the paths 3-16-17-20-1,
3-19-7-10-1, 5-13-11-18-1 and 5-12-14-15-1 show four jumps to the gateway. This
means the data sent by these devices suffer a delay when received by the gateway, since
it will need to be relayed through three intermediate nodes. With respect to the degree
of fault, the intermediate nodes 7, 10, 11, 12, 13, 14, 16, 17 and 19 have a 30 % degree
of fault, and the other router nodes have an index lower than 30 %. So 70 % of the paths
from the sensors to the gateway, continue to exist even after the removal of a node.
 Applying Artificial Immune Systems for Deploying Router Nodes 181

Case Study 3
1

0.9

0.8 2

0.7 14 11
13 9 17 16
15
0.6 12 18 20
Distance

0.5 5 1 3

0.4
8 6 19
10 7
0.3

0.2 4

0.1

0
0 0.2 0.4 0.6 0.8 1
Distance

Fig. 8. Best router nodes positioning for case study 3 in PosC configuration. Node 1 is the
gateway, nodes 2 to 5 are the sensors, and nodes 6 to 20 are the router nodes.

Figure 9 shows the best configuration out of ten experiments for case study 4 and
Table 8 shows the network performance for case study 4.
Figure 9 indicates that for sensor nodes 3 and 6, the paths 3-20-22-24-7-40-63-53-36-1,
3-50-49-61-4-52-15-56-39-1 and 6-58-61-4-60-31-15-37-26-1 show nine hops to the
gateway. This means that the data sent by these devices suffer a delay in the gateway, since
it will need to be relayed by eight intermediate nodes. With respect to the degree of fault, the
router node 32 have 21 % degree of fault, and the other router nodes have an index lower than
21 %. This means that 79 % of the paths from the sensors are still present even after a node
removal.

Table 8. Network performance for case study 4. PosD configuration.

Network Min. Average Max. Standard deviation
Number of nodes 59 60.50 63 1,18
Number of routers 51 52.50 55 1.18
Number of critical sensors 0 0 0 –
Number a router is used 5 5,4 8 0.97
182 P.H.G. Coelho et al.

Case Study 4
1 14 17 21 13
27 29
0.9 2 16 41
23 45
0.8
35
1 44
0.7 26 59 62
57 39 42
60 25 34
0.6 4 38 11 36
Distance

33 52 31 37 56 28
0.5 15
53 18
61 32 63 55 30
0.4 46 47
19
5
58 51 43
0.3 49 6 54 12
40 9
48
0.2 10
7 8
50
0.1 3

0 20 22 24
0 0.2 0.4 0.6 0.8 1
Distance

Fig. 9. Best router nodes positioning for case study 4 in PosD configuration. Node 1 is the
gateway, nodes 2 to 8 are the sensors, and nodes 9 to 63 are the router nodes.

This work proposed a positioning algorithm for router nodes in wireless network using
immune systems techniques. The algorithm creates redundant paths to the data collected
by the sensors to be sent to the gateway by any two or more paths, meeting the criteria of
degree of failure, the number of retransmission by routers and number of sensors to neigh‐
boring routers. The algorithm allows each criterion is enabled at a time or that they be
combined with weights. The affinity function, which works as an objective function, is
multi-objective, so several other objectives could be jointly considered.

References

1. Zheng, J., Lee, M.J.: A comprehensive performance study of IEEE 802.15.4. Sensor Network
Operations, Chapter 4, pp. 218–237. IEEE Press, Wiley InterScience (2006)
2. Costa, M.S., Amaral, J. L.M.: A tool for node positioning analysis in wireless networks for
industrial automation. In: XVIII Automation Brazilian Congress. Bonito, pp. 1521–1527, in
portuguese (2010)
3. Moyne, J.R., Tilbury, D.M.: The emergence of industrial control, diagnostics, and safety data.
Proc. IEEE 95(1), 29–47 (2007)
4. Cannons, J., Milstein, L.B., Zeger, K.: An algorithm for wireless relay placement. IEEE Trans.
Wirel. Commun. 8(11), 5564–5574 (2006)
 Applying Artificial Immune Systems for Deploying Router Nodes 183

5. Coelho, P.H.G., Amaral, J.L.M., Amaral, J.F.M., Barreira, L.F.A., Barros, A.V.: Deploying
nodes for industrial wireless networks by artificial immune systems techniques. In: 15th
International Conference on Enterprise Information Systems, Angers, France (2013)
6. Gersho, A., Gray, R.M.: Vector quantization and signal compression. Kluwer Academic
Publishers, Norwell (1992)
7. Hoffert, J., Klues, K., Orjih, O.: Configuring the IEEE 802.15.4 MAC Layer for Single-sink
Wireless Sensor Network Applications, Technical report (2007). http://www.dre.
vanderbilt.edu/~jhoffert/802_15_4_Eval_Report.pdf
8. Youssef, W., Younis, M.: Intelligent gateways placement for reduced data latency in wireless
sensor networks. In: ICC 2007 International Conference on Communications, Glasgow, pp.
3805–3810 (2007)
9. Molina, G., Alba, E., Talbi, E.G.: Optimal sensor network layout using multi objective
metaheuristics. J. Univers. Computer Sci. 15(15), 2549–2565 (2008)
10. Coelho, P.H.G., Amaral, J.L.M., Amaral, J.F.M.: Node positioning in industrial plants
wireless networks. In: WES 2012 International Conference on Communications, Rio Grande,
R.S., in Portuguese (2012)
11. Shi, Y., Jia, F., Hai-Tao, Y.: An improved router placement algorithm base on energy efficient
strategy for wireless network. In: ISECS International Colloquium on Computing,
Communication, Control and Management (CCCM 2009), pp. 421–423 (2009)
12. Silva, L.N.C.: Immune Engineering: Development and Application of Computational Tools
Inspired by Artificial Immune Systems, Ph. D. thesis, State University of Campinas,
Campinas, in Portuguese (2001)
13. Amaral, J.L.M.: Artificial Immune Systems Applied to Fault Detection, Ph. D. thesis,
Pontifical Catholic University of Rio de Janeiro, Rio de Janeiro, in Portuguese (2006)
14. Jerne, N.K.: Towards a Network Theory of the Immune System. Ann. Immunol. (Inst.
Pasteur) 125C, 373–389 (1974)
15. Howard, A., Mataric, M.J., Sukhatme, G.S.: Mobile sensor network deployment using
potential fields: a distributed, scalable solution to the area coverage problem heuristics. In:
Asama, H., Arai, T., Fukuda, T., Hasegawa, T. (eds.) DARS 2002, 6th International
Symposium on Distributed Autonomous Robotics Systems, vol. 2002, pp. 299–308. Springer,
Japan (2002)
16. Poduri, S., Pattem, S., Krishnamachari, B., Sukhatme, G.: Controlled Deployments of Sensor
Networks. In Press (2006)
Information Systems Analysis
and Specification
 Performance Evaluation of Aspect-Oriented
Programming Weavers

Michel S. Soares1(B) , Marcelo A. Maia2 , and Rodrigo F.G. Silva2

1
Computing Department, Federal University of Sergipe, São Cristóvão, Brazil
[email protected]
2
Computing Faculty, Federal University of Uberlândia, Uberlândia, Brazil

Abstract. Aspect-oriented programming (AOP) was proposed with the

purpose of improving software modularization by treating crosscutting
concerns. Since its introduction, there is no consensus about the impact
on performance of the use of AOP techniques to deal with crosscutting
concerns. This article explores further the evaluation of performance by
proposing a systematic literature review to find out how performance
is affected by the introduction of aspects. Then, an experiment is per-
formed to find results about the performance of AOP and weavers. This
experiment showed the assessment of several versions of an application.
According to this study, the difference concerning resource consumption
through variation of weavers can be considered irrelevant considering
web applications.

Keywords: Aspect-oriented programming · Systematic literature

review · Crosscuting concerns · Performance · MVC framework

1 Introduction

Aspect Oriented Programming (AOP) [1] was proposed as an attempt to deal

with crosscutting concerns aiming to improve modularization. AOP has gained
importance since its introduction to implement crosscutting concerns, with vary-
ing degree of success [2–5]. Within AOP, crosscutting concerns are implemented
as aspects and are further weaved into code. The way aspects are weaved into
code may affect performance as the weaving process introduces new code to the
original programs.
The impact on performance, caused by AOP techniques, has motivated pre-
vious works in scientific literature. Liu [6] showed that the aspect-oriented
approach does not have significant effect on performance, and that in some
cases, aspect-oriented software even outperform the non-aspect one. Addition-
ally, introduction of a large number of join points does not have significant effect
on performance. Remko [7] assessed the performance effects between programs
created by a weaver and a hand-coded version. This work came to the conclusion
that simple advices give no real performance penalties, but the more sophisti-
cated advices are, the slower they become. Kirsten [8] compared the four leading
c Springer International Publishing Switzerland 2015
J. Cordeiro et al. (Eds.): ICEIS 2014, LNBIP 227, pp. 187–203, 2015.
DOI: 10.1007/978-3-319-22348-3 11
188 M.S. Soares et al.

AOP tools at the time (2004) and, when it comes to performance, he postu-
lated that, in general, code with aspects performs similarly to that of a purely
object-oriented solution, where the crosscutting code is scattered throughout the
system.
The use of AOP to implement crosscutting concerns and its impact on per-
formance is the motivation for this study. First, a systematic literature review is
performed. The goal of this systematic review is to understand the extent of the
impact of AOP on the performance of software systems, if there is an impact.
The purpose of the experiment described following the review is to find results
about performance and aspect-oriented implementations. Considering the fact
that AOP techniques add additional complexity in software in order to treat
cross-cutting concerns, this addition can generate an overhead in software exe-
cution, leading to impact on performance.
Considered factors or independent variables for evaluating performance in
AOP software include the weaver, the type of weaving, the type of advice, the
number of join points, the size measured in lines of code, and the number of
classes to be loaded in the load-time weaving process. In order to evaluate the
impact of AOP techniques on performance, this article proposes the evaluation
of a major factor that may impact these techniques: the type of weaving. The
factors number of lines of code, types of advices, number of joinpoints and number
of classes to be loaded during load-time weaving will be considered in future
research. Therefore, the hypotheses of this article is that changing the weaving
process sensitize the outcomes.

2 Protocol for Systematic Review

The main question that motivated the systematic literature review [9] we con-
ducted in this paper is: Does the use of aspect-oriented techniques to implement
crosscutting concerns impact software performance ? A derived research ques-
tion is “If the impact exists, how meaningful is it ? ”. The answer to both ques-
tions could help developers to reason about the feasibility of the use of AOP
techniques to handle crosscutting concerns on architectures where performance
is itself a concern. The systematic review started by searching in a number of
software engineering conferences and journals. The search was performed con-
sidering publications in the past 6 years. The chosen conferences were: AOSD
(International Conference on Aspect-Oriented Software Development) and ICSE
(International Conference on Software Engineering). The chosen journals were:
JSS (Journal of Systems and Software), IST (Information and Software Technol-
ogy), SCP (Science of Computer Programming), TSE IEEE (IEEE Transactions
on Software Engineering), TOSEM (ACM Transactions on Software Engineer-
ing Methodology). ENTCS (Electronic Notes in Theoretical Computer Science),
which can be considered a series, was also included.
The search string was (“Aspect-oriented programming” AND “perfor-
mance”). The search has retrieved 338 papers. From these 338 papers, a sub-
selection has been made with the purpose of separating those relating AOP with
 Performance Evaluation of Aspect-Oriented Programming Weavers 189

Table 1. Search results and selected papers.

Publication Retrieved papers Relevant papers Selected papers Data source

JSS 38 2 1 ScienceDirect
IST 32 10 5 ScienceDirect
SCP 32 2 1 ScienceDirect
TSE 1 1 1 IEEExplorer
TOSEM 21 3 1 ACM Digital Library
ENTCS 26 3 1 ScienceDirect
AOSD 127 9 3 ACM Digital Library
ICSE 61 2 2 ACM Digital Library
Total 338 32 15

Table 2. Selected papers.

Venue/year Reference
1 ICSE/07 [10]
2 SCP/08 [11]
3 IST/09 [12]
4 ENTCS/09 [13]
5 IST/09 [14]
6 ICSE/09 [15]
7 AOSD/09 [16]
8 JSS/10 [17]
9 ACM/10 [18]
10 IST/10 [19]
11 AOSD/10 [20]
12 IST/10 [21]
13 AOSD/10 [22]
14 IEEE/12 [4]
15 IST/12 [23]

any performance metrics. In a first step, 32 papers were selected and classified
as relevant. For the first selection, the title, the keywords and the abstract were
read. If the subject was pertinent to AOP and performance, the introduction
and the conclusion were read as well. In case of doubt about the relevance of the
paper, specific keywords were searched in the paper, such as aspect, crosscutting
and performance. There were also relevant papers which used other terms, includ-
ing cost, payload and overhead when considering assessment of performance of
some AOP technique, and in those cases, they were selected too.
190 M.S. Soares et al.

In a second step, of the 32 relevant papers, only those ones which assess
the performance of implementation of some crosscutting concern were selected
to be fully read. As a result, 15 papers were selected in total. Several types of
concerns have been classified by the papers as crosscutting concerns, even though
some of them were domain specific. However, papers which had crosscutting
concerns implemented through some AOP technique but which did not consider
any assessment of the used technique(s), or this assessment was incomplete, were
discarded. The summary of the filtering process can be seen in Table 1 and the
final selection of papers is presented in Table 2.

3 Systematic Review Results

All 15 selected papers were fully read for the evaluation. The selected papers were
evaluated based on two sets of criteria: Application Type and Performance.

3.1 Application Type Criteria

The first set of criteria concerns about Application Type and encompass the
following metrics: number of assessed studies, lines of code (size, in LOCs),
original programming language (Original PL), aspect programming language
(Aspect PL) and application domain.
The application type is related to the type of application of the case studies
or experiments which have been assessed by the papers. The following types were
retrieved from the papers: Middleware, Web Service, Embedded, Platform, Sys-
tem or Application, Language or Extension (Language) and Framework. Cases
where their case studies were described as Monitoring Systems were classified
as System or Application. Papers which did not mention the application type of
their experiments have been classified under the closest definition of these ones
already mentioned. The number of assessed studies indicates only those studies
that were implemented by some AOP technique and were assessed by some kind
of metric.
The application domain includes: e-commerce, industrial application, Office,
Bank and Generic. Cases where there is no specific domain, for example a toolkit
or a language extension, were classified as Generic. Some papers, mostly in appli-
cation type, did not mention the application domain and were also classified by
proximity.
The summary of studies is presented in Table 3. Cases where no metric was
presented or in which it was not possible to identify were classified as not avail-
able (NA).

3.2 Performance Criteria

The second set of criteria concerns Performance. Four metrics were extracted
from the papers: weaving type, implemented crosscutting concerns, used perfor-
mance method, and performance overhead.
 Performance Evaluation of Aspect-Oriented Programming Weavers 191

Table 3. Summary of studies.

The weaving type indicates the type of weaving performed by the studies
in the papers. Two main kinds were considered: Compile-time and Runtime
weaving.
Several kinds of crosscutting concerns were retrieved from the papers. There
were cases where the crosscutting concerns were domain specific. Papers which
treated only one concern in the study prevailed, but there were cases where more
than one concern was considered, one of them domain specific [4].
The performance methods retrieved were the measurements of running or
execution time (ext), business operations per second (bos), average memory over-
head (avmo), CPU usage (cpu), qualitative observation of the overall execution
(obs), parsing time (pat), and average number of method calls per second (met).
Some papers presented more than one assessed variable. In these cases, when
there was performance reduction, the considered measurements were based on
the worst case. Some cases related the performance overhead as negligible (negl).
The results of the performance assessment performed by the papers are
presented in Table 4. In the Performance Overhead column, the “+” and the
“−” signs means decrease and increase in performance, respectively. If a sign is
followed by negl, it means that the paper reported a degradation or gain in
performance, but this result is negligible according to the authors.
192 M.S. Soares et al.

Table 4. Performance analysis of studies.

3.3 On the Target Applications

From the first set of criteria, related to Application type, it is possible to con-
clude that most papers, 10 out of 15, assessed only one study or experiment.
However, only four of them showed the LOC or size of their assessed stud-
ies. The two studies that have evaluated more systems, evaluated three small-
scale systems (at most 50KLOC or 190 Kb). The larger evaluated system had
118KLOC. One hypothesis for such lack of large scale studies is that AOP is not
extensively adopted such as OOP or procedural programming. Therefore, the
low adoption from the community restricts the availability of large systems for
experimentation.
 Performance Evaluation of Aspect-Oriented Programming Weavers 193

The prevailing application type was System or Application. That is reasonable

to expect because in general this kind of applications are more frequent and
more accessible. The prevailing application domain was Industrial applications
followed by applications with no specific domain, hereby classified as Generic. We
can observe that there is reasonable variability in terms of Application Type and
Application domain. We could observe that in Middleware software the overhead
was negligible. In the category System or Application, there is a tendency of more
impact in the performance.
The LOC seems not to influence because the larger studies systems had negli-
gible impact on execution time and CPU performance. The Application Domain
also seems not to influence the performance because of the high variation in
the results. The application domain did not present a clear influence in perfor-
mance. The Industrial domain, which has the larger number of studies, had also
presented negligible and positive impact in performance.
Finally, concerning the implemented crosscutting concerns in the applica-
tions, there was no prevailing concern in the studies, and surprisingly, none of
the studies implemented common concerns such as Logging or Exception Han-
dling. This can be an indicative that the studied cases were not representative
in terms of typical aspect-oriented software.

3.4 On the Used Programming Languages

The prevailing original programming language was Java with 11 out of 15 studies
and the prevailing aspect programming language was AspectJ with 6 out of 15
studies. JBoss AOP was present only in two cases, and Spring AOP was not
used in any of them. Considering the impact of the programming language in
performance, we can observe that the original programming language that has
significant number of studies is Java, but there was no clear indication that Java
is an influence factor. In the same way, AspectJ, which is the prevailing aspect
language, has shown no direct influence on the performance because it presented
either negligible or positive impact in performance. Although only two studies
were carried out with JBoss AOP, both studies have shown a positive impact in
performance.

3.5 On the Type of Weaving

From the second set of criteria, it is possible to conclude that run-time is the most
common weaving type process presented by papers in the experiments. One of the
reasons for this choice, instead of compile-time weaving, is the fact that runtime
weaving allows aspects to be added to the base program dynamically, which
is better for a context-aware adaptation of the applications [13]. Some papers
not only used the runtime weaving process but also extended it to adequate
the process to their studies. Also concerning the weaving process, the papers
in general postulated that runtime weaving requires more effort at runtime,
impacting on performance, but no proofs about this assumption were found in
this research.
194 M.S. Soares et al.

3.6 On the Experimental Setting of Reviewed Papers

Papers assessed their experiments in different ways, but the prevailing perfor-
mance metric was measurement of execution time (ext). The performance over-
head varied according to a set of variables such as the used approach, imple-
mented concern, the aspect programming language, weaving type process and
the used aspect weaver.
The workload is strongly influenced by the target application, which defines
several other sub-factors: the kind of join points, pointcuts and advices, the
ratio of occurrence of AOP constructs and the other non-AOP constructs, the
requirement of the application for specific type of weaving. Considering the space
for combination of levels for these factors, it is challenging (if not impossible)
to find a real world application (or a set of them) that can have all the possible
levels. Therefore, an alternative could be the design of a synthetic application
that could be use as a benchmark for performance evaluation of AOP techniques.
This benchmark would need to be meaningful to mimic real world scenarios and
would need to be comprehensive to guarantee that all important factors and
their respective levels would be considered.

4 Experimental Settings

In order to produce further evidences on the impact of the type of weavers in

the performance of AOP programs, we decided to conduct our own experiment.
This section explains the conditions of the experiment and how the experiment
was executed. The section is organized as follows. In Subsect. 4.1 the environ-
ment in which the experiment was performed is explained. Subsection 4.2 shows
the versions of the case study. Subsection 4.3 explains about the scenario of the
experiment. Subsection 4.4 shows the plan of the experiment designed for exe-
cuting the tests.

4.1 Environment

The original software in which the experiment was performed is a web application
system called SIGE. Designed in 2011, it has the purpose of automating the
management of information of academic and administrative units at Federal
University of Uberlândia. The System was developed in Java. SIGE uses two
frameworks, Struts (version 2.3.3) and Spring (version 3.0.0). SIGE also access
a DB2 database.
The crosscutting concern implemented by AOP is Logging, responsible for
logging some methods in service and DAO layers of the system for purposes
of auditing. In order to perform the experiment, a copy of the last version of
SIGE was produced and named as version V0. Version V0 has as size 297.915
LOC, considering files of the following extensions: java, jsp, js, html, xml and css.
From version V0, other versions were built. This work adopted the same idea
of inheritance from Object Orientation for extended versions. Each extended
 Performance Evaluation of Aspect-Oriented Programming Weavers 195

versions preserve the same properties of the parents and override the interested
properties so the proposed factors could be measured and analyzed.
Three weaving processes were considered regarding AOP. These are compile-
time weaving, load-time weaving, and runtime weaving. The class responsible for
implementing the Logging aspect is the AspectProfiler class. The adopted style
for enabling AOP into the class was the AspectJ annotation Style, provided
by the Spring AOP AspectJ support. Each method is composed of an advice
annotation where a pointcut expression is used to match the join points, which
is a method execution for all expressions. Despite the fact that AspectJ has
several definitions for join points, those used in the AspectProfiler class always
represent a method execution because of the Spring AOP framework restriction
which limits join points to be only method executions. This kind of join point was
adopted as default in order to enable comparisons among the different considered
weavers. This class was the only class used to implement the Logging aspect and
had its content varied in versions, according to the purpose of each version. Each
phase of the experiment explains how this content was modified for its purposes.
Two types of invocations were considered to perform the experiment, internal
and external invocations. External invocations are generated by the JMeter1 tool
while internal invocations are generated by a JUnit class, which are explained
in details in Sect. 4.4.

Fig. 1. Versions for the case study.

Concerning system settings, all tests were executed under the same hardware
and software conditions. The hardware was composed of one notebook with Intel
Core i5-2410M 2.30 GHz processor and 4 GB of RAM memory running Ubuntu
12.04 LTS operational system. The used Java version was 1.7.0. Before all test
measurements, memory and swap conditions were verified through Ubuntu Sys-
tem Monitor and were always under 55 % and 2 % respectively. The running
processes during applications were, besides regular Ubuntu processes, Eclipse,
System Monitor, LibreOffice calc, VisualVM and, in specific stages, JMeter.
1
https://jmeter.apache.org/.
196 M.S. Soares et al.

Despite the fact that only 4 assessed versions of the experiment access the data-
base, a running DB2 database was necessary for the application context to run.
The version of the running DB2 database was 9.7. Network was disabled during
tests in order to avoid CPU interruptions.

4.2 Versions
Figure 1 shows the design of proposed versions. The image is divided into versions
that will be compared to each other. It contains the original version (version V0),
one version with no AOP involved (version V1) and three other versions weaved
by three different weaving processes, which are compile-time weaving (CTW),
load-time weaving (LTW) and runtime weaving (RTW), respectively versions
V2, V3 and V4.

4.3 Scenario

The chosen scenario for the experiment was the user authentication scenario.
This scenario was chosen because it is composed of several method calls in dif-
ferent system layers and it does not present complex business rules.
The scenario was refactored so that the invocations do not access the data-
base. The reason for not using the database access is due to the fact that the
resources consumed by I/O operations would influence the measurements of the
logging aspect. Likewise, at the data access object (DAO) layer, all methods of
the scenario returns with the minimum of processing and simulate a successful
authentication. This refactoring did not alter the purpose of the methods or
the scenario sequence of operations. The resulting version, named V1, was used
to build all other versions where factors were varied so the outcomes could be
measured.

4.4 Plan of the Experiment

In order to measure the outcomes, a default experiment plan was made. This
plan is composed by a parameter with fixed value and parameters with variable
values, which are set in each system version. The parameter with fixed value is
the number of test executions, which defines how many times the test is executed
to generate outcomes data and was set to 15.
One of the parameters with variable value is the loopCount variable, present
in the JUnit test class for internal invocations and configured in the JMeter
experiment plan for external invocations. The loop count defines how many
times each internal invocations or HTTP request is executed on the scenario in
a test. It was initially set to 10.000. The criteria to choose is because this number
should be high enough to emphasize the small impact on performance caused
by some interested factor, but low enough to allow the assessment of the set of
invocations in an acceptable time. From a set of considered options, which were
1, 10, 100, 1.000, 10.000 and 100.000, the number 10.000 was the best choice for
 Performance Evaluation of Aspect-Oriented Programming Weavers 197

versions with disabled database and which used internal invocations as default
measuring method.
CPU, memory and time were measured in two forms, depending on the type
of invocation. For internal invocations, the execution of each test is launched
through Eclipse. A breakpoint is set at the beginning of test execution in JUnit
test class, where the start method of StopWatch is invoked and starts time count-
ing. This breakpoint is necessary for the process inspection through VisualVM
tool. When the JUnit running thread stops at the breakpoint, VisualVM is con-
figured to inspect only memory and CPU in the application process. After these
steps, breakpoint is released, the process run and the current aspect mechanism
perform over the scenario method executions. After the end of the process exe-
cution, StopWatch class calculate the elapsed time in milliseconds and shows it
at Eclipse console while VisualVM generates the real time report of memory and
CPU history. CPU and memory are manually retrieved and, the highest values
of the graphics, respectively, in percentage and megabytes (MB), were always
considered.
For external invocations, no breakpoint is necessary for VisualVM to inspect
CPU and memory. After the launch of tomcat through Eclipse debug, VisualVM
is capable of inspecting the running process of tomcat before the beginning of
the invocations. After configuring VisualVM to inspect CPU and memory on
tomcat process, the JMeter experiment plan is invoked. After the execution of
the test, JMeter report provides the time report while CPU and memory are
retrieved manually, as done for the internal invocations.

5 Design of the Experiment

The experiment plan was run in each phase with the proper parameters. This
Section explains the design of the experiment, showing the phases, their details
and purposes.

5.1 Versions of the Software

The purpose of this phase is the generation of versions containing different weav-
ing processes and serves as template to be used for building the other versions.
In this Section the building process of versions V1, V2, V3 and V4 is presented.

Version V1. The purpose of this version is to establish baselines of values

related to the dependent variables to be compared with values obtained in other
phases of the experiment.

Version V2. From version V1, version V2 was built. This version has the
purpose of assessing the impact on performance by adopting the compile-time
weaving process, provided by the AspectJ weaver. The results of this phase
are compared to the results of version V1, this way it is possible to assess if
198 M.S. Soares et al.

the weaver AspectJ sensitizes the dependent variables through the compile-time
weaving process. In this version all log generation messages were removed from
the scenario methods. The weaver was provided by the plugin of the AspectJ
Development Eclipse Tools. This plugin, when enabled, forces the project to
re-compile, thus injecting the aspect code bytecodes to the classes at compile
time.
The following parameters were set:

– Weaver: AspectJ compile-time weaver;

– Number of advices: 6;
– Type of advices: 2 before, 2 around, 2 after, having log messages invoked only
on before advices;
– PointCut Expressions: interception of all methods present in service and DAO
packages.

Version V3. From version V2, version V3 was built. This version has the
purpose of assessing the impact on performance by adopting the load-time
weaving process, provided by the AspectJ weaver. Like in version V2, the
results of this phase are compared to the results of version V1 in order to verify
if this weaver, using this weaving process, sensitizes the dependent variables. In
this version, the AspectJ plugin, which is responsible for compiling the aspects,
was disabled and the weaving type paremeter was overridden to the load-time
weaving process. The load-time weaving process is used in the context of Spring
Framework. AspectJ is responsible for weaving aspects into the application’s
class files as they are being loaded into JVM. From an aop.xml file created in
META-INF folder, which was added to the build path of the project, with the
necessary configuration to weave aspects of AspectProfiler class into the project
classes, this file specifies the loading of two classes to the JVM, and one DAO
class containing the methods of the scenario.

Version V4. From version V2, version V4 was built. This version has the
purpose of assess the impact on performance by adopting the runtime weaving
process, provided by the Spring AOP framework. Like in versions V3 and V2,
the results of this phase are compared to the results of version V1 in order
to verify if this weaver, using this weaving process, sensitizes the dependent
variables. In this version, the AspectJ capability provided by the AspectJ plugin
was disabled and the weaver overridden to Spring AOP.
This phase has the purpose of analyzing AOP together with two layers which
impacts performance significantly, which are the chain of mechanisms present in
the front end controller provided by the MVC framework, in this case Struts 2,
and the access to the database, here represented by DB2.
Previous tests reported that the time spent to assess each external invoca-
tion of this version varied between 6000 and 7000 ms. The time consumed to
assess the initial value for the loop count parameter, which is 10.000, would be
impracticable, which addresses the loop count value to be overridden to a new
 Performance Evaluation of Aspect-Oriented Programming Weavers 199

value, for less. Besides this difficulty on measuring time, the maximum values of
CPU and memory are manually selected on the chosen profiling tool, which is in
this case VisualVM, and the higher this value, the higher the difficulty to select
these maximums. However, in terms of assessing the impact on performance, the
lower this value, the less the impact on performance would be noticed, which
addressed this number to be 50.

6 Results
This Section presents the results of versions for each region of the map of versions
for the proposed factors, as follows.
The weaver factor was varied in versions of the software. The values of the
measures are shown in box plot graphs and grouped by dependent variable. In
Figs. 2, 3 and 4 graphs are represented containing the consumption of time, CPU
and memory for versions V1, V2, V3 and V4 respectively.
Results of versions V2, V3 and V4 were compared with results of version V1.
Table 5 shows this comparison to the averages of the dependent variables.

Fig. 2. Time consumption for versions V1, V2, V3 and V4.

Fig. 3. CPU consumption for versions V1, V2, V3 and V4.

200 M.S. Soares et al.

Fig. 4. Memory consumption for versions V1, V2, V3 and V4.

Table 5. Comparison between averages of dependent variables of versions V2, V3 and

V4 with V1.

Compared to V1 Time(%) CPU(%) Memory(%)

V2 +3, 41 −8, 5 +3, 24
V3 +7, 32 −4, 56 +30, 69
V4 +41, 40 −2, 58 +39, 93

7 Conclusions
The systematic review and the further analysis of the retrieved papers pre-
sented in this article show that there are few experiments concerning AOP and
performance in scientific literature. More specifically, too few experiments were
reported about the performance of AOP techniques when implementing crosscut-
ting concerns. From the results, it is clear that there is no prevailing implemented
concern in the studies. On the contrary, most of the implemented concerns were
domain specific. Most papers postulated that runtime weaving requires more
effort at run-time, impacting on performance. In order to produce further evi-
dences on the impact of the type of weavers in the performance of AOP programs,
we decided to conduct our own experiment. The experiment showed the assess-
ment of several versions of an application, where measurements were made with
and without the presence of heavy system layers in terms of resource consump-
tion which are the database and the whole chain of mechanisms of the MVC
framework.
The variation of weavers showed that the weaver is a factor which sensitizes
performance. When analyzed isolated, in other words, without the presence of
the resources consuming mentioned elements, weavers showed a performance
variation of about 40 % from one to another in two outcomes in certain circum-
stances where the pointcuts match were high.
Results of this experiment have shown that, in general, in applications which
have database calls and/or MVC frameworks, the impact on performance caused
by AOP solely could be considered to be negligible. However, this impact could
 Performance Evaluation of Aspect-Oriented Programming Weavers 201

be meaningful for applications without these layers and where AOP would be
widely used. According to this study, the difference concerning resource con-
sumption through variation of weavers can be considered irrelevant considering
common web applications, but nothing can be confirmed about embedded appli-
cations where resource consumption might be limited.
The analyzed crosscutting concern was Logging. However, the focus of the
analysis was on the AOP interception mechanism of each tool and not in the
AOP implementation of the crosscutting concern itself. The implementation
through AOP of crosscutting concerns by frameworks such as Spring could be
more studied and compared to manual implementations with regard to perfor-
mance. Future works could also study the performance of AOP techniques for
embedded applications, where hardware conditions are limited.
Threats to validity were considered as follows. The considered outcomes were
measured through three different tools. Time outcome was provided by Stop-
Watch class on internal invocations and displayed at Eclipse console, and by
JMeter on external invocations, displayed on a summarized report. CPU and
memory, however, were not provided directly by any tool. Differently from time,
which could be measured in absolute values, CPU and memory varies during the
tests. As VisualVM did not provide the average of each one in a time shift, the
used metric for obtaining these values was a manually selection of the maximum
values of the graphs generated by VisualVM for each one. Measurements where
the maximum values were far higher from the average values represent a threat
to validity, once that these top points could not represent the real behaviour of
the outcomes related to the application. Besides, mistakes on manual selection
could have influenced wrong results.
The external invocations generated by JMeter were configured to be sequen-
tial. Despite the fact that on real production environments calls to the appli-
cation generate multiple threads working simultaneously, sequential calls were
more proper for this experiment because they do not require a possible extra
effort to be scheduled. Besides, in order to achieve a possible impact on perfor-
mance, the number of calls needed to be high in most cases, considering that the
impact provoked by AOP instrumentation was discrete. This high number of calls
would not be supported by the web container in case these calls were generated
by multiple threads simultaneously. This restriction addressed the experiment
to be based on sequence calls only.

Acknowledgements. The authors would like to thank CAPES (www.capes.gov.br),

CNPq grant 475519/2012-4, CNPq Grant 445500/2014-0, FAPEMIG Grant APQ-
01589-11, and FAPEMIG grant APQ-2086/2011 for the financial support.

References
1. Kiczales, G., Lamping, J., Mendhekar, A., Maeda, C., Lopes, C., Loingtier, J.-M.,
Irwin, J.: Aspect-oriented programming. In: Akşit, M., Matsuoka, S. (eds.) ECOOP
1997. LNCS, vol. 1241, pp. 220–242. Springer, Heidelberg (1997)
202 M.S. Soares et al.

2. Ali, M.S., Ali, M., Babar, L., Chen, K.-J., Stol, A.: A systematic review of compar-
ative evidence of aspect-oriented programming. Inf. Softw. Technol. 52, 871–887
(2010)
3. Przybylek, A.: Impact of aspect-oriented programming on software modularity.
In: Proceedings of the 15th European Conference on Software Maintenance and
Reengineering, pp. 369–372 (2011)
4. Mortensen, M., Ghosh, S., Bieman, J.: Aspect-oriented refactoring of legacy appli-
cations: an evaluation. IEEE Trans. Softw. Eng. 38(1), 118–140 (2012)
5. França, J.M.S., Soares, M.S.: A systematic review on evaluation of aspect oriented
programming using software metrics. In: ICEIS 2012 - Proceedings of the 14th
International Conference on Enterprise Information Systems, vol. 2, pp. 77–83
(2012)
6. Liu, W.-L., Lung, C.-H., Ajila, S.: Impact of aspect-oriented programming on soft-
ware performance: a case study of leader/followers and half-sync/half-async archi-
tectures. In: Proceedings of the 2011 IEEE 35th Annual Computer Software and
Applications Conference, COMPSAC 2011, pp. 662–667 (2011)
7. Bijker, R.: Performance effects of Aspect Oriented Programming. Twente Univer-
sity, Technical report, Enchede, The Netherlands (2005)
8. Kirsten, M.: AOP@Work: AOP tools comparison, Part 1: Language mecha-
nisms, Technical report, IBM Developer Works (2005). http://www-128.ibm.com/
developerworks/java/library/j-aopwork1/
9. Kitchenham, B.: Procedures for Performing Systematic Reviews, Keele university.
Technical report tr/se-0401, Department of Computer Science, Keele University,
UK (2004)
10. Froihofer, L., Glos, G., Osrael, J., Goeschka, K.M.: Overview and evaluation of
constraint validation approaches in java. In: Proceedings of the 29th International
Conference on Software Engineering, ICSE 2007, pp. 313–322, Washington (2007)
11. Fabry, J., Tanter, É., D’Hondt, T.: KALA: kernel aspect language for advanced
transactions. Sci. Comput. Program. 71(3), 165–180 (2008)
12. Georg, G., Ray, I., Anastasakis, K., Bordbar, B., Toahchoodee, M., Houmb, S.H.:
An aspect-oriented methodology for designing secure applications. Inf. Softw. Tech-
nol. 51(5), 846–864 (2009)
13. Hundt, C., Glesner, S.: Optimizing aspectual execution mechanisms for embedded
applications. Electron. Notes Theor. Comput. Sci. 238(2), 35–45 (2009)
14. Ganesan, D., Keuler, T., Nishimura, Y.: Architecture compliance checking at run-
time. Inf. Softw. Technol. 51(11), 1586–1600 (2009)
15. Zhang, C.: FlexSync: an aspect-oriented approach to java synchronization. In: Pro-
ceedings of the 31st International Conference on Software Engineering, ICSE 2009,
pp. 375–385, Washington (2009)
16. Cannon, B., Wohlstadter, E.: Enforcing security for desktop clients using authority
aspects. In: Proceedings of the 8th ACM International Conference on Aspect-
Oriented Software Development, AOSD 2009, pp. 255–266. New York (2009)
17. Malek, S., Krishnan, H.R., Srinivasan, J.: Enhancing middleware support for
architecture-based development through compositional weaving of styles. J. Syst.
Softw. 83(12), 2513–2527 (2010)
18. Dyer, R., Rajan, H.: Supporting dynamic aspect-oriented features. ACM Trans.
Softw. Eng. Methodol. 20(2), 1–34 (2010)
19. Ortiz, G., Prado, A.G.D.: Improving device-aware web services and their mobile
clients through an aspect-oriented model-driven approach. Inf. Softw. Technol.
52(10), 1080–1093 (2010)
 Performance Evaluation of Aspect-Oriented Programming Weavers 203

20. Toledo, R., Leger, P., Tanter, E.: AspectScript: expressive aspects for the web.
In: Proceedings of the 9th International Conference on Aspect-Oriented Software
Development, AOSD 2010, pp. 13–24 (2010)
21. Janik, A., Zielinski, K.: AAOP-based dynamically reconfigurable monitoring sys-
tem. Inf. Softw. Technol. 52(4), 380–396 (2010)
22. Ansaloni, D., Binder, W., Villazón, A., Moret, P.: Parallel dynamic analysis on
multicores with aspect-oriented programming. In: Proceedings of the 9th Inter-
national Conference on Aspect-Oriented Software Development, AOSD 2010, pp.
1–12 (2010)
23. de Roo, A., Sozer, H., Aksit, M.: Verification and analysis of domain-specific mod-
els of physical characteristics in embedded control software. Inf. Softw. Technol.
54(12), 1432–1453 (2012)
 ABOR: An Automatic Framework for Buffer
Overflow Removal in C/C++Programs

Sun Ding1(&), Hee Beng Kuan Tan1, and Hongyu Zhang2

1
School of Electrical and Electronic Engineering,
Nanyang Technological University, Singapore city, Singapore
{ding0037,ibktan}@ntu.edu.sg
2
School of Software, Tsinghua University, Beijing, China
[email protected]

Abstract. Buffer overflow vulnerability is one of the commonly found sig-

nificant security vulnerabilities. This vulnerability may occur if a program does
not sufficiently prevent input from exceeding intended size and accessing
unintended memory locations. Researchers have put effort in different directions
to address this vulnerability. How, authorized reports and data showed that as
more sophisticated attack vectors are being discovered, efforts on a single
direction are not sufficient to resolve this critical issue well. In this paper, we
characterize buffer overflow vulnerability in four patterns and propose ABOR,
a framework to remove buffer overflow vulnerabilities from source code auto-
matically. It only patches identified code segments, which means it is an opti-
mized solution that eliminates buffer overflows at the maximum while adds
runtime overhead at the minimum. We have implemented the proposed
approach and evaluated ABOR over a set of real world C/C++ applications. The
results prove ABOR’s effectiveness in practice.

Keywords: Buffer overflow
Static analysis
Automatic bug fixing
Security

vulnerability

1 Introduction

Buffer overflow in C/C++ is still ranked as one of the major security vulnerabilities [1],
though it has been 20 years since this vulnerability was first exploited. This problem
has never been fully resolved and has caused enormous losses due to information
leakage or customer dissatisfaction [1].
To mitigate the threats of buffer overflow attacks, a number of approaches have
been proposed. Existing approaches and tools focus mainly on three directions [2]:
• Prevent buffer overflow attacks by creating a run-time environment, like a sandbox,
so that taint input could not directly affect certain key memory locations;
• Detect buffer overflows in programs by applying program analysis techniques to
analyze source code;
• Transform the original program by adding additional verification code or external
annotations.

© Springer International Publishing Switzerland 2015

J. Cordeiro et al. (Eds.): ICEIS 2014, LNBIP 227, pp. 204–221, 2015.
DOI: 10.1007/978-3-319-22348-3_12
 ABOR: An Automatic Framework for Buffer Overflow Removal 205

For approaches in the first direction, as modern programs are becoming more complex,
it is difficult to develop a universal run-time defense [2]. For approaches in the second
direction, even if buffer overflow vulnerabilities are detected, the vulnerable programs
are still being used until new patches are released. For the third direction, though it is
well-motivated to add extra validation to guard critical variables and operations, the
existing approaches will add considerable runtime overhead. For example, a recent
novel approach that adds extra bounds checking for every pointer may increase the
runtime overhead by 67 % on average [3].
We noticed though none of the existing methods can resolve the problem fully, they
share many commonalities and also have differences. In this paper, we first integrate
existing methods and characterize buffer overflow vulnerability in the form of four
patterns. We then propose a framework—ABOR that combines detection and removal
techniques together to improve the state-of-the-art. ABOR iteratively detects and
removes buffer overflows in a path-sensitive manner, until all the detected vulnera-
bilities are eliminated. Unlike the related methods [3–6], ABOR only patches identified
code segments; thus it can eliminate buffer overflows while keeping a minimum run-
time overhead.
We have evaluated the proposed approach on a set of benchmarks and three
industrial C/C++ applications. The results show that the proposed approach is effective.
First it can remove all the detected buffer overflow vulnerabilities in the studied sub-
jects. Second we also compare ABOR with methods that focus on buffer overflow
removal. On average, it removes 58.28 % more vulnerabilities than methods that apply
a straight-forward “search and replace” strategy; it inserted 72.06 % fewer predicates
than a customized bounds checker.
The contribution of the paper is as following:
• The proposed approach integrates and extends existing techniques to remove buffer
overflows automatically in a path-sensitive manner.
• The proposed approach guarantees a high removal accuracy while could keep a low
runtime overhead.
• The proposed approach contains an exhaustive lookup table that covers most of the
common buffer overflow vulnerabilities.
The paper is organized as follows. Section 2 provides background on buffer overflows
vulnerability. Section 3 covers the proposed approach that detects buffer overflows and
removes detected vulnerability automatically. Section 4 evaluates the proposed
approach and Sect. 5 reviews the related techniques that mitigate buffer overflow
attacks. Section 6 concludes the paper.

2 Background

In this section, we give background information about buffer overflow vulnerability

and attack. We also introduce a collection of rules for preventing such attacks.
Buffer overflow based attacks usually share a lot in common: they occur anytime
when a program fails to prevent input from exceeding intended buffer size(s) and
accessing critical memory locations. The attacker usually starts with the following
206 S. Ding et al.

attempts [1, 7]: they first exploit a memory location in the code segment that stores
operations accessing memory without proper boundary protection. For example,
a piece of code allows writing arbitrary length of user input to memory. Then they
Locate a desired memory location in data segment that stores (a) an important local
variable or (b) an address that is about to be loaded into the CPU’s Extended instruction
Counter (EIP register).
Attackers attempt to calculate the distance between the above two memory loca-
tions. Once such locations and distance are discovered, attackers construct a piece of
data of length (x1 + x2). The first x1 bytes of data can be any characters and is used to
fill in the gap between the exploited location and the desired location. The second x2
bytes of data is the attacking code which could be (a) an operation overwriting a local
variable, (b) a piece of shell script hijacking the system or (c) a handle redirecting to a
malicious procedure.
Therefore, to prevent buffer overflow attack, it is necessary to ensure buffer writing
operations are accessible only after proper validations.
Figure 1 shows an example of one buffer overflow vulnerability. For the sample
procedure _encode, node s1 uses the C string function strcpy to copy a bulk of data to a
destination buffer. Such a function call is vulnerable because strcpy does not have any
built-in mechanism that prevents over-sized data from being written to the destination
buffer. And before node s1, there is no explicit validation to constraint the relationship
between the pvpbuf’s size and req_bytes’s length. At node14, the variable req_bytes is
defined as an array length of 1024 and it receives a bulk of data size of 1024 from user
input via the command-line. Along the path (n9, n10, n11, n12, n13, n14, n15, n16, s1),
the variable pvpbuf is defined as an array size of c2. However there is no validation to

Fig. 1. A sample code with exploited buffer overflow vulnerability.

 ABOR: An Automatic Framework for Buffer Overflow Removal 207

ensure c2 is larger than 1024, in order to prevent pvpbuf being overwritten. Therefore
the system running this procedure _encode becomes vulnerable due to the existence of
(n9, n10, n11, n12, n13, n14, n15, n16, s1).

3 The Proposed Approach

In this paper, we propose Automatic Buffer Overflow Repairing (ABOR), a framework

that integrates and extends existing techniques to resolve the buffer overflow vulner-
abilities in a given program automatically. Figure 2 demonstrates the overall structure
of the framework. ABOR consists of two modules: vulnerability detection and vul-
nerability removal. ABOR works in an iterative way: once the vulnerability detection
module captures a vulnerable code segment, the segment is fed to the removal module;
the fixed segment will be patched back to the original program. ABOR repeats the
above procedure until the program is buffer-overflow-free. In this section, we introduce
the two modules of ABOR in detail.

3.1 Buffer Overflow Patterns

We first review some basic definitions of static analysis [8, 9]. A control flow graph
(CFG) for a procedure is a graph that visually presents the control flow among program
statements. A path is one single trace of executing a sequential of program statements.
A variable in a program is called an input variable if it is not defined in the program
solely from constants and variables. An abstract syntax tree (AST) is a tree structure
that represents the abstract syntax of source code written in a particular programming
language. We call a program operation that may cause potential buffer overflows as a
buffer overflow sensitive sink (bo-sensitive sink). In this paper, we shorten the name
“bo-sensitive sink” to sink. The node that contains a sink is called a sink node.
There are many methods to protect sinks from being exploited [10–14]. One
general way is to add extra protection constraints to protect sinks [2]. After a careful
review, we collect a list of common sinks and the corresponding protection constraints.
We characterize them in a form of four patterns:

Fig. 2. Overview of ABOR.

208 S. Ding et al.

(1) Pattern#1: A statement is a bo-sensitive sink when it defines or updates a des-

tination buffer with an input from either (1) a C stream input function which is
declared in < stdio.h > , or (2) a C ++ input function inherited from the base class
istream. The protection constraint shall ensure the length of the destination buffer
is not smaller than the input data.
(2) Pattern#2: A statement that copies/moves the content of a block of memory to
another block of memory is a bo-sensitive sink. The protection constraint shall
ensure that the copied/moved data is not larger than the length of the destination
memory block.
(3) Pattern#3: A statement is a bo-sensitive sink when it calls a C stream output
function when (1) this function is declared in < stdio.h > ; or (2) this function
contains a format string that mismatches its corresponding output data; or (3) the
function’s output is data dependent on its parameters [12]. The protection con-
straint shall ensure that:
• the output data should not contain any string derived from the prototype [15]:
%[flags][width][.precision][length]specifier;
• or all the character “%” in the output data has been encoded in a backslash
escape style, such as “\%”.
(4) Pattern#4: A statement other than the above cases but referencing a pointer or
array is a bo-sensitive sink. Before accessing this statement, there should be a
protection constraint to do boundary checking for this pointer or array.
We use the above four patterns as a guideline to construct the buffer overflow
detection and removal modules of ABOR. In order to specify these cases clearly, we
use metadata to describe them exhaustively at the AST level. The metadata is main-
tained in Table 1 (Here we only show a fragment of Table 1; the full table is presented
on our website). Each row in Table 1 stands for a concrete buffer overflow case. The
column Sink lists the AST structure of sinks. The column Protection Constraint
specifies the AST structure of the constraint which could prevent the sinks being
exploited. Additionally, in order to concrete the protection constraints ABOR needs to
substitute in constants, local variables, expressions, and also two more critical data
structures: the length of a buffer and the index of a buffer.

Table 1. Detail of ABOR pattern lookup table (a fragment).

Pattern Sink Protection Constraint Required
Intermediate Variable
1 gets(dst) dst_length ≥ SIZE_MAXa /*The destination buffer dst’s length
(bytes).*/ dst_length
2 memcpy dst_length ≥ t; //The destination buffer dst’s length,
(dst, src_length ≥ t; in terms of bytes. dst_length,
src, t) //The source buffer src’s length, in
terms of bytes. src_length
a
SIZE_MAX stands for the max value of unsigned long
 ABOR: An Automatic Framework for Buffer Overflow Removal 209

In C/C++, there is no universal way to retrieve a buffer’s length and index easily.
To resolve this, ABOR creates intermediate variables to represent them. In Table 1, the
last column Required Intermediate Variable records the required intermediate variables
for each constraint.

3.2 Buffer Overflow Detection

Among the existing detection methods, approaches working in a path sensitive manner
offer higher accuracy because they target at modeling the runtime behavior for each
execution path: for each path, path sensitive approaches eventually generate a path
constraint to reflect the relationship between the external input and target buffer, and
the path’s vulnerability is verified through validating the generated path constraint.
In the current implementation of ABOR, we modified a recent buffer overflow
detection method called Marple [16] and integrated it into ABOR. We chose Marple
mainly because it detects buffer overflows in a path sensitive manner, which offers high
precision.
Marple maintained a lookup table to store the common bo-sensitive sinks’ syntax
structure. For each recognized bo-sensitive sink node and the path passing through the
sink node, Marple generates an initial constraint, called query and backward propagates
the query along the path. The constraint will be updated through symbolic execution
when encountering nodes that could affect the data flow information related to the
constraint. Once the constraint is updated, Marple tries to validate it by invoking its
theorem prover. If it is proved the constraint is unsatisfiable, Marple concludes this path
buffer-overflow-vulnerable. Figure 3 demonstrates how ABOR integrates and extends
Marple:
(1) ABOR replaces Marple’s lookup table with Table 1. We enforce Marple to search
for the syntax structures listed in the column Sink of Table 1. Additionally,
Marple will raise a query based on the column Protection Constraint.
(2) ABOR uses a depth-first search to traverse a given procedure’s control flow
graph: each branch will be traversed once. If a bo-sensitive sink is found, the

A path with sink

Symbolic Execution
Constraint propagation
Next
path Constraint Solving: Z3
Constraint solving

Vulnerability Confirmation

Fig. 3. Vulnerability detection in ABOR.

210 S. Ding et al.

segment starting from the sink back to the procedure entry will be constructed to
be a set of paths and each path will be fed to Marple for processing.
(3) ABOR replaces Marple’s constraint solver with Z3 (Z3, 2013), a latest SMT
solver from Microsoft with strong solvability.
(4) If one path is identified vulnerable, ABOR records the sub-path that causes the
vulnerability or infeasibility [16]. Later, paths containing any of such sub-paths
will not be examined.
(5) We follow the way Marple handles loop structures: we treat a loop structure as a
unit and try to compute each loop’s impact on the propagating constraint, if and
only if such impact is linear.
We illustrate the above procedure with an example. In Table 4, s1 is a sink node.
Therefore, Marple raises a constraint as pvpbuf_length ≥ req_bytes_length. It propa-
gates backward along the paths that pass through s1 and tries to evaluate the constraint.
For example, along the path (n1, n2, n3, n4, n5, n6, n7, n9, n10, n11, n12, n9, n13,
n14, n15, n16, s1), the constraint is updated at node n13 and becomes c2 ≥ 1024. The
variable c2 is affected by the loop [n9, n10, n11, n12, n9]. The variable c2 is used in the
loop, and it is data dependent on the input–ADDRSIZE. There exists a counterexample
to violate the constraint c2 ≥ 1024. So this path is identified as being vulnerable. The
method introduced in Sect. 3.3 will be used to remove this vulnerability.

3.3 Buffer Overflow Removal

The removal module takes an identified vulnerable path, analyzes the sink’s AST and
picks the corresponding constraint to protect the sink. The main challenge is to concrete
the selected protection constraint into valid C ++ code.
ABOR propagates backward along the given path to enable the intermediate
variables simulating their semantics. It maintains Table 2 to assist this propagation.

Table 2. Update intermediate variable during propagation.

Syntax Structure Update Operation
Buffer definition
•buffer = new wildcard T [n]; T buffer buffer_length = n * sizeof(T);
[n];
Buffer referencing
•T * p = buffer; p_length = buffer_length; &&
p_index = buffer_index;
Array index subscription
•buffer[i] = wildcard buffer_index = i;
Pointer arithmetic
•p ++; p–; p_index ++; p_index–;
•p = p+n; p_index = p_index +n;
Free memory
•free (p); delete [] p; p_length = 0; p_index = NULL;
 ABOR: An Automatic Framework for Buffer Overflow Removal 211

In Table 2, the column Syntax Structure stands for the AST structure ABOR searches
for during the propagation. The column Update Operation stands for how ABOR
updates the corresponding intermediate variables.
The algorithm removeVul shown in Table 3 shows the workflow of ABOR’s
removal module. This algorithm takes an identified vulnerable path pth and its control
flow graph G as parameters. It inserts defensive code into G to remove the vulnera-
bility. The defensive code includes statements to update intermediate variables and a
predicate to protect the sink. After this, removeVul concretes the protection constraint
by substituting required local variables, expressions, constants and intermediate vari-
ables. The protection constraint is used to construct a predicate node. This predicate
node wraps the sink node to provide protections. At last, removeVul converts the
modified control flow graph back to source code. Therefore, the vulnerability caused by
the sink has been removed by ABOR.
Table 4 presents a full example of using ABOR. The vulnerable program is in the
left side column. The procedure of ABOR is as follows:

Table 3. Vulnerability removal in ABOR.

212 S. Ding et al.

Table 4. An example of applying ABOR.

(1) ABOR’s buffer overflow detection module finds that the path (n1, n2, n3, n4, n5,
n6, n7, n9, n10, n11, n12, n13, n14, n15, n16, s1) is vulnerable. The path will be
passed to removeVul.
(2) removeVul traverses the sink node s1’s AST and determines that the first pattern
shall be applied. The constraint is to validate that the length of pvpbuf is larger
than or equal to the length of req_bytes.
(3) ABOR inserts one node p1 into the CFG and creates two intermediate variables
with an initial value of 0: pvpbuf_temp0_size for the length of pvpbuf and
req_bytes_temp0_size for the length of req_bytes. ABOR inserts another three
 ABOR: An Automatic Framework for Buffer Overflow Removal 213

nodes p2, p3, and p4 into CFG, to manipulate the two intermediate variables to
track the lengths of pvpbuf and req_bytes.
(4) ABOR constructs the protection constraint as pvpbuf_temp0_size ≥ req_bytes_
temp0_size and transforms the original program by wrapping statement s1 with
statements p5, and p6.
(5) At last, ABOR converts the modified CFG back to source code, which is listed in
the right side column of Table 4. Therefore, the vulnerability has been removed.

4 Evaluation

4.1 Experiment Design

We have implemented the proposed approach as a prototype system based on the
CodeSurfer infrastructure [20]. The prototype has two parts: Program Analyzer and
ABOR. The Program Analyzer receives C/C++ programs as input and utilizes Code-
Surfer to build an inline inter-procedural CFG. This CFG is then sent to the ABOR for
the vulnerability detection and removal.
Nine systems are selected to evaluate ABOR’s performance: (a) six benchmark
programs from Buffer Overflow Benchmark [17] and BugBench [18] are selected for
the evaluation, namely Polymorph, Ncompress, Gzip, Bc, Wu-ftdp and Sendmail; and
(b) three industrial C/C++ applications, namely RouterCore, PathFinder, and RFID-
Scan, are selected. This is to further evaluate the effectiveness of the proposed approach
on real-world programs.
For each system, we first run ABOR and then manually validate the results. The
experiments are carried out on a desktop computer with Intel Duo E6750 2-core
processor, 2.66 GHz, 4 GB memory and Windows XP system.

4.2 Experimental Results

4.2.1 System Performance
We evaluate ABOR’s performance in terms of removal accuracy and time cost.

Removal Accuracy. For benchmark systems, the experimental results are shown in
Table 5(a). A false negative case occurs if we manually find that the proposed method

Table 5(a). Vulnerabilities removed by ABOR in benchmarks.

System KLOC #Reported #Repaired #FP #FN
Polymorph-0.4.0 1.7 15 15 0 0
Ncompress-4.2.4 2.0 38 38 0 0
Gzip-1.2.4 8.2 38 38 0 0
Bc-1.06 17.7 245 245 0 0
Wu- ftdp-2.6.2 0.4 13 13 0 0
Sendmail-8.7.5 0.7 21 21 0 0
Total 30.7 370 370 0 0
214 S. Ding et al.

Table 5(b). Vulnerabilities removed by ABOR in industrial programs.

System KLOC #Detected #Repaired #Manual #FP #FN ErrorRate(%)
RouterC* 137.15 217 217 309 3 41 14.23
Pathfinder 104.23 79 79 103 1 25 25.24
RFIDscan 219.36 312 312 406 2 96 24.13
Total 460.74 608 608 818 6 162 20.53

failed to remove one buffer overflow vulnerability. A false positive case occurs if we
manually find the proposed method patched a piece of code that is actually buffer
overflow free. We calculate the error rate of the proposed method by dividing the total
number of vulnerabilities by the sum of the number of false positives and false neg-
atives. The column KLOC stands for thousands of lines of code. The column #Reported
records the real reported number of vulnerabilities while column #Repaired records the
number of vulnerabilities removed by ABOR. The columns #FP and #FN stand for the
numbers of false positives and false negatives respectively. For all the systems, the total
number of reported buffer overflow vulnerabilities is 370. Therefore, ABOR can cor-
rectly remove all the vulnerabilities reported in previous work [17, 18].
For industrial programs, the results are shown in Table 5(b). The columns KLOC,
#Repaired, #FN and #FP have the same meaning with Table 5(a). Additionally, the
column #Detected stands for the number of detected vulnerabilities; and the column
#Manual stands for the number of vulnerabilities discovered from manual investiga-
tions. The manual investigation double checked the detection result and analyzed the
reason behind the cases that ABOR failed to proceed. As shown in Table 5(b), our
approach detects 608 buffer overflow vulnerabilities and can successfully remove all of
them. The results confirm the effectiveness of the proposed approach in removing
buffer overflow vulnerabilities. However, due to implementation limitations, ABOR’s
detection modules didn’t capture all the buffer overflow vulnerabilities in the industrial
programs. On average, the error rate of our proposed method is 20.53 %, which
consists of 19.80 % false negative cases and 0.73 % false positive cases. The details of
the error cases are discussed in Sect. 4.2.3.

Time Cost. We measured the time performance of the proposed approach on both
benchmarks and industrial programs. Table 6 records the time spent on processing each
program individually. ABOR is scalable to process large programs. The time cost over
the entire 9 systems is 4480 s, which is nearly 75 min. It is also discovered that a large
amount of time is spent on vulnerability detection, which is 77.77 % of the total time.
The vulnerability removal process is relatively lightweight, which costs only 22.23 %
of the total time.

4.2.2 Comparison
There are another two types of commonly used removal methods [2], which are “search
& replace” and “bounds checker”.

Search and Replace. The first category of methods replaces those common vulnerable
C string functions with safe versions. If a program contains a large number of C string
 ABOR: An Automatic Framework for Buffer Overflow Removal 215

Table 6. The time performance of ABOR.

System Total time (ms) Detection time Removal time
Time(ms) % Time (ms) %
Polymorph 95.25 81.91 85.99 13.34 14.01
Ncompress 214.32 160.81 75.03 53.51 24.98
Gzip 3698.16 2388.71 64.59 1309.45 35.41
Bc 149469.60 132026.90 88.33 17442.66 11.67
Wu- ftdp 221.13 185.33 83.81 35.80 16.19
Sendmail 134.82 103.76 76.96 31.06 23.04
Routercore 2446656.17 1781649.13 72.82 665007.07 27.18
Pathfinder 654987.43 564359.17 86.16 90628.22 13.84
RFIDscan 1224366.67 1003880.72 81.99 220485.98 18.01
Total 4479843.55 3484836.44 77.77 995007.09 22.23

functions, this category of methods can achieve a good effect. Additionally, they are
straight-forward for implementation [19]. But as the fast development of attacking
techniques based on buffer overflows [2], the “search and replace”, methods will miss
many buffer overflows in real code.
Bounds checker: The second category of methods chases high precision by
inserting effective validation before every memory access. In practice, they are usually
used by mission-critical systems [2, 3]. However, they normally bring in high runtime
overhead as a number of inserted validations are redundant.
ABOR is the method that only patches identified detected vulnerable code segment
in a path-sensitive way. So it guarantees the removal precision while it can keep a low
runtime overhead. Using the same benchmarks and industrial programs, we compare
ABOR with the other two main categories of removal methods.
First, we compare ABOR with the “search and replace” category. In Fig. 4(a), we
compare ABOR with a recent “search and replace” method from Hafiz and Johnson [6].
It maintains a static database that stores common C/C++ vulnerable functions with their
safe versions.

Fig. 4. Compare ABOR with existing methods.

216 S. Ding et al.

Figure 4(a) represents the histograms on the number of removed buffer overflows.
It contains a pair of bars for each test case. The shadowed ones are for ABOR while the
while bars are for the applied “search and replace” method. The Longer bars are better
as they represent the higher removal precision. For the entire 9 systems, in total, the
“search and replace” method removes 570 vulnerabilities while ABOR removes 978
vulnerabilities (41.72 % more). This is mainly because (1) many vulnerable codes are
not covered by the static database of the “search and replace” method; (2) even after
using a safe version of certain C/C++ functions, the vulnerabilities are not removed due
to improper function parameters.
Second, we compare ABOR with the bounds checker” category. At the current
stage, we implemented a customized bounds checker following a novel approach from
Nagarakatte et al. [3]. It will insert predicates to protect every suspicious sink.
Figure 4(b) represents the histograms on the number of inserted predicates. As with
more inserted predicates, runtime overhead will increase. Figure 4(b) contains a pair of
bars for each test case. The shadowed ones are for ABOR while the white ones are for
the applied bounds checker. Shorter bars are better as they represent the fewer number
of inserted predicates. For the entire 9 systems, in total, the applied bounds checker
inserts 3500 predicates. ABOR only patches confirmed sinks so it only inserts 978
predicates, which are 72.05 % less than the applied bounds checker.
Last but not least, though detection of buffer overflow is not a new research, till
now, no approaches can detect buffer overflow with full coverage and precision. As the
proposed approach uses these approaches, it is also limited by the accuracy of these
approaches.

4.2.3 Discussion
It is also found that ABOR caused some false positive and false negative cases when
processing the industrial programs. We further investigate these error cases and found
the errors are mainly caused by implementation limitations. We categorize them into
three types:
Error 1- inaccuracy from alias analysis: it is difficult to implement a comprehensive
alias analysis (CodeSurfer, [20]). Table 7(a) shows a false-negative example from

Table 7. Examples of errors from processing the industrial programs.

 ABOR: An Automatic Framework for Buffer Overflow Removal 217

RouterCore that is caused by inaccurate pointer analysis. In the loop, pointer ptr is
incremented by one in each iteration until it equals to the address of pointer slt. So after
the loop, ptr and slt are actually aliased. However, currently we cannot detect such alias
relationship. The node n2 could overwrite the value of *ptr. Though node n1 does the
boundary checking, it no longer protects node n3.
Error 2- inaccuracy from loop structure. So far only variables that are linearly
updated within an iteration are handled by ABOR. Table 7(b) shows an example found
in RouterCore. The variable x is non-linearly updated by using a bitwise operation.
Therefore the corresponding constraint, which compares the size between buff and
input, is beyond the solvability of our current implementation.
Error 3- platform-based data types: the industrial programs PathFinder and
RFIDScan both involve external data types of Microsoft Windows SDK (e.g., WORD,
DWORD, DWORD_PTR, etc.). This requires extra implementations to interpret them.
Additionally the value ranges of these data types specify implicit constraints. Table 7(c)
shows a false-positive case from PathFinder. In this example, although accessing the
buffer element via index at node n2 has no protection, there will be no buffer overflow
vulnerability because index ranges only from 0 to 65535.
We summarize the error analysis in Table 8. In the future, further engineering effort
is required to address these implementation difficulties.

5 Related Work

We reviewed the recent techniques in addressing the buffer overflow vulnerability. We

categorize them into three types: buffer overflow detection, runtime defense and vul-
nerability removal.

5.1 Buffer Overflow Detection

Buffer overflow vulnerability could be effectively detected by well-organized program
analysis. The analysis could be performed either on source code or binary code. The
current detection methods can be classified into path-sensitive approaches and
non-path-sensitive approaches.
Path-sensitive approaches analyze given paths and generate path constraints
according to the properties that ensure the paths are not exploitable for any buffer
overflow attack. The path constraints are extracted using symbolic evaluation through

Table 8. Accuracy and error analysis.

System Total error E1 E2 E3
# % # % # %
Routercore 44 23 52.27 10 22.73 11 25.00
Pathfinder 26 16 61.54 0 0 10 38.47
RFIDscan 98 31 31.63 22 22.45 45 45.92
Total 168 70 41.67 32 19.05 66 39.29
218 S. Ding et al.

either forward or backward propagation. A theorem prover or customized constraint

solver is instrumented to evaluate the constraints. If a constraint is determined as
unsolvable, the path is concluded as vulnerable. These methods pursue soundness and
precision but usually include heavy overhead due to the use of symbolic evaluation.
Typical examples include ARCHER [21], Marple [16], etc.
Non-path-sensitive approaches try to avoid complex analysis. They usually rely on
general data flow analysis. These methods compute the environment, which is a
mapping of program variables to values at those suspected locations. The environment
captures all the possible values at a location L. Therefore, if any values are found to
violate the buffer boundary at L, a buffer overflow vulnerability is detected. Choosing
proper locations and making safe approximation requires heuristics and sacrifices
precision. However, such a design gains more scalable performance, which is highly
essential when dealing with large-scale applications. Typical examples include the
from Larochelle and Evans [22], CSSV [23], etc.

5.2 Run-Time Defense

This type of approaches adopt run-time defense to prevent exploiting potential vul-
nerabilities of any installed programs. These approaches aim to provide an extra pro-
tection regardless of what the source program is. Normally such protections are
implemented using three different kinds of techniques: code instrumentation, infra-
structure modification, and network assistance.
The first aspect is to simulate a run-time environment partially, execute the sus-
picious program in the virtual environment and check whether malicious actions have
taken place. Examples include StakeGuard and RAD [24], which are tools that create
virtual variables to simulate function’s return address. Suspicious code will be redi-
rected to act over the virtual variables first.
The second aspect is to modify the underlying mechanism to eliminate the root
of attacks: Xu et al. [25] proposed a method to split stack and store data and control
information separately; Ozdoganoglu et al. [26] proposed to implement a non-executable
stack. More details and corresponding tools of run-time defense could be found in recent
surveys [2, 27].
The third aspect is to do a taint analysis for the input data from any the untrusted
network source. This analysis compares network data with vulnerability signatures of
the recorded attacks, or inspects payload for shell code for detecting and preventing
exploitation. TaintCheck [28], proposed by Newsome and Song, tracks the propagation
of tainted data that comes from un-trusted network sources. If a vulnerability signature
is found, the attack is detected. Similar approaches include PASAN from Smirnov and
Chiueh [29], and Vigilante from Costa et al. [30].

5.3 Attack Prevention Through Auto Patching

The last category aims to removal the vulnerability from the source code. The objective
is to add defensive code to wrap sink nodes so that no taint data could access the sinks
 ABOR: An Automatic Framework for Buffer Overflow Removal 219

directly. There are three strategies to design defensive code and insert them: search and
replace, bounds checker, and detect and transform [2].
The first strategy is to search common vulnerable C string functions and I/O
operations and replaces the found operations with a safe version. For example, vul-
nerable functions strcpy and strcat could be replaced as strncpy and strncat or strlcpy
and strlcat [19], or even with a customized version. Munawar and Ralph [6] proposed a
reliable approach to replace strcpy and strcat with a customized version based on
heuristics. This strategy is straight-forward and easy to implement. However, it misses
many complex situations.
The second strategy is aiming to insert effective validation before each memory
access to perform an extra bounds checking. A typical design is to add validation
before each pointer operation, named “pointer-based approach”. These approaches
track the base and bound information for every pointer and validate each pointer
manipulation operation against the tracked information. Examples include CCured,
MSCC, SafeC, and Softbound [3]. These approaches are designed to provide high
precision. However, as nearly every pointer operation will be wrapped with additional
checking, the code may grow large and the runtime performance could be downgraded.
The third strategy is to locate the vulnerability first and then transform the vul-
nerable code segment into a safe version. Comparing with the second strategy, this
helps reduce the size of added code without compromising the removal precision.
Relatively few efforts have been put in this direction. Lin et al. [11] proposed to slice
the code based on a taint analysis, and then detect the buffer overflow vulnerability over
each slice. The detected vulnerable slice will be inserted in additional code for vul-
nerability removal. However, their approach is based on path-insensitive information
only. Wang et al. [10] proposed a method to add extra protection constraints to protect
sink nodes. The method called model checker to verify the satisfiability of the inserted
protection constraints. If the constraint fails to hold, that sink node will be recorded
vulnerable, and the corresponding constraints will be left to protect the sink node.
However, this method is path-insensitive.
ABOR follows the third strategy and pushes the state-of-the-art one step ahead. It
first detects the buffer overflows based on path sensitive information and then only add
defensive code to repair the vulnerable code segment. Therefore, it adds limited code
and has a low runtime overhead.

6 Conclusions

In this paper, we have presented an approach to remove buffer overflow vulnerabilities

in C/C++ programs automatically. We first characterize buffer overflow vulnerability in
the form of four patterns. We then integrate and extend existing techniques to propose a
framework—ABOR that removes buffer overflow vulnerability automatically: ABOR
iteratively detects and removes buffer overflows in a path-sensitive manner, until all the
detected vulnerabilities are eliminated. Additionally, ABOR only patches vulnerable
code segment. Therefore, it keeps a lightweight runtime overhead. Using a set of
benchmark programs and three industrial programs written in C/C++, we experimentally
220 S. Ding et al.

show that the proposed approach is effective and scalable for removing all the detected
buffer overflow vulnerabilities.
In the future, we will improve our approach and tool to minimize the number of
wrongly detected cases. We will also integrate ABOR into existing testing frameworks,
such as CUnit, GoogleTest, to further demonstrate its practicality.

Acknowledgements. The authors thank the Jiangsu Celestvision from China for assisting this
study and providing their internal programs for our experiment.

References
1. US-CERT (2014). http://www.us-cert.gov/
2. Younan, Y., Joosen, W., Piessens, F.: Runtime countermeasures for code injection attacks
against C and C ++ programs. ACM Comput. Surv. 44, 1–28 (2012)
3. Nagarakatte, S., Zhao, J., Martin, M.M.K., Zdancewic, S.: SoftBound: highly compatible
and complete spatial memory safety for C. In: Proceedings of the 2009 ACM SIGPLAN
Conference on Programming Language Design and Implementation, pp. 245–258. ACM,
Dublin, Ireland (2009)
4. Criswell, J., Lenharth, A., Dhurjati, D., Adve, V.: Secure virtual architecture: a safe
execution environment for commodity operating systems. SIGOPS Oper. Syst. Rev. 41,
351–366 (2007)
5. Dhurjati, D., Adve, V.: Backwards-compatible array bounds checking for C with very low
overhead. In: Proceedings of the 28th international conference on Software engineering,
pp. 162–171. ACM, Shanghai, China (2006)
6. Hafiz, M., Johnson, R.E.: Security-oriented program transformations. In: Proceedings of the
5th Annual Workshop on Cyber Security and Information Intelligence Research: Cyber
Security and Information Intelligence Challenges and Strategies, pp. 1–4. ACM, Oak Ridge,
Tennessee (2009)
7. Vallentin, M.: On the Evolution of Buffer Overflows. Addison-Wesley Longman Publishing
Co., Boston (2007)
8. Sinha, S., Harrold, M.J., Rothermel, G.: Interprocedural control dependence. ACM Trans.
Softw. Eng. Methodol. 10, 209–254 (2001)
9. en.wikipedia.org/wiki/Abstract_syntax_tree
10. Lei, W., Qiang, Z., Pengchao, Z.: Automated detection of code vulnerabilities based on
program analysis and model checking. In: Eighth IEEE International Working Conference
on Source Code Analysis and Manipulation 2008, pp. 165–173 (2008)
11. Lin, Z., Jiang, X., Xu, D., Mao, B., Xie, L.: AutoPaG: towards automated software patch
generation with source code root cause identification and repair. In: Proceedings of the 2nd
ACM symposium on Information, Computer and Communications Security, pp. 329–340.
ACM, Singapore (2007)
12. Lhee, K.-S., Chapin, S.J.: Buffer overflow and format string overflow vulnerabilities. Softw.
Pract. Exper. 33, 423–460 (2003)
13. Necula, G.C., Condit, J., Harren, M., McPeak, S., Weimer, W.: CCured: type-safe
retrofitting of legacy software. ACM Trans. Program. Lang. Syst. 27, 477–526 (2005)
14. Kundu, A., Bertino, E.: A new class of buffer overflow attacks. In: Proceedings of the 2011
31st International Conference on Distributed Computing Systems, pp. 730–739. IEEE
Computer Society (2011)
 ABOR: An Automatic Framework for Buffer Overflow Removal 221

15. C ++ Ref (2014). http://www.cplusplus.com/reference/

16. Le, W., Soffa, M.L.: Marple: a demand-driven path-sensitive buffer overflow detector. In:
Proceedings of the 16th ACM SIGSOFT International Symposium on Foundations of
Software Engineering, pp. 272–282. ACM, Atlanta, Georgia (2008)
17. Zitser, M., Lippmann, R., Leek, T.: Testing static analysis tools using exploitable buffer
overflows from open source code. SIGSOFT Softw. Eng. Notes 29, 97–106 (2004)
18. Lu, S., Li, Z., Qin, F., Tan, L., Zhou, P., Zhou, Y.: Bugbench: benchmarks for evaluating
bug detection tools. In: Workshop on the Evaluation of Software Defect Detection Tools.
(2005)
19. Miller, T.C., Raadt, T.D.: Strlcpy and strlcat: consistent, safe, string copy and concatenation.
In: Proceedings of the Annual Conference on USENIX Annual Technical Conference,
pp. 41–41. USENIX Association, Monterey, California (1999)
20. GrammaTech (2014). http://www.grammatech.com/products/codesurfer
21. Xie, Y., Chou, A., Engler, D.: ARCHER: using symbolic, path-sensitive analysis to detect
memory access errors. In: ESEC/FSE-11: Proceedings of the 9th European Software
Engineering Conference Held Jointly with 11th ACM SIGSOFT International Symposium
On Foundations Of Software Engineering, pp. 327–336. ACM, (2004)
22. Larochelle, D., Evans, D.: Statically detecting likely buffer overflow vulnerabilities. In:
Proceedings of the 10th Conference on USENIX Security Symposium, vol. 10, pp. 14–14.
USENIX Association, Washington, D.C. (2001)
23. Dor, N., Rodeh, M., Sagiv, M.: CSSV: towards a realistic tool for statically detecting all
buffer overflows in C. In: PLDI 2003: Proceedings of the ACM SIGPLAN 2003 Conference
on Programming Language Design and Implementation, pp. 155–167. ACM, (2003)
24. Wilander, J., Kamkar, M.: A comparison of publicly available tools for dynamic buffer
overflow prevention. In: Network and Distributed System Security Symposium (NDSS),
pp. 149–162 (2003)
25. Xu, J., Kalbarczyk, Z., Patel, S., Ravishankar, I.: Architecture support for defending against
buffer overflow attacks. In: Second Workshop on Evaluating and Architecting System
Dependability, pp. 55–62 (2002)
26. Ozdoganoglu, H., Vijaykumar, T.N., Brodley, C.E., Kuperman, B.A., Jalote, A.:
SmashGuard: a hardware solution to prevent security attacks on the function return
address. IEEE Trans. Comput. 55, 1271–1285 (2006)
27. Padmanabhuni, B., Tan, H.: Techniques for Defending from Buffer Overflow Vulnerability
Security Exploits. Internet Computing, IEEE PP, 1–1 (2011)
28. Newsome, J., Song, D.: Dynamic taint analysis for automatic detection, analysis, and
signature generation of exploits on commodity software. In: Proceedings of the Network and
Distributed System Security Symposium (2005)
29. Smirnov, A., Tzi-cker, C.: Automatic patch generation for buffer overflow attacks. In: Third
International Symposium on Information Assurance and Security, IAS 2007, pp. 165–170
(2007)
30. Costa, M., Crowcroft, J., Castro, M., Rowstron, A., Zhou, L., Zhang, L., Barham, P.:
Vigilante: end-to-end containment of internet worm epidemics. ACM Trans. Comput. Syst.
26, 1–68 (2008)
31. Automatic Buffer Overflow Repairing (2014).http://sunshine-nanyang.com/index.html
 ValueApping: An Analysis Method to Identify
Value-Adding Mobile Enterprise Apps in Business
Processes

Eva Hoos1 ✉ , Christoph Gröger1, Stefan Kramer2, and Bernhard Mitschang1

( )

1
Graduate School of Excellence Advanced Manufacturing Engineering (GSaME),
University of Stuttgart, Universitätsstraße 38, 70569 Stuttgart, Germany
[email protected]
2
Daimler AG, 71034 Böblingen, Germany

Abstract. Mobile enterprise apps provide novel possibilities for the optimization
and redesign of business processes, e.g., by the elimination of paper-based data
acquisitioning or ubiquitous access to up-to-date information. To leverage these
business potentials, a critical success factor is the identification and evaluation of
value-adding MEAs based on an analysis of the business process. For this
purpose, we present ValueApping, a systematic analysis method to identify usage
scenarios for value-adding mobile enterprise apps in business processes and to
analyze their business benefits. We describe the different analysis steps and corre‐
sponding analysis artifacts of ValueApping and discuss the results of a case-
oriented evaluation in the automotive industry.

Keywords: Business processes · Analysis framework · Mobile application

1 Introduction

As Apple released the first iPhone in 2007, the success story of mobile applications
running on smartphones and tablets, so called mobile apps, began. In 2014, there will
be over 138 Billion app downloads [1]. Nowadays, the majority of target users are
consumers, but there is an increasing focus on mobile technology in enterprises. Mobile
apps used in business are called mobile enterprise apps (MEA) and they are one of the
top ten strategic technology trends for enterprises in 2014 [2]. The unique features of
MEAs like intuitive touchscreen-based handling, anywhere and anytime usage as well
as sensor capabilities enable a large spectrum of opportunities for new business models
and novel business processes. The employment of MEAs can lead to higher productivity,
higher employee satisfaction, integration of mobile process activities, elimination of
paper-based data acquisitioning and ubiquitous information access [3].
Typically, the introduction of MEAs is realized in a technology-driven manner, which
means MEAs are developed because it is technically feasible and fancy. However, a
benefit from a business point of view cannot be guaranteed. In order to realize the busi‐
ness value of MEAs, the enterprise has to establish a mobile strategy. Thereby, a critical
success factor is the business-driven identification and evaluation of value-adding MEAs

© Springer International Publishing Switzerland 2015

J. Cordeiro et al. (Eds.): ICEIS 2014, LNBIP 227, pp. 222–243, 2015.
DOI: 10.1007/978-3-319-22348-3_13
ValueApping: An Analysis Method to Identify Value-Adding Mobile Enterprise Apps 223

based on a comprehensive analysis of the business process to be optimized. The goal is to

identify process activities suited for MEAs and the derivation of corresponding usage
scenarios. This includes the analysis of business benefits of desired MEAs and their
communication with the corporate management and the employees [4].
A key problem in the identification of value-adding MEAs is that the spectrum of
new fields of application enabled by mobile apps in business is not yet well-understood,
especially the conceptual differences compared to existing mobile IT systems like
laptops have not been analyzed sufficiently. Furthermore, there is a methodological lack
to systematically identify value-adding MEA usage scenarios. Previous methodologies
consider mobile applications in general but do not focus on mobile apps and their unique
features specifically [5–7]. Moreover, there is no clear picture of the business benefits
of mobile app usage as these benefits comprise monetary and non-monetary factors [8].
To address these problems, we present ValueApping, an analysis method to identify
value-adding MEA usage scenarios in business processes. It contains a criteria catalogue
which combines technological and business-oriented aspects of mobile app usage and
comprises several systematic analysis steps. By applying this method, it can be decided,
which type of IT technology fits best to support a particular process activity. Moreover,
ValueApping enables the analysis of the business benefits of a MEA usage scenario for
a certain process activity.
The remainder of this paper is structured as follows: Sect. 2 characterizes mobile
enterprise apps in general. Section 3 gives an overview of ValueApping including
addressed requirements, analysis steps and the underlying analysis artifacts. Section 3
details on the analysis artifacts whereas the analysis steps are described in Sect. 4. An
evaluation of ValueApping is presented in Sect. 5 based on a case-oriented application
in the automotive industry. Section 6 reviews related work. Finally, Sect. 7 concludes
the paper and highlights future work.
This paper represents a significantly revised and extended version of our work
presented in [9]. In particular, we extend our method by several components for a busi‐
ness benefit analysis. Furthermore, we investigate the special characteristics of mobile
enterprise apps compared to traditional mobile enterprise applications.

2 Mobile Enterprise Apps

In this section, MEAs are defined and characterized based on the term “mobile apps”.
According to [10], mobile apps are applications running on smart mobile touch-based
devices (SMTD) such as smartphones and tablets. Thereby, applications comprise all
types of executable programs as well as browser-based web applications. MEAs are
mobile apps used in business, whereas we specifically focus on mobile apps for
employees. MEAs differ from traditional mobile applications running on mobile devices
such as laptops, PDAs and mobile phones by the technical capabilities of SMTDs as
well as the way they are developed, distributed and consumed which is known as the
mobile ecosystem. Thereby, several technical and organizational challenges have to be
addressed to leverage MEAs. These aspects are explained in the following.
224 E. Hoos et al.

2.1 Mobile Devices and Technical Capabilities

Originally, there were two different types of mobile devices [11]: Mobile PCs such as
laptops and mobile handhelds such as mobile phones and PDAs. Mobile PCs have
similar capabilities as desktop PCs. In contrast, mobile handhelds are highly portable,
always-on and they are equipped with various sensors, but they have limited resources
such as computing power. The evolution of mobile technology lead to the combination
of these two device types into smart mobile touch-based devices (SMTD) such as smart‐
phones and tablets [12–15] as shown in Fig. 1. SMTDs are more and more replacing
mobile handhelds and due to their increasing technical capabilities they are becoming
an alternative to mobile PCs [14], as well. In contrast to mobile PCs, SMTDs are char‐
acterized by a unique feature set [10, 13–17] comprising anywhere and anytime usage,
intuitive touch-based interaction as well as various device sensors like a GPS receiver.
In the following, these features are discussed in detail.

Fig. 1. Evolution of mobile devices.

Due to mobile networks, SMTDs allow for anywhere and anytime usage. This
enables employees to access the enterprise back-end whenever it is required. Besides,
the small and handy form factor of SMTDs leads to a higher portability. Moreover,
SMTDs have no long boot process, because they are designed to be always on. This
increases the reachability and working ability of the user.
Due to their touchscreen handling and multi-modal input capabilities, SMTDs enable
an intuitive touch-based interaction using touch events and gestures [18]. This is more
appropriate in mobile environments than the usage of mouse and keyboard. Due to the
small screen size, the interaction design has to be tailored towards the functionality
relevant for the task. We call this a task-oriented design. Moreover, the new interaction
paradigm enables people who are generally uncomfortable with computers to interact
with SMTDs without prior training [19]. In addition, sensors enable further input capa‐
bilities as described in the following.
SMTD are equipped with various device sensors including camera, GPS receiver,
and accelerometer. They can be used for interaction as well as context sensing and
enrichment. Regarding context sensing, sensors can be used for taking a photo, for voice
recording, and for positioning. Furthermore, sensors can be used for interaction using
voice commands or motion gestures. The camera can also be used for sensing the envi‐
ronment and in augmented reality application.
ValueApping: An Analysis Method to Identify Value-Adding Mobile Enterprise Apps 225

2.2 Mobile Ecosystems and Challenges to Leverage Mobile Enterprise Apps

Generally, mobile apps are part of complex mobile ecosystems which define the way
apps are developed and distributed. A mobile ecosystem consists of the SMTD as a
hardware component, the mobile operating system, native apps running on the operating
system as well as a storefront and supplementary online services [20]. Google’s Android,
Apple’s iOS, and Microsoft’s Windows 8 are the three major mobile ecosystems
competing nowadays. Each defines software development kits and user interaction
guidelines as well as app store concepts to distributed mobile apps.
The unique features of SMTDs as well as the preexistence of mobile ecosystems
lead to five core challenges to leverage MEAs according to [10]:
• MEA Portfolio: Systematic identification of usage scenarios for MEAs in business
processes to derive value-adding MEAs and define a corresponding MEA portfolio.
• MEA Development: The development of MEAs is challenged by the requirements of
high usability and restricted computing resources as well as limited network capaci‐
ties [4]. In particular, a task-oriented interaction design is a critical success factor for
usability. Moreover, the coexistence of different mobile ecosystems requires cross-
platform development approaches.
• MEA Infrastructure: In order to distribute and manage MEAs, enterprises have to
design company-internal app stores oriented towards apps store for consumers.
Moreover, a unified device management is necessary, e.g., to define policies and
control operating system updates across various types of SMTDs and mobile oper‐
ating systems.
• MEA Security and Privacy: The use of MEAs poses new risks on privacy and security.
Business-critical information has to be secured even in case of loss of SMTDs [21].
Moreover, personal data and context data on SMTDs have to be protected to avoid
employee tracking.
• MEA IT Architecture: The use of MEAs on top of existing back-end IT systems
requires an adaption and extension of the IT architecture especially with respect to
the design of lightweight and context-aware back-end services.
With ValueApping, we address the MEA portfolio challenge by providing a system‐
atic analysis method to derive value-adding MEA usage scenarios in business processes.

3 Requirements and Overview of ValueApping

In this section, we first define the requirements of an analysis method to identify value-
adding MEA usage scenarios. Then, we give an overview of ValueApping and its parts.

3.1 Requirements

In order to identify value-adding MEA usage scenarios in business processes, for each
process activity, the type of IT technology which fits best has to be selected. The tech‐
nologies range from PCs as stationary IT technology and laptops as mobile PCs to
SMTDs such as smartphones and tablets as a basis for MEAs.
226 E. Hoos et al.

The corresponding decision making process is complex, because there are several
issues and requirements to consider. On the basis of shortcomings of existing analysis
approaches (see Sect. 7), we identified the following three major requirements Ri our
method has to address:
Potential of mobile technology (R1): A central question is whether there is a business
benefit of using mobile technology. Generally, mobile technology can have two different
effects on business processes [7]:
• Supporting existing mobility given by the process
• Enabling novel mobility in processes where none existed before
However, not every employment of mobile technology leads to an improvement of
the business processes in terms of efficiency and effectiveness. Hence, activities that
profit from one of the two effects have to be identified systematically.
Types of mobile devices (R2): There are a lot of different devices for mobile technology
such as laptops, smartphones, tablets, PDAs, and mobile phones differing in hardware
and software characteristics. In this work, we are considering the following types of
mobile devices as described in Sect. 2.1:
• Mobile PCs like laptops
• Smart mobile touch-based devices like smartphones and tablets
Holistic Point of View (R3). The combination of business-oriented and technology-
oriented aspects avoids a purely technology-driven introduction of mobile technology.
The latter typically focuses on porting existing back-end applications on mobile apps
without a detailed business analysis. Besides, business aspects do not only refer to the
mobility of process activities but further contextual factors like the elimination of manual
data acquisition. In addition, not only aspects of the process activity but also infrastruc‐
tural and organizational issues of the enterprise, e.g., the availability of a mobile
network, have to be considered.

3.2 Overview of ValueApping

The purpose of ValueApping is to systematically analyze process activities with respect

to their improvement potential using mobile technology in order to support enterprises
in the decision which IT technology fits best. At this, ValueApping incorporates the
above requirements and takes a holistic view on both technological and business-related
aspects (R3), differentiates between SMTDs and mobile PCs (R1) and investigates both
the support of existing mobility and the enablement of novel mobility in the process
(R2). Process improvements can be determined according to the goal dimensions time,
quality, and flexibility. The major result is a portfolio of analyzed process activities
which are categorized according to the IT technology which fits best. On this basis, a
business benefit radar chart for each process activity suited for SMTD support is derived.
The radar chart represents the business benefits of a corresponding MEA usage scenario
for the activity according to the goal dimen-sions. This provides a starting point to define
corresponding MEA development projects and IT investments.
ValueApping: An Analysis Method to Identify Value-Adding Mobile Enterprise Apps 227

ValueApping is made up of two major parts, namely analysis steps and analysis
artifacts (see Fig. 2). The analysis steps are executed in a sequence and create different
analysis artifacts as input and output. Thereby, we distinguish between four groups of
analysis artifacts, namely
• the criteria catalogue and criteria values,
• the app potential as a metric,
• the app management portfolio, and
• the business benefit radar chart.

Fig. 2. Overview of ValueApping.

The criteria catalogue reflects the different aspects for the usage of mobile technology
in enterprises. The app potential is a metric to operationalize the improvement potential
of each activity with respect to MEA. This means, the higher the app potential the more
the activity can be improved using a MEA. The app management portfolio enables the
classification and ranking of the activities according to the IT technology which fits best.
The business benefit radar chart is based on a business benefit breakdown structure. It
indicates on which goal dimensions the usage of a MEA has a positive impact for a
particular process activity. The chart enables a comparison with other activities
according to the goal dimensions and constitutes a graphical representation to easily
communicate potential benefits.
ValueApping comprises four analysis steps, namely
• process analysis,
• app potential evaluation,
• recommendation generation, and
• business benefit analysis.
The two starting points represent different application variants of the method. The
user point of view enables employees to validate improvement suggestions for selected
activities across different processes. The process point of view considers improvements
of an entire process including all activities.
Process analysis refers to a procedure to determine the value of each criterion in the
criteria catalog. The input is the criteria catalogue and the output comprises a criteria
value for each analyzed criterion. These values represent in turn the input for the eval‐
uation of the app potential. The latter defines a procedure to calculate the app potential
for each activity as a metric. Then, recommendation generation reveals the app manage‐
ment portfolio according to the app potential of each activity. On this basis, recommen‐
dations are deduced according to the IT technology which fits best for each activity in
228 E. Hoos et al.

the portfolio. At last, the business benefit analysis step reveals the business benefits of
corresponding MEAs for all process activities which are suited for SMTD support
according to the app management portfolio.

4 Analysis Artifacts

This section describes the analysis artifacts, namely the criteria catalogue, the app
potential, the app management portfolio, as well as the business benefit breakdown
structure and radar chart.

4.1 Criteria Catalogue

The criteria catalogue is based on multi-criteria analysis techniques. With these techni‐
ques, complex decision problems with multiple options and restrictions can be structured
[22]. As a basis for the criteria definition, we conducted literature analyses [5–7, 23–28]
Moreover, we carried out expert interviews with employees of a German car manufac‐
turer to refine the identified criteria.
The criteria catalogue reflects the different aspects of mobile app usage in enterprises
including the requirements R1, R2, and R3. The criteria are grouped into four categories:
mobility, process, technology requirement, and corporate conditions. Each criterion has
predefined ordinal values following a qualitative approach. In addition, some criteria
are complemented by indicators to ease the determination of their value. Table 1 shows
the structure of the criteria catalogue. In the following, an overview of the different
categories and the corresponding criteria is given.

Mobility of the Activity. This category includes two criteria: task and actor. These
criteria consider the aspects given in R1. The criterion task is based on the definition of
mobile processes given in [5] and has the predefined values of high, medium and low.
The indicators are a stationary workplace, the uncertainty of the execution space, moving
actor or multiple execution places. The uncertainty of the execution space emerges if
the execution space is unknown at the start of the process or it differs in multiple instances
of the process. For example, the value of the criterion task is high, if there is a high
uncertainty of the execution space, a moving actor or multiple execution spaces. The
value is low if the task is executed on a stationary workspace. This criterion investigates
whether mobile technology can be employed to support existing mobility in the process.
In contrast, the criterion actor considers if there is a benefit by enabling the location
independent execution of a stationary activity. Therefore, the cross-process mobility of
the actor is investigated on the basis of the definition of mobile workers given in [7].
The predefined values of the criterion actor are high, medium and low. The indicators
are stationary workspace, mobile workforce, and frequent business trips. For example,
the value is high if the actor is part of a mobile workforce, rarely at his stationary work‐
space or often on business trips.

Process. The category process considers aspects given by the process itself. This
comprises, on the one hand, the effects of the improvement of the activity on the entire
process and, on the other hand, the improvement potential of the underlying information
ValueApping: An Analysis Method to Identify Value-Adding Mobile Enterprise Apps 229

Table 1. Criteria catalogue.

system. Therefore, the category is divided into two subcategories: relevance and current
information system. The category relevance contains the criteria frequency and acute‐
ness. Based on these criteria, the impact on the process by improving the respective
activity is analyzed. The criterion frequency refers to the frequency of execution of an
activity. Thereby, it is not differentiated if the activity is executed multiple times in one
process instance or if multiple process instance lead to frequent activity executions as
the potential impact of the activity is higher the more often it is executed in general. The
predefined values are often, regularly, and rarely. There are no concrete numbers as
these depend on industry-specific process conditions. The subcategory current infor‐
mation system considers the improvement potential regarding the current information
230 E. Hoos et al.

system. The criteria are digitalization, existence of devices, usability and sensors. For
instance, the criterion sensors investigates if the use of sensors has the potential to
improve the activity, e.g., by taking photo of a situation instead of describing it textually.

Technology Requirements. The category technology requirements analyzes techno‐

logical aspects of the application used in the activity. They are deduced from [23–25,
28]. The category is divided into performance aspects and software quality aspects. The
performance subcategory contains the following criteria: Data Volume of send and
receive, computing power, presentation and type of input. With these criteria, the
required performance can be matched with the different types of mobile technology. For
instance, the criterion presentation refers to the characteristics of small screens. It is
investigated if it is possible to present the data on small screens. Indicators are type of
the data, e.g., text or picture, and number of data sets. The subcategory software quality
refers to non-functional properties and contains the criteria availability and security.
Security is one of the biggest barriers to introduce mobile technology in enterprises [10].
In this paper, security refers to data security which can be divided into confidentiality,
integrity, authenticity, non-repudiation. The predefined values are high, medium and
low. For the determination, the risks of violating each aspect have to be considered.

Corporate Conditions. The category corporate conditions combines general organi‐

zational and technological conditions for the use of mobile technology in the enterprise.
Thereby, aspects of mobile readiness as well as the context of the usage have to be
considered [29]. Thus, the subcategories are individual, organizational and infrastruc‐
tural. Individual considers the user and the management and their readiness to use and
accept MEAs. For instance, the criterion user estimates if the users have a general affinity
for mobile devices. Indicators are technical interests of the user and whether he already
uses SMTDs. The predefined values are high, medium and low. If the value is high, then
the possibility that the user would use the devices is high. The subcategory organiza‐
tional refers to organizational aspects of the enterprises and includes the criteria mobile
devices and guidelines. The criterion mobile devices investigates if the actor already
employs mobile devices that he can reuse for other applications. Guidelines may
prescribe, for instance, that in some restricted company areas mobile device are not
allowed. Infrastructural contains one criterion, data communication. It represents the
availability of mobile networks.

4.2 App Potential

The app potential is a metric representing the improvement potentials for a process
activity when supported by a MEA. The app potential has two dimensions, mobilization
potential and app capability.
The mobilization potential refers to the aspect whether a mobile execution of the
activity is beneficial: the higher the mobilization potential, the higher the advantages of
using mobile technology in general. The app capability refers to the question, whether the
application supporting the activity is suited to be realized as an application on SMTDs.
ValueApping: An Analysis Method to Identify Value-Adding Mobile Enterprise Apps 231

Fig. 3. App management portfolio.

In order to determine the app potential, the criteria of the catalogue are mapped to
the two dimensions of the app potential. The numerical calculation is then based on
scored and weighted criteria values as explained in Sect. 5.2.
The app potential metric enables the ranking and prioritization of process activities
in a portfolio (see Sect. 4.3) and makes them comparable regarding their improvement
potential using MEAs.

4.3 App Management Portfolio

The app management portfolio is based on portfolio analysis concepts. The latter are
typically used for evaluating, selecting and managing research and development projects
in order to make strategic choices [30–32]. We adapted these concepts to the evaluation
and selection of process activities regarding mobile technology. The app management
portfolio groups the process activities into four categories according to their mobiliza‐
tion potential and their app capability. The goal is to define action recommendations for
each category. These recommendations focus on the type of IT technology which fits
best for each category. The four categories are flexible & easy-on-the-go, complex &
mobile, legacy & fixed, and fancy & pointless. The resulting portfolio is shown in
Fig. 3. The higher the app potential of an activity, the more it is positioned further up
on the right of the portfolio.
Activities in the flexible & easy-on-the-go category have a high mobilization poten‐
tial and a high app capability. That is, process improvements are high when using apps
for this activity. It is highly recommended to deduce a corresponding usage scenario for
a mobile app. For instance, if a mobile worker needs current information of an enterprise
backend system or has to record information on-the-go, these activities may be in the
flexible & easy-on-the-go category. A corresponding app could not only provide mobile
access but easily enrich the information by sensor data, e.g., photos, location, voice or
video as provided by most smartphones. The recorded information can be transmitted
directly to the backend instead of describing the situation textually on paper and trans‐
ferring it manually.
The complex & mobile category is characterized by a high mobilization potential and a
low app capability. That is, activities in that category can be improved, if their applica‐
232 E. Hoos et al.

tions run on mobile devices. However, the application is not suitable for running on SMTDs
due to, e.g., high performance requirements of the application. Hence, the actors of these
activities should be equipped with laptops enabled to connect to the enterprise IT backend.
For example, if a simulation model should be compared to the real world, the employee has
to go to this area with his mobile device. Simulation needs a lot of computing power, hence
a notebook might be suited. Writing a long report at the point of action is another example
for a notebook application because writing a text on touchscreens is not appropriate.
Low mobilization potential and low app capabilities are the characteristics of activities
positioned in the legacy & fixed quadrant. This implies that there are no improvements
when using mobile technology. Thus, there is a clear suggestion to refer to traditional
stationary technology like PCs.
The fancy & pointless category has low mobilization potential and high app capa‐
bilities. That is, it is possible to create an app for this application but the app does not
add value, because the execution of the activity is not improved. For instance, an engi‐
neer might use an app for mobile product data management without having mobile tasks.
Technology-driven approaches are in danger of producing apps for this type of process
activities. Activities in this category should be supported by stationary IT technology
although it is technologically possible to employ apps.
The boundaries of the quadrants can be varied according to the enterprise strategy.
By default, boundaries are based on half of the maximum values for mobilization poten‐
tial and app capability revealing quadrats of equal size. The numerical calculation of
these values is described in Sect. 5.3 and the categorization of activities in the portfolio
is detailed in Sect. 5.4.

4.4 Business Benefit Breakdown Structure and Business Benefit Radar Chart
The business benefit breakdown structure offers a mechanism to analyze the business
benefits of MEA usage for a certain process activity. The structure is based on the three
goal dimensions of process improvement, namely flexibility, time, and quality. For each
dimension, we identified major business benefits which can be achieved by the usage of
a MEA in a particular activity, e.g., a reduction of reaction times in the time dimension.
It has to be remarked that we do not consider the cost dimension, because a profound
cost analysis requires additional investment calculations regarding the use of informa‐
tion systems in organizations [33]. However, the benefits can be used as basis for a
traditional cost-benefit analysis.
The benefits were identified on the basis of both a comprehensive literature study
[26, 34–36] and a case-oriented investigation of mobile apps in the engineering domain
presented in [37]. To determine, whether a certain benefit can be realized for an activity,
we mapped related criteria of the criteria catalogue to each benefit. In the following, we
give an overview of the business benefit breakdown structure with major benefits according
to the three goal dimensions. Related criteria for each benefit are written in brackets.
Flexibility. In the context of ValueApping, flexibility means that the execution of the
activity can be modified or adapted at process runtime.
ValueApping: An Analysis Method to Identify Value-Adding Mobile Enterprise Apps 233

• Location Independent Task Execution ( actor): Due to the anywhere and anytime
characteristics, the actor is not restricted to perform the task in a specific place and
is enabled to perform his work even on-the-go. Hence, this is determined by the
criterion actor.
• Task Scheduling ( mobility of the task, availability): Different circumstances can
necessitate a rescheduling of a task during mobile work. For example, a service
engineer may receive an urgent service request while being at the customer. When
he is able to access the required information relevant for the new task in the back-
end system, he can reschedule his work immediately and reprioritize his tasks.
Therefore, the corresponding criteria are availability and task.
Quality. Quality refers to the data quality in the process. A higher quality refers to
fewer input errors as well as more precise information representation [38].
• Reduction of Input Errors ( digitalization, usability): Due to digitalization, media
breaks can be avoided, because the data are entered directly into the IT system. The
corresponding criterion is digitalization. In addition, the intuitive touch-based inter‐
action and targeted functionality of MEAs can avoid errors during activity execution
in general. This benefit can be realized if the usability of the existing IT system is
low. Thus, the corresponding criterion is usability.
• Contextualization and Enrichment of Information ( sensor): Device sensors can be
used to record data, which was not available before such as taking a photo or deter‐
mining the geographic location of the user. This enables an informational enrichment
by multimedia data as well as the additional contextualization of input data. The
relevant criterion for this benefit is sensor.
• Access to Real Time Data ( availability): Due to the mobile connection to the back-
end system, the user can access current data and base his decision on up-to-date
information. This is determined by the criterion availability.
Time. Time refers to the reduction of the entire lead time of the process. That is, the
time from the start of the first activity until the end of the last activity of the process.
• Reduction of Activity Execution Time ( sensor, usability): Due to a higher usability
and the sensor-based contextualization, the human computer interaction can be
improved. For example, taking a picture is faster than describing a situation textually.
Moreover, due to their strictly task-oriented user interface, MEAs only provide the
functionality which is actually necessary to perform an activity in contrast to complex
multi-functioned desktop applications. This leads to a faster execution of the activity,
especially for untrained employees. Hence, the corresponding criteria are sensor and
usability.
• Reduction of Waiting Time ( actor, availability): MEAs increase the reachability
and working ability of employees as described in Sect. 2.1. This is important, if the
start of an activity depends on receiving an event from a back-end system. With a
MEA the actor receives the notification of the event immediately and waiting times
between process activities can be reduced. This benefit is determined by the criteria
actor and availability.
234 E. Hoos et al.

• Elimination of Activities ( digitalization, sensor): One of the main optimization

potentials of MEAs is the elimination of activities for paper-based data acquisitioning
through digitalization. In addition, due to the integration of multiple IT functions in
one mobile device, further data processing activities can be eliminated, e.g., due to
the integrated camera, it is not necessary to use a separate camera and transfer the
data from the camera to the IT system. The corresponding criteria are sensor and
digitalization.
On the basis of the benefit breakdown structure, the business benefit radar chart (see
Fig. 4) aggregates the benefits for each analyzed activity according to the goal dimen‐
sions time, quality, and flexibility in a graphical manner. At this, the chart does not
represent a detailed quantified statement but an indicator for the achievable benefits and
can be used to relatively compare particular MEA usage scenarios of different process
activities. The value for each dimension ranges between 0 and 1 and represents the
proportion of achievable benefits compared to the maximum. The calculation of the
values is detailed in Sect. 5.4.

5 Analysis Steps

In this section, the analysis steps of ValueApping are explained, namely process analysis,
evaluation of app potential, recommendation generation and business benefit analysis.

5.1 Process Analysis

The process analysis refers to the application of the criteria catalogue and the determi‐
nation of the criteria values for a given process activity. It comprises four analysis
activities, one for each category of criteria. The entire procedure for process analysis is
shown in Fig. 5.
The input for the activity analysis of mobility depends on the application variants.
In the user-driven approach, the input is one activity whereas in the process-driven
approach the input is the entire process. Then, each activity is analyzed by determining
the values of the criteria from the category mobility of activity. To minimize the effort,
there is a condition for early termination after the analysis of mobility: If no mobility is
detected, then the analysis of the activity is terminated because mobility is the prereq‐
uisite for the use of mobile devices. No mobility is given, if the values of the criteria
actor and task are both low.

Fig. 4. Business benefit radar chart.

ValueApping: An Analysis Method to Identify Value-Adding Mobile Enterprise Apps 235

Fig. 5. Procedure and activities for process analysis.

After this step, the activities for the analysis of process aspects, the analysis of tech‐
nology requirements and the analysis of corporate conditions follow. Thereby, these
activities are executed in parallel. The advantages of dividing the process analysis into
four sub analyses are that the entire procedure is clearly structured and the results can
be reused. For example, if two activities are executed in the same environment, the
corporate conditions have to be analyzed only once and the results are used for both
activities.

5.2 App Potential Evaluation

In order to evaluate the app potential, the criteria and their values have to be mapped to
the dimensions of the app potential as explained in Sect. 4.2. For this purpose, the
influence of the criteria on the dimensions has to be examined. For example, the criterion
task in the category mobility of the activity has an influence on the mobilization potential
due to the fact that a mobile task would benefit from mobilization. Hence, the criterion
task is assigned to the dimension mobilization potential ( . In contrast, the crite‐
rion computing power is assigned to the dimension app capability ( , because
this differentiates laptops from SMTDs.
The next step is to specify the concrete influence of a criterion value on the dimension
it belongs to. Therefore, a scoring function maps the ordinal value kc of a
criterion C to a numerical value. The scoring function is based on a scoring matrix as
shown in Table 2. For example, if the criterion actor has the value high, then
and in case the value is low it is .

Table 2. Excerpt of the scoring matrix.

Score 3 2 1
Task High Medium Low
Actor High Medium Low
Frequency Often Regularly Rarely
… ….

In addition, the influence of individual criteria on the app potential can be adapted
by weighting each scored criterion C with weight wc as in . The
236 E. Hoos et al.

weighting enables enterprises to adapt the impact of the criteria according to their mobile
strategy. For example, if data security issues are very important, such as with product
data for manufacturing cars, the weight can be increased.
On this basis, the numerical values for the app potential of a process activity are
calculated as follows:

5.3 Recommendation Generation

The step recommendation generation positions the activities in the app management
portfolio and defines action recommendations for each portfolio category (see Sect. 4.3).
Process activities are positioned according to their values for app capability and mobi‐
lization potential. For example, activities with the app potential (0,0) belong to the cate‐
gory legacy & fixed. The higher the app potential of an activity, the more it is positioned
further up on the right of the portfolio.
Using this portfolio, the stakeholders can decide which activities should be supported
by apps and prioritize corresponding development projects. Hence, the enterprise gets
a structured overview about the app potential across various processes.

5.4 Business Benefit Analysis

In this step, the business benefits of MEA usage scenarios for activities located in the
flexible & easy-on-the-go quadrant in the portfolio are analyzed based on the business
benefit breakdown structure and the criteria (see Sect. 4.4): if one of the scored criteria
values related to a benefit is larger than zero, we assume that this benefit is likely to be
realized with the corresponding MEA. For example, the benefit reduction of execution
time has the corresponding criteria sensor and usability. Hence, this benefit will occur
if the value of the criterion sensor is yes or the value of the criterion usability is medium
or high. For the business benefit radar chart, all scored criteria values for all benefits are
then aggregated for each goal dimension. The calculation of the value for each goal
dimension is defined as follows:

It is important to remark that the value of does not denote the absolute quantified
improvement of a goal dimension but enables a relative comparison of different MEA
ValueApping: An Analysis Method to Identify Value-Adding Mobile Enterprise Apps 237

usage scenarios for certain process activities. For instance, does not repre‐
sent an improvement of the process quality by 50 % but, compared to another MEA
usage scenario with , the first has a higher impact on the quality dimension.

6 Case-Oriented Evaluation in the Automotive Industry

As an initial evaluation, we applied ValueApping in a real case at a large German car

manufacturer. At this, we used the method to analyze a concrete process in the engi‐
neering domain. In the following, we briefly describe the process and the analysis
procedure and discuss major results.

6.1 Modification Approval Process

The modification approval process is part of the car development process. During the
development of a car, a lot of change requests arise. For instance, the design of the seat
is changed or another breaking system should be used. However, single changes have
impacts on the whole car. For instance, it has to be checked whether the new seat design
fits the car’s interior. The modification approval ensures that the product data in the
product data management (PDM) system is in a consistent state despite modifications.
In general, a faster execution of the process is desirable to reduce development times.
For our analysis, a process description is needed. Therefore, we conducted interviews
with the organizational owners of the process to get a high level overview of the process
and deduce a simple process model. The process model is shown in Fig. 6. It consists of six
sequential activities. The process starts if product data is modified. Product data comprises
both product geometry representation in terms of computer-aided-design models and the
product structure in form of a bill of materials. When the modification is done, the engi‐
neer has to create a modification document including all relevant changes. Once the docu‐
ment is checked into the PDM system, the process starts. Then, the system forwards the
document to various persons with different responsibilities following a pre-determined
order. At first, the responsible person for this component, the creator himself or his boss,
has to perform the check modification record activity. This includes checking the docu‐
ment for correctness and completeness. After that, the activity verify packaging is
performed by the packaging manager. A package is a higher level component built from
multiple parts. For example, the worker checks if there is an installation space collision,
e.g., whether the new engine fits in the engine compartment. After that, the design vali‐
dator performs the activity verify design to ensure data quality. Then, the activity verify and
approve modification has to be executed by the technically responsible persons. First, the
team lead has to give his approval and then the department leader approves as well. If the
document received all required approvals, the documentarian performs the activity create
entries in PDM. With that, the modification is completely documented in the PDM and the
modification approval process finishes.
238 E. Hoos et al.

Fig. 6. Process model of the modification approval process.

This simple modelling is sufficient for our analysis, because all other important
aspects for mobile IT support, e.g., location and roles, are covered in the criteria cata‐
logue. Yet, for further stages like the development of suitable apps for the process, the
process model has to be extended by other process characteristics such as location,
actors, business domains and resources [39, 40].

6.2 Method Application and Results

On the basis of the process model described above, we applied ValueApping in the
process-driven approach in order to analyze the entire process. Thereby, we conducted
interviews with process experts to determine the criteria values.
On this basis, we investigated the mobility of each activity according to the procedure
described in Fig. 5. To this end, the criteria task and actor were used. We observed that
all tasks have a low mobility. The reason is that they are all executed at the actor’s
stationary workspace. However, during the evaluation of the criterion actor, two groups
of activities were identified. One group has actors with a low mobility and the other one
has actors with a high mobility. The activities create modification record document,
check document, verify packaging, verify design, and create entries in PDM have actors
with a low mobility because they are most of their working time at their stationary
workspaces. In contrast, the activities check record and verify and approve modification
have actors who are rarely at their workplaces. Thus, according to the termination
condition, we further analyzed only the activities from group two, check record and
verify and approve modification, and skip process analysis for group one.
Our analysis results of these activities revealed that the values of the sub–categories
process, performance, and individual have a positive influence on the app potential of
these activities, because the process is very important, so enhancement is beneficial for
the enterprise and the performance requirements make it possible to run the application
on SMTDs. In addition, employees and the corporate management welcome the usage
of SMTDs. However, security requirements are a big challenge. Product data are highly
sensitive and no unauthorized person should be able to access them.
After performing the app potential evaluation (see Fig. 7), two activities were posi‐
tioned in the category flexible & easy-on-the-go, namely check record and verify and
approve modification. For these activities, an app usage scenario was defined as a basis
for the development of a concrete app within the car manufacturer, called ApprovalApp.
The other activities create entries in PDM, check package, check design, and create
modification cannot be improved through mobile technology due to a low mobilization
potential.
The business benefit analysis for the ApprovalApp usage scenario revealed the
following benefits:
ValueApping: An Analysis Method to Identify Value-Adding Mobile Enterprise Apps 239

Fig. 7. App management portfolio of the modification approval process.

• Location-independent task execution

• Reduction of waiting time
• Access to real time data
The resulting business benefit radar chart (see Fig. 8) reveals that the ApprovalApp
improves flexibility, time, and quality of the modification approval process. Many actors
of the verify and approve modification activity are senior managers. Hence, they are
rarely at their stationary workplaces, most of the time they are on business trips and
meetings. With the ApprovalApp, they are informed immediately if a new approval task
occurs and they can perform the task right on-the-go. Furthermore, the decision can be
made on the basis on up-to-date product data.

6.3 Discussion
We discussed both the procedure of applying ValueApping as well as the concrete results
for the modification approval process with experts on mobile technology within the
industry partner and summarize major results in the following.
It was underlined that the strict structure and the systematic procedure of analysis
steps make the results of ValueApping comprehensible and transparent. Moreover, it

Fig. 8. Benefit radar chart of the ApprovalApp.

240 E. Hoos et al.

was emphasized that the portfolio-oriented visualization enables an easy communication

and representation of the analysis results especially for corporate management. Before,
various ideas for new MEAs were discussed within the industry partner without clear
prioritization. The portfolio helped to get an overview of all analyzed activities and
corresponding possibilities for new apps. This provided a sound basis for decision
making and prioritization of investments in mobile technology. On the one hand, poten‐
tial users could be convinced that their app ideas in the category fancy & pointless should
not be realized. On the other hand, IT responsibles developed a deeper understanding
for a business-driven view on mobile technology.
With respect to the analysis steps, the termination condition was recognized as
helpful because it decreased the analysis effort significantly. The approval modification
process comprised six activities and the analysis of four was terminated using the termi‐
nation condition. Yet, with respect to the criteria, additional indicators revealed to be
helpful in order to precisely determine the value of each criterion. At this, more fine-
grained values for some criteria like security and data volume would be helpful, too.
Considering the usage of MEAs in the modification approval process, the need for
supporting the activities check record and verify and approve modification through a
MEA was recognized by the industry partner. It was stated that an app has the potential
to reduce execution times and enhance flexibility of the process significantly as high‐
lighted by the business benefit radar chart.

7 Related Work

In this section, we give an overview of work related to our approach. A further qualitative
evaluation of ValueApping based on a comparison with similar approaches can be found
in our previous work [9]. For the discussion of related work, we differentiate three groups
of work with respect to mobile technology in business processes.
The first group comprises work on the general potential and impact as well as the
basic conditions for the use of mobile technology in business processes [26, 29, 34, 41].
These works discuss different high level aspects of mobile technology in enterprises such
as benefits of mobilizing processes, transformational impact of mobile technology and
mobile enterprise readiness. Yet, they do not address issues of a methodology to system‐
atically realize the benefits of mobile technology.
The second group comprises concepts which are similar to our ValueApping method.
Gumpp and Pousttchi propose a framework to evaluate mobile applications according
to their potential business benefits [7]. The framework is based on the theory of infor‐
mational added values and its application to mobile business. It constitutes a high level
approach and misses the detailed analysis of processes to derive concrete usage
scenarios. Gruhn et al. present a method called Mobile Process Landscaping to choose
a suitable mobile application to enhance business processes [5]. The authors make use
of typical return on investment concepts to analyze mobility in processes and evaluate
different mobile applications. Yet, they neither incorporate technological aspects, e.g.,
the complexity of data input, nor do they focus on the specific characteristics of MEAs.
Scherz defines criteria to identify mobile potential in business processes during a
ValueApping: An Analysis Method to Identify Value-Adding Mobile Enterprise Apps 241

condition-analysis as part of a classical system analysis [6]. These criteria are divided
into four categories, namely actor, process classification, data and information system
as well as devices. Yet, mobile apps are not addressed specifically.
The third group of work considers the usage of mobile apps in enterprises [10, 42, 43].
They point out that apps have a great potential to improve business processes, suggest
general application areas for apps and discuss selected app-oriented aspects, e.g., tech‐
nical requirements for the IT back-end. Yet, they do not focus on an analysis method‐
ology to identify concrete usage scenarios.

8 Conclusions and Future Work

In this work, we presented ValueApping, an analysis method to identify value-adding

usage scenarios for mobile enterprise apps in order to improve business processes.
ValueApping helps stakeholders to decide which type of IT technology fits best for a
given process activity. It is based on a comprehensive criteria catalog to systematically
analyze business processes and reveals an app management portfolio, which categorizes
process activities according to their improvement potential using MEAs Furthermore,
the business benefits of the resulting app usage scenarios are evaluated and represented
in a graphical manner.
ValueApping can not only be employed to identify usage scenarios for one process.
It can also be used to get a general view on mobile potentials of several processes in an
enterprise in order to identify cross-process synergies and prioritize company-wide
investments. On this basis, ValueApping enables both a systematic prioritization of IT
investments in mobile technology and a transparent IT portfolio management as part of
a mobile enterprise strategy.
Our future work comprises two aspects. On the one hand, we plan to implement
ValueApping as a software tool to ease the application of the method, especially the
determination of the criteria values. On the other hand, we want to extend ValueApping
in order to apply it at the business modelling stage and determine usage scenarios for
mobile enterprise apps during the initial design of a new business process.

Acknowlegements. The authors would like to thank the German Research Foundation (DFG)
for financial support of this project as part of the Graduate School of Excellence advanced
Manufacturing Engineering (GSaME) at the University of Stuttgart.

References

1. Gartner Inc.: Gartner Says Mobile App Stores Will See Annual Downloads Reach 102 Billion
in 2013. http://www.gartner.com/newsroom/id/2592315
2. Gartner Inc.: Gartner Identifies the Top 10 Strategic Technology Trends for 2014. http://
www.gartner.com/newsroom/id/2603623
3. Stieglitz, S., Brockmann, T.: Mobile enterprise. HMD Prax. der Wirtschaftsinformatik 49,
6–14 (2012)
4. Stieglitz, S., Brockmann, T.: Increasing organizational performance by transforming into a
mobile enterprise. MIS Q. Executive 11, 189–201 (2012)
242 E. Hoos et al.

5. Gruhn, V., Köhler, A., Klawes, R.: Modeling and analysis of mobile business processes. J.
Enterp. Inf. Managementscher 20, 657–676 (2007)
6. Scherz, M.: Mobile Business. Schaffung eines Bewusstseins für mobile Potenziale im
Geschäftsprozesskontext. VDM Verlag, Munich (2008). Dr. Müller
7. Gumpp, A., Pousttchi, K.: The “Mobility-M”-framework for application of mobile
technology in business processes. In: Proceedings of 35th GI-Jahrestagung, pp. 523–527
(2005)
8. Pousttchi, K., Becker, F.: Gestaltung mobil-integrierter Geschäftsprozesse. In: Meinhardt, S.,
Reich, S. (eds.) Mobile Computing. Dpunkt.verlag, Heidelberg (2012)
9. Hoos, E., Gröger, C., Kramer, S., Mitschang, B.: Improving business processes through
mobile apps - an analysis framework to identify value-added app usage scenarios. In:
Proceedings of the 16th International Conference on Enterprise Information Systems (ICEIS).
SciTePress (2014)
10. Gröger, C., Silcher, S., Westkämper, E., Mitschang, B.: Leveraging apps in manufacturing.
Framework App. Technol. Enterp. 7, 664–669 (2013). Procedia CIRP
11. Coulouris, G.F.: Distributed Systems. Concepts and Design. Addison-Wesley, Boston (2012)
12. Kornak, A., Teutloff, J., Welin-Berger, M.: Enterprise Guide to Gaining Business value from
Mobile Technologies. Wiley, Hoboken (2004)
13. Abolfazli, S., Sanaei, Z., Gani, A., Xia, F., Yang, L.T.: Rich mobile applications: genesis,
taxonomy, and open issues. J. Network Comput. Appl. 40, 345–362 (2014)
14. Pitt, L., Berthon, P., Robson, K.: Deciding when to use tablets for business applications. MIS
Q. Executive 10, 133–139 (2011)
15. Pitt, L.F., Parent, M., Junglas, I., Chan, A., Spyropoulou, S.: Integrating the smartphone into
a sound environmental information systems strategy: principles, practices and a research
agenda. J. Strateg. Inf. Syst. 20, 27–37 (2011)
16. Andersson, B., Henningsson, S.: Accentuated factors of handheld computing. In: Pooley, R.,
Coady, J., Schneider, C., Linger, H., Barry, C., Lang, M. (eds.) Information Systems
Development, pp. 293–304. Springer, New York (2013)
17. Sammer, T., Brechbühl, H., Back, A.: The new enterprise mobility: seizing the opportunities
and challenges in corporate mobile IT. In: Proceedings of the 19th Americas Conference on
Information Systems (2013)
18. Bragdon, A., Nelson, E., Li, Y., Hinckley, K.: Experimental analysis of touch-screen gesture
designs in mobile environments. In: Tan, D., Fitzpatrick, G., Gutwin, C., Begole, B., Kellogg,
W.A. (eds.) Proceedings of the 9th SIGCHI Conference on Human Factors in Computing
Systems, p. 403
19. Häikiö, J., Wallin, A., Isomursu, M., Ailisto, H., Matinmikko, T., Huomo, T.: Touch-based
user interface for elderly users. In: Cheok, A. D., Chittaro, L. (eds.) Proceedings of the 9th
International Conference on Human Computer Interaction with Mobile Devices and Services,
pp. 289–296
20. Kenney, M., Pon, B.: Structuring the smartphone industry: is the mobile internet OS platform
the key? J. Ind. Competition Trade 11, 239–261 (2011)
21. Oberheide, J., Jahanian, F.: When mobile is harder than fixed (and Vice Versa): demystifying
security challenges in mobile environments. In: Proceedings of the 11th Workshop on Mobile
Computing Systems and Applications, pp. 43–48. ACM, New York, NY, USA (2010)
22. Cansando, F., Vasconcelos, A., Santos, G.: Using multi-criteria analysis to evaluate enterprise
architecture scenarios. In: Proceedings of 14th International Conference on Enterprise
Information Systems (ICEIS), pp. 232–237 (2012)
23. Forman, G., Zahorjan, J.: The challenges of mobile computing. Computer 27, 38–47 (1994)
ValueApping: An Analysis Method to Identify Value-Adding Mobile Enterprise Apps 243

24. Krogstie, J.: Requirement engineering for mobile information systems. In: Proceedings of the
Seventh International Workshop on Requirements Engineering: Foundations for Software
Quality (REFSQ 2001) (2001)
25. Murugesan, S., Venkatakrishnan, B.: Addressing the challenges of Web applications on
mobile handheld devices. In: Proceedings of the International Conference on Mobile
Business, pp. 199–205. IEEE Computer Society, Los Alamitos, Calif (2005)
26. Nah, F.F.-H., Siau, K., Sheng, H.: The value of mobile applications: a utility company study.
Commun. ACM 48, 85–90 (2005)
27. Sarker, S., Wells, J.D.: Understanding mobile handheld device use and adoption. Commun.
ACM 46, 35–40 (2003)
28. Wasserman, A.I.: Software engineering issues for mobile application development. In:
Proceedings of the FSE/SDP Workshop on Future of Software Engineering Research, pp.
397–400. ACM, New York, NY, USA (2010)
29. Basole, R.: Mobilizing the enterprise: a conceptual model of transformational value and
enterprise readiness. In: 26th ASEM National Conference Proceedings (2005)
30. Bohanec, M., Rajkovič, V., Semolić, B., Pogačnik, A.: Knowledge-based portfolio analysis
for project evaluation. Inf. Manage. 28, 293–302 (1995)
31. Mikkola, J.H.: Portfolio management of R&D projects: implications for innovation
management. Technovation 21, 423–435 (2001)
32. Killen, C.P., Hunt, R.A., Kleinschmidt, E.J.: Project portfolio management for product
innovation. Int. J. Q. Reliab. Manage. 25, 24–38 (2008)
33. Ward, J., Peppard, J.: Strategic planning for information systems. J. Wiley, Chichester (2002)
34. Basole, R.C.: The value and impact of mobile information and communication technologies.
In: Proceedings of the IFAC Symposium on Analysis, Modeling and Evaluation of Human-
Machine Systems, pp. 1–7 (2004)
35. der Heijden, H., van, Valiente, P.: The value of mobility for business process performance:
evidence from sweden and the netherlands. In: Proceeding of European Conference on
Information Systems, p. 34 (2002)
36. Picoto, W.N., Palma-dos-Reis, A., Bélanger, F.: How does mobile business create value for
firms? In: 2010 Ninth International Conference on Mobile Business and 2010 Ninth Global
Mobility Roundtable (ICMB-GMR), pp. 9–16
37. Hoos, E., Gröger, C., Mitschang, B.: Mobile apps in engineering: a process-driven analysis
of business potentials and technical challenges. In: Proceedings of 9th CIRP Conference on
Intelligent Computation in Manufacturing Engineering (to be published)
38. Schmelzer, H.J., Sesselmann, W.: Geschäftsprozessmanagement in der Praxis. Kunden
zufrieden stellen - Produktivität steigern - Wert erhöhen. Hanser, Carl, München (2010)
39. Gao, S., Krogstie, J.: Capturing process knowledge for multi-channel information systems:
a case study. Int. J. Inf. Syst. Model. Design (IJISMD) 3, 78–98 (2012)
40. Gopalakrishnan, S., Krogstie, J., Sindre, G.: Capturing location in process models comparing
small adaptations of mainstream notation. Int. J. Inf. Syst. Model. Design 3, 24–45 (2012)
41. Gebauer, J., Shaw, M.J.: Success factors and impacts of mobile business applications: results
from a mobile e-procurement study. Int. J. Electron. Commer. 8, 19–41 (2004)
42. Lunani, M.: Enterprise Mobile Apps (2011)
43. Clevenger, N.C.: iPad in the Enterprise. Developing and deploying business applications.
Wiley, Indianapolis (2011)
 An Overview on Aspect-Oriented
Requirements Engineering Area

Paulo Afonso Parreira Júnior1,2(&) and Rosângela Dellosso Penteado1

1
Departament of Computer Science,
Federal University of São Carlos, São Carlos, Brazil
[email protected], [email protected]
2
Computer Science Course, Federal University of Goiás, Jataí, Goiás, Brazil

Abstract. Background: Aspect-Oriented Requirements Engineering (AORE) is

a research field that aims to provide appropriate strategies for identification,
modularization and composition of crosscutting concerns (also called
early-aspects). Several AORE approaches have been developed recently, with
different features, strengths and limitations. Goals: the aim of this paper is
threefold: (i) cataloguing existing AORE approaches based on the activities
encompassed by them; (ii) describing what types of techniques have been used
by these approaches for “Concern Identification and Classification” – a bottle-
neck activity; and (iii) identifying which are the most used means of publication
of AORE-based studies and how it has been the progress of these studies over
the years. Results: we have selected and analyzed 60 (sixty) papers and among
them, we identified 38 (thirty-eight) AORE distinct approaches. Some inter-
esting results are: (i) few approaches lead to “Conflict Identification and Res-
olution”, an activity responsible for discovering and treating the mutual
influence between different concerns existing in a software; (ii) there is a lack of
evaluation studies about already existing AORE approaches; (iii) the most
productive research institutions on AORE in the world are located in Lancaster
(UK) and Nova Lisboa (Portugal); among other.

Keywords: Systematic mapping
Concern identification and classification

Aspect-oriented requirements engineering
Crosscutting concerns

Early-Aspects

1 Introduction

The Requirements Engineering (RE) encompasses activities related to the elicitation

and analysis of information about the software: its requirements. Each sub-activity (or
task) performed during requirements elicitation will result in a document with the
textual description of all the software requirements. This document is then analyzed
and requirements are structured in individual units, such as viewpoints, goals, use
cases, scenarios. This is done in order to promote the Separation of Concerns - SoC [1],
i.e., the identification and modularization of pieces of the software that are relevant for
a particular purpose.
In the context of RE, a “concern” can be understood as a set of one or more
software requirements for a given purpose. For example, a security concern can
© Springer International Publishing Switzerland 2015
J. Cordeiro et al. (Eds.): ICEIS 2014, LNBIP 227, pp. 244–264, 2015.
DOI: 10.1007/978-3-319-22348-3_14
 An Overview on Aspect-Oriented Requirements Engineering Area 245

encompass several requirements related to the following goal: “ensuring that software
is secure”.
In an ideal scenario of software development, each concern should be allocated in a
specific module, which achieves its goals. When it occurs, the software is called
well-modularized, because all their concerns are clearly separated [2]. However, there are
some types of concerns, for which, this allocation is not possible, only using traditional
software engineering abstractions, such as viewpoints, goals, use cases, scenarios, among
others. These concerns are called “crosscutting concerns” or “early aspects” and are
defined as software requirements that are spread and tangled within other requirements.
Some examples of common crosscutting concerns include: Persistence, Security, Caching,
and Synchronization. The existence of crosscutting concerns can lead to lack of modu-
larization and make harder the software maintenance and evolution activities.
Aspect-Oriented Requirements Engineering (AORE) is a research field that pro-
vides the most appropriate strategies for identification, modularization and composition
of crosscutting concerns. A concern, in the context of AORE, encapsulates one or more
requirements specified by stakeholders, and a crosscutting concern is a concern whose
requirements cut across requirements of other software concerns. For example,
a security concern may contain a requirement related to “encryption” and another one
related to “checking access permissions”. In addition, this set of security requirements
can affect other software requirements, such as the requirement of sending registration
information to a customer, which is related to another software concern. Hence, the
security concern is called “crosscutting concern”.
Several AORE approaches have been proposed recently, although with different
features, strengths and limitations. However, there are few studies in the literature that
describe: (i) the amount of studies produced about this subject; (ii) the location of these
studies and in what time they were produced; (iii) which are the main AORE activities
explored by researchers; among others.
This paper shows the planning and execution of a Systematic Mapping (SM) [3, 4],
conducted with a focus on AORE, aiming to catalogue, identify and classify approa-
ches related to this subject. A SM can be understood as a wider review of primary
studies available in the literature, in order to identify the amount and types of studies
about a particular subject. It also may indicate the evolution of the published studies
about this subject over the years [3].
In this SM, we have selected and analyzed 60 (sixty) papers and among them, we
identified 38 (thirty-eight) AORE distinct approaches. Some interesting results are:
(i) few approaches lead to “Conflict Identification and Resolution”, an activity
responsible for discovering and treating the mutual influence between different con-
cerns existing in a software; (ii) the most of sixty studies consist of presenting new
AORE approaches or extensions of previous approaches - therefore, there is a lack of
evaluation studies on already existing approaches; (iii) few studies have been published
in journals, what can be a consequence of the item (ii); (iv) the most productive
research institutions on AORE in the world are located in Lancaster (UK) and Nova
Lisboa (Portugal); among other.
The obtained results in this SM can help other researchers to conduct further studies
from this work, proposing new methods/techniques/tools for AORE as well as com-
paring their proposals with the catalogue present in this paper. The remainder of this
246 P.A.P. Júnior and R.D. Penteado

paper is organized as follows: Sect. 2 presents an overview about Aspect-Oriented

Requirements Engineering; Sect. 3 illustrates the planning of the Systematic Mapping,
along with the research questions for which we have found answers in this work. In
Sect. 4, the answers to the research questions are given and discussed. In Sect. 5, some
threats to validity are discussed and, finally, Sect. 6 presents the final remarks and
proposals for future works.

2 Background

The SoC principle is based on the identification and modularization of pieces of the
software relating to a particular concept, goal or purpose [1]. Several traditional
approaches for software development, such as Object-Orientation, were created based
on this principle; however, some broad scope concerns (e.g., security, synchronization,
and logging) are not easy to be modularized and maintained separately during the
development of software. When these concerns are not appropriately modularized, the
software can contain tangled and scattered representations, making its understanding
and evolution harder.
An effective approach for RE must take into account the SoC principle and the need
to satisfy broad scope concerns [5]. AORE emerges as an attempt to encompass this
goal through the usage of specific strategies to modularize concerns that are difficult to
be isolated in individual modules (crosscutting concerns). The concern identification on
requirements level allows software engineers to think about them in an isolation way
from the beginning of software development, hence facilitating the creation/usage of
strategies to modularization.
Figure 1 shows a generic process for AORE, proposed by Chitchyan et al. [6],
which was developed based on other approaches available in the literature [5, 7–9]. The
rounded-corner rectangles represent the process activities.

Fig. 1. A generic process for AORE (Chitchyan et al. [6]).

 An Overview on Aspect-Oriented Requirements Engineering Area 247

From an initial available set of requirements, the activity “Concern Identification”

identifies and classifies software concerns as basis or crosscutting ones. The software
engineer knows the influences and constraints imposed by crosscutting concerns on
other software concerns, through the activity “Concern Relationship Identification”.
The activity “Concern Screening Out” aid software engineers to identify if there is
repetition in the list of identified concerns and decide which of these concerns are
relevant to the software. The activity “Concern Refinement” happens when there is a
need to change the set of already identified concern and relationships.
During the activity “Concern Representation”, the concerns are then represented in
a particular template. This template can vary according to the used AORE approach,
e.g., it can be a text, a use case model, viewpoints, among others. For example, the
approach developed by Rashid et al. [8, 10] represents base concerns using viewpoints;
in the Baniassad and Clarke’s approach [7, 11] it is defined themes as a new concept for
representation of base and crosscutting concerns. Still in the “Concern Representation”
activity, the software engineer can identify the need for refinement, for example, for
addition/removal concerns and/or relationships. Therefore, he/she can return to the
previous activities (Fig. 1). Finally, the base and crosscutting concerns represented in a
specific template must be composed and then analysed to identify conflicts between
them. These tasks are performed in the “Concern Composition” and “Conflict Identi-
fication and Resolution” activities. Then, identified conflicts are solved by the software
engineers with the help of stakeholders.
In general, the activities described in the process presented in Fig. 1 are aggregated
into four major activities, namely: “Concern Identification and Classification”, “Concern
Representation”, “Concern Composition” and “Conflict Identification and Resolution”.
These activities are used as a basis for cataloguing the AORE approaches (Sect. 4).

3 Systematic Mapping Planning

Kitchenham et al. [4] argue that a systematic review should be carried out following the
steps of planning, execution and documentation of the review and these steps can be used
in the context of a Systematic Mapping (SM). This section shows the planning and
strategy of execution of the SM performed in this work, according to the model of
Kitchenham et al. Further, a discussion about the results of this SM is presented in Sect. 4.

3.1 Research Questions

The SM conducted aims to answer the questions presented in Table 1. The first column
shows the code of the research question, which will be referenced throughout this text,
and the second one, shows its description.
The goal of the question Q1 is discovering the AORE approaches existing in the
literature and what activities they encompass. This question is important for at least two
reasons, it allows to: (i) catalogue existing approaches based on the activities encom-
passed by them; and (ii) indicate the approaches that deal with concern identification and
classification - which will facilitate obtaining data to answer the question Q3.
248 P.A.P. Júnior and R.D. Penteado

Table 1. Research questions for the SM.

# Description
Q1 What are the AORE approaches available in the literature and which activities they
cover?
Q2 What are the types of studies (Validation Study, Evaluation Study, Original Solution,
Adapted Solution, Philosophical Study, Opinion Papers and Experience Papers) that
have been proposed regarding the approaches identified in Question Q1?
Q3 What are the types of techniques that have been used by the approaches listed in
Question Q1 for concern identification and classification?
Q4 Which are the events (conferences, workshops, among others), journals, book chapters,
among others, where the approaches listed in Question Q1 have been published and
when this happened?

The question Q2 classifies the studies analysed in the SM based on the type of
study conducted by the authors: “Validation Study”, “Evaluation Study”, “Original
Solution”, “Adapted Solution”, “Philosophical Study”, “Opinion Papers” and “Expe-
rience Papers”. This classification was initially defined by Wieringa et al. [13] and is
used to guide the development of SMs proposed by Kai et al. [3]. An adaptation of the
original classification is presented in Table 2.

Table 2. Classification of the types of studies (adapted from Wieringa et al., [13]).
Classification Description
Validation It presents an evaluation of a proposed approach in simulated environments
Study (laboratories), through controlled experiments, case studies or proof of
concept.
Evaluation It presents a practical evaluation of a proposed approach, through
Study experiments on real industrial environment. This type of study, in
general, is conducted on more mature approaches, whose strengths have
been evaluated by means of “validation studies”.
Original It presents the description of an original solution to a given problem. The
Solution potential benefits and applicability of the proposed solution are presented
by small examples and good arguments by the authors of the study.
Adapted It presents a description of a solution to a given problem, but it is an
Solution adaptation of an existing solution. An adaptation may be considered as a
supplementary solution or a solution that minimize certain limitations of
the original approach. Similarly, the type of study “original solution”, the
potential benefits and applicability of the proposed solution are shown by
small examples and good argumentation from the authors of the study.
Philosophical It delineates a new way to look at existing approaches and structures them
Study in the form of a taxonomy, conceptual framework or catalogue.
Opinion Papers This type of study expresses the personal opinion of a (some) researcher(s)
about the benefits and/or limitations of a particular approach or how the
approach should be used.
Experience It consists in testimonials expressed by professionals/researches about how
Papers the approaches can be used in practice. It is the personal experience of
the author(s) from the usage of a particular approach.
 An Overview on Aspect-Oriented Requirements Engineering Area 249

It is important to highlight few points regarding the classification presented in

Table 2:
• This work adapts the classification proposed by Wieringa et al. in the following
way: the category “proposed solution” [13] was subdivided into “original solution”
and “adapted solution”. This adaptation let we know which approaches are new and
which ones are extensions of existing approaches;
• In the context of this paper, a study is classified as an “adapted solution” when it
comes to an extension of an existing AORE approach. For example, a study that
describes the development of a computational support for an approach proposed in
another study. However, studies that show solutions non-based on AORE
approaches are considered original solutions, because the extensions made in the
original approach, generally, are more significant. For example, a study that pre-
sents an approach for concern identification and classification based on use cases,
which is a traditional approach for software development, is classified as “original
solution”; and
• A study can be classified in more than one class described in Table 2. For example,
a study may provide an original solution to a problem while presenting a description
of a controlled experiment for evaluating this approach (“validation study”).
Regarding to the question Q3, some studies [2, 14] describe that concern identi-
fication and classification activity is a bottleneck in the AORE process. While this
activity serves as a basis for execution of the other activities, it is important to know:
(i) what types of techniques have been used for concern identification and classifica-
tion; (ii) what are the strengths and limitations of these techniques; and (iii) which of
them has been more used.
Finally, the question Q4 was proposed in order to know which are the most used
means of publication of AORE-based studies and how it has been the progress of these
studies over the years.
To answer these questions, it is necessary to conduct an investigation in the liter-
ature aiming to recovery primary studies, as full papers, experience reports, among
others. Kitchenham et al. [4] describe certain criteria to lead to an appropriate selection
of primary studies, they are: population, intervention and outcomes.
The population refers to the group of studies that will be observed. In this work, the
population consists of publications (full papers published in conference proceedings,
journals, among others) with a focus on Aspect-Oriented Requirements Engineering.
The intervention refers to what will be observed in SM. In this case, all type of AORE
approaches, techniques, methods and tools was observed. The outcomes refer to the
expected results at the end of the SM. In this case, the expected results are: (i) a
catalogue of AORE approaches available in the literature; (ii) the classification of the
main AORE activities encompassed by the identified approaches; (iii) a catalogue of
the main techniques used for concerns identification and classification; and (iv) the
presentation of the evolution of the publications related to AORE over the years, as
well as vehicles in which they have been published.
250 P.A.P. Júnior and R.D. Penteado

3.2 Search String and Keywords

The keywords used in the search string to obtain the primary studies of this SM are:
“requirements engineering”, “approach”, “aspect-oriented”, “aspect orientation”,
“tool”, “method” and “technique”. Based on this set of keywords, the search string was
generated: ((approach OR approaches OR technique OR techniques OR tool OR tools
OR method OR methods) AND (“aspect-orientation” OR “aspect-oriented”) AND
(“requirements engineering”)).
Some studies that were not retrieved by the search engines used in this SM were
manually added to the repository of the studies. These works were mainly obtained
from references found in publications considered relevant for this work.

3.3 Criteria for Inclusion of the Sources and Method

for the Search of Primary Studies
The IEEE Xplorer (ieeexplore.ieee.org) and Scopus (www.scopus.com) search sources
of primary studies were select based on the following criteria: (i) the source must index
publications in the field of Computer Science; (ii) the source should allow searches to
studies published in conference proceedings and journals via web; and (iii) the source
must provide advanced search engines, using the keywords and filters. The method
used to search for primary studies was the search engines available for these sources.
Beside, a manual review of the references from studies returned by these sources was
performed to obtain publications that were not retrieved by the search engines and that
are relevant for this SM.

3.4 Criteria for Inclusion of Primary Studies, Quality Criteria

and Methods for Evaluation of Primary Studies
The inclusion and exclusion criteria for primary studies are presented in Table 3.
Quality Criteria: as a way to assess the quality of selected primary studies, we have
considered only publications that present a complete and detailed description of the

Table 3. Inclusion and exclusion criteria for the primary studies.

Inclusion (I) and Exclusion (E) Criteria
(I1) The text of the study is written in English.
(I2) The complete version of the text of the study is available on web.
(I3) The study treats to the usage, adaptation, and/or creation of AORE approaches.
(E1) The study text is written in another language, not in English.
(E2) The complete version of the study text is not available on web.
(E3) The study does not treat to the usage, adaptation, and/or creation of AORE approaches.
(E4) The study is a duplicated version of another study.
(E5) The study is an older version of another study.
(E6) The study is a short paper or a poster up to 2 pages.
 An Overview on Aspect-Oriented Requirements Engineering Area 251

proposed approach. Thus, short papers and posters up to two pages were not consid-
ered. Moreover, only the newer versions of studies were analysed.
The Evaluation Method: consists in selecting primary studies according to the
inclusion criteria described in Table 3 and the quality criteria, as previously comment.
This protocol was applied for each study obtained from the research method described
in Sect. 3.3 and the selected papers were stored for later analysis.

3.5 Data Extraction from the Selected Primary Studies

The data extraction from the selected primary studies for this SM was performed in
four steps.
Step 1. One of the researchers has applied the research method to identify potential
primary studies. Based on the preliminary identified studies, a researcher read the title,
the abstract and the keywords of the publication, applying the criteria described in
Table 3. It was recovered 217 studies: 162 coming from the source Scopus and 55 from
the IEEE Xplorer; 112 of these studies were accepted and then, completely analyzed in
the second step of this SM.
The other ones were considered duplicated (48) or rejected (57). Duplicated studies
are those ones that consist of exactly the same publication, without any extension. This
occurs because the source Scopus can also index publications available in other
sources, such as IEEE, among others. Fifty-seven studies were rejected mainly due to
the exclusion criteria E3 (the study does not treat to the usage, adaptation, and/or
creation of AORE approaches.). Examples of rejected studies are those that propose the
usage of aspect orientation to create tools for requirements management. In this case,
the meaning of keyword “requirements engineering” was not directly related to AORE.
Step 2. In this step, the same researcher who has completed Step 1 also has read the full
text of the 112 accepted studies. The criteria described in Table 3 were reapplied, as
well as the quality criteria (Sect. 3.4). Several studies were rejected because they were
short papers (exclusion criterion E6) without enough information for answering the
research questions of this work.
The classification of the primary studies after finishing Step 2 can be seen in Fig. 2.
It can be noticed that some duplicated studies were identified yet. This occurred
because the researcher has not detected this situation, while performing Step 1. The
obtained results reinforces that Systematic Mapping must be done at stages, as well as
by more than one researcher [3, 4].
Due to the manual insertion of some relevant studies referenced in the selected
papers of the Step 1, the amount of studies has increased in 14 (fourteen). It is
important to mention these studies also were submitted to the same set of inclusion and
quality criteria already discussed in the paper.
Step 3. The results of the Step 2 were reviewed by another researcher involved in this
study, so that any disagreements were discussed and resolved. There was not need to
change the previously selected set of studies, but the interaction between the
researchers was important for the next step (Step 4), as it will be explained below.
252 P.A.P. Júnior and R.D. Penteado

Step 4. Finally, the resulting set of primary studies was used to extract the information
required to answer the questions listed at the beginning of this study (Table 1). In this
step, the collaboration between the researches was important to reduce the interpreta-
tion errors about some data extracted to the studies. The results of this step are pre-
sented in detail in Sect. 4.

4 Results

In this section, answers to the research questions for this SM (Table 1) are presented.
The data needed to answer the question Q1 are in Table 4; there are 38 approaches
identified and analysed in this SM. The columns 1 and 2 show the code of the
approach, used to identify it in other parts of this paper, the name of the approach and
the reference of the study(ies) that present(s) it, respectively.
If there is not a specific name for an approach, we have used the title of the study in
which this approach was presented. Then, the reader can find, at any time, what are the
AORE approaches analysed in this SM and which studies are related to them. Columns
3, 4, 5 and 6 of Table 4 describe which AORE activities, as discussed in Sect. 2, are
encompassed by the identified approaches.
It is possible to notice that many approaches include the concern identification
and classification, representation and composition activities. However, there is a lack of
approaches related to the “Conflict Identification and Resolution” activity. Other interesting
points are that only 16 % of the analysed approaches (7, 8, 13, 15, 18 and 23) are complete,
i.e., include all activities related to AORE and 55 % of them encompass just one or two
activities. This provides indications that conducting studies on a specific AORE activity or a
small subset of activities can be an interesting strategy of research instead of trying to
develop approaches that deal with all activities. A final point to be emphasized with respect
to Table 4 is that not every 60 studies analysed in this SM are referenced in this table. This
occurs because some studies are related to the usage and/or comparison of some AORE
approaches, i.e., they did not develop or extend any approach.
In order to answer the question Q2, Fig. 3 presents the classification of primary
studies analysed in this SM according to the classes described in Table 2.

Fig. 2. Classification of the Fig. 3. Classification of Primary Studies.

studies.
 An Overview on Aspect-Oriented Requirements Engineering Area 253

Table 4. AORE approaches.

# Approach AORE Activities
CIC CR CC CIR
1 An approach for crosscutting concern identification at requirements X
level using NLP [16–18]
2 ACE - Aspect Clustering Engine [19] X
3 AWC - Aspect Weaving Connector [20] X X
4 RDL - Requirements Description Language [21, 22] X
5 ASSD - Aspects Specification for the Space Domain [23] X X
6 A semi-automatic strategy to identify crosscutting concerns in X
PL-AOVgraph requirement models [24]
7 EA-Miner [24, 25] X X X X
8 NFR/AUC [26, 27] X X X X
9 DERAF [28] X X
10 An evolutionary model of requirements correctness with early X X X
aspects [29]
11 AORE/XML [30] X X X
12 Theme [7, 11] X X X
13 AspOrAs [31, 32] X X X X
14 EA-Analyzer [33] X
15 AORE with Arcade [8, 10] X X X X
16 PROBE [34] X
17 MAST - Modeling Aspectual Scenarios with Theme [35] X X X
18 Integrating Problem Frames with Aspects [36] X X X X
19 Interaction Analysis in Aspect-Oriented Models [37] X
20 Isolating and relating concerns in requirements using latent semantic X
analysis [38]
21 ADORA [39] X
22 Multi-ComBO [40] X X
23 Multi-Dimensional Separation of Concerns in Requirements X X X X
Engineering [5]
24 Concern Interaction Graph (CIG) [41, 42] X
25 On the discovery of candidate aspects in software requirements [43] X
26 Promoting the software evolution in AOSD with early aspects: X X
Architecture-oriented model-based pointcuts [44]
27 RCT - Requirements Composition Table [45] X X X
28 AOZCL [46] X
29 Scenario Modeling with Aspects [15] X X X
30 VisualAORE [47] X
31 Aspectual i* Model [48] X X
32 AO-ADL [49] X
33 AoUCM-to-RAM [50, 51] X X
(Continued)
254 P.A.P. Júnior and R.D. Penteado

Table 4. (Continued)
# Approach AORE Activities
CIC CR CC CIR
34 Using tagging to identify and organize concerns during X
pre-requirements analysis [52]
35 VGraph - From Goals to Aspects [9] X X X
36 AORA - Aspect-Oriented Requirements Analysis [53] X X X
37 AoURN - Aspect-oriented User Requirements Notation [54] X X X
38 RAM - Reusable Aspect Models [55] X X
Amount of approaches for each AORE activity 22 26 22 12
Subtitle: CIC – Concern Identification and Classification; CR – Concern
Representation; CC – Concern Composition; CDR– Conflict Identification and Resolution.

Seventy nine percent of the studies were classified as an “Original Solution” or an

“Adapted Solution” and there were few validation studies and none evaluation studies.
This fact calls our attention, because many approaches are being used/adapted without
having been submitted to evaluation studies. In addition, new approaches are being
proposed without knowing the real accuracy of the existing approaches.
Another evidence about the previously affirmation is presented in Table 5, that
presents: (i) the code of the AORE approach; (ii) the year of the first publication of this
approach; (iii) the years of publication corresponding to adaptations of this approach;
(iv) the references of the approaches used as basis for development of this approach;
and (v) the references of the studies that evaluate this approach.

Table 5. Evolution of the AORE approaches.

Performed Performed
Evolution Base Based
# Proposal Evaluation # Proposal Evolutions Evaluation
s d on on
s s
1 2008 2011 12 - 20 2006 - 12 -

2 2007 - 20 - 21 2007 - - -

3 2007 - - - 22 2008 - 36 -

4 2007 2008 - [22] 23 2005 - - [2]

5 2008 - 36 - 24 2009 - 29 -
6 2013 - 35 - 25 2009 - - -
7 2005 2006 - [14] 26 2009 - - -
8 2009 2010 - - 27 2012 - - [22]
9 2007 - - - 28 2007 - - -
10 2007 - - - 29 2004 - - -
11 2006 - - [2] 30 2010 - - [47]
12 2004 2005 - [14] 31 2010 - - -
13 2005 2008 - - 32 2007 - - -
14 2009 - - - 33 2007 2011 37; 38 -
15 2002 2003 - [2, 22] 34 2009 - - -
16 2004 - - - 35 2004 - - [2, 22]
12;
17 2010 -
29
- 36 2003 - - -
18 2009 - - - 37 2010 - - [54]
19 2006 - - - 38 2009 - - -
 An Overview on Aspect-Oriented Requirements Engineering Area 255

Approximately 74 % of the approaches (28 of 38) were not evaluated through case
studies, controlled experiments, among others, performed in a laboratory or an
industrial environment. In addition, many approaches (55 %: 2, 3, 5, 6, 9, 10, 14, 16,
17, 18, 19, 21, 22, 23, 24, 25, 26, 28, 31, 32 and 34) have been proposed and then they
have not been adapted, evaluated or used as a basis for other approaches anymore. In
other hand, some approaches that have been evaluated, adapted and/or used as a basis
for other approaches are: 4, 7, 11, 12, 15, 20, 23, 27, 29, 30, 35, 36, 37 and 38. The
approaches 15 [8, 10] and 35 [9] have been evaluated in more than one experimental
study. The approaches 12 [7, 11] and 29 [15] were used as basis for, at least, two other
approaches.
With regard to the question Q3, Table 6 presents the name and description of five
concern identification and classification techniques and the approaches that use them.
Some experimental studies [2, 14] describe that concern identification and classification
activity is a bottleneck in the AORE process. Then, knowing the techniques used in this
activity can help professionals and researchers to obtain better strategies to perform this
activity.
As can be seen the most used technique in different approaches is “Manual
Analysis of the Requirements Document by the Software Engineers with the Aid of
Guidelines”. Despite being limited to large scale software, this technique has promising
benefits, such as minimizing the dependence of users’ experience during the applica-
tion of the approach.
This is an indication that this technique has significant benefits for the concern
identification and classification and should be studied more carefully. Another
important point to notice is that few approaches (7, 23 and 27) use more than one
technique for concern identification and classification. This fact can be an interest
research field, because the usage of combined techniques can lead to higher accuracy of
AORE approaches.
Finally, to answer the question Q4, two bubble charts were built and they are
presented, respectively, in Figs. 4 and 5. Regarding to these figures, it is important to
comment that:
(i) the distribution of the published studies on conference proceedings is in Y-axis
of the graph of the Fig. 4;
(ii) the distribution of the published studies on journals, books and other vehicles of
scientific publication is in Y-axis of the graph of the Fig. 5. Aiming to simplify
the visualization of these graphs, only the initials of the events/journals was used
to identify them; and
(iii) the amount of publications per event, journal or book and the year in which they
occurred are presented in X-axis of the graphs of Figs. 4 and 5.
It may be notice that most publications (eleven) comes from the Workshop in
Aspect-Oriented Requirements Engineering and Architecture Design (EA). This makes
sense, because this is an event dedicated to publishing works related to AORE. Another
event that has a relevant amount of publications related to AORE (eight) is the
International Requirements Engineering (RE), a good and well-known conference in
the RE field.
256 P.A.P. Júnior and R.D. Penteado

Another important point to be observed in Figs. 4 and 5 is the evolution of AORE

publications over the years. It is possible to notice that there was a great amount of
publications between 2007 and 2009. In this period, the scientific community have
published 52 % of all studies published from 2002 to 2013. Finally, we also can
observe that most of studies have been published in conference proceedings and only
eight studies (13 %) were published in journals.
This indication is consistent to what we have said about the lack of evaluation
studies on the existing approaches. Since the approaches are not mature enough, i.e.,
the evidences of the robustness and accuracy of such approaches are fragile, publishing
them in good journals may be a hard task.
Beside the answers for the questions of Table 1, we aimed to know what are the
major research institutions and countries involved with AORE. This information is
important because it can help to improve: (i) the dissemination of the work carried out
by research groups; and (ii) the cooperation between different research institutions,
interested in the same area, i.e., AORE.
Regarding the distribution of studies among research institutions, the graph of
Fig. 6 presents the name of the research institution and the percentage of studies
published by this institution. Lancaster University (UK) and University of Nova
Lisboa (Portugal) are the most productive research institutions on AORE in the
world, with a total of 18.5 % and 21 % of the studies published on this area, respec-
tively. All other research institutions, with only one published study, were grouped in
the category “Other ones”.
Once the institutions located in Lancaster and Nova Lisboa have published the most
of studies related to AORE, countries that had the greatest amount of published studies
in this area were Portugal and the UK, as can be seen in the graph of Fig. 7 (besides the
University of Nova Lisboa, Polytechnic Institute of Beja also had publications on
AORE and is located in Portugal). In third place is Brazil, with 13.6 % of the amount of
published studies on AORE. The main universities in Brazil responsible for these
publications are Federal University of Rio Grande do Norte and Federal University of
Rio de Janeiro (Fig. 6). All other countries, with only one published study, were
grouped in the category “Other ones”.
Another relevant point that can be raised from this SM, but that was not system-
atized through research questions, is the identification of the main researchers involved
with AORE in the world. The graph in Fig. 8 presents the word cloud (generated by
Word Cloud Generator tool, available at: http://www.jasondavies.com/wordcloud/#)
generated from the names of the authors of the studies analyzed in this SM. From this
graph is possible to identify the main researchers involved with AORE as João Araújo
(University of Nova Lisboa), Ana Moreira (University of Nova Lisboa), Awais Rashid
(Lancaster University), among others. This information is important because it can
serve as a “control” for performing other secondary studies, such as Systematic
Mappings or Systematic Reviews. A control can be used to judge whether the search
string allows recovering relevant studies or not.
 An Overview on Aspect-Oriented Requirements Engineering Area 257

Table 6. Techniques for concern identification and classification.

Concern Identification Description
and Classification
Technique
Natural Language It is based on NLP techniques such as part-of-speech,
Processing (NLP) lemmatization (approach 7), among others to find keywords of
the text of requirements document that are related to with some
kind of concern. According to the analyzed studies, this
technique was not good for identification of implicit concerns,
i.e., concern that are not explicitly described in the text of the
requirements document.
Approaches that use this technique: 1, 7 and 23.
Probabilistic Models and It is based on statistical models, such as (Latent Semantic Analysis
Clustering - approach 20) and clustering techniques such as the use of tags
(approach 34) to find concern candidates. As this is a technique
based on statistical analysis, it does not usually bring good
results when the requirements document is small, i.e., when the
sample is small.
Approaches that use this technique: 2, 20, 25 and 34.
Manual Analysis of In this type of technique, the software engineer performs a manual
Requirements inspection in the requirements document trying to discover the
Document by Software software concern. As limitations, we have: (i) the results
Engineers obtained with this technique are strongly dependent on the
experience of those who apply it; and (ii) this technique is
error-prone, difficult to replicate its application and has high cost
of execution when requirements documents of large software are
used.
Approaches that use this technique: 10, 13, 15 and 17.
Manual Analysis of It is similar to the technique of manual analysis, but differs by the
Requirements fact that users of this technique have guidelines that can assist
Document by Software them during the process of concern identification. One type of
Engineers guidelines quite common is non-functional requirements
with the Aid of catalogs, such as proposed by Chung and Leite [57]. This type of
Guidelines technique can minimizing the dependence of the user experience
that applies it, however, it remains costly to be performed on
large software. In addition, it takes a certain user experience to
understand and follow the guidelines.
Approaches that use this technique: 5, 7, 8, 11, 23, 27, 29, 31, 36 and 37.
Software Visualization It is based on visualization techniques to help the user identify the
software concerns. A type of well-known visualization if “Action
Views”, proposed by Baniassad and Clarke (approach - 12).
A limitation of this technique is that for building the
visualizations, usually, the user must perform a manual
inspection from the requirements document, i.e., it suffer the
same problems of technical based on manual analysis of
requirements document, cited above.
Approaches that use this technique: 6, 18, 12 and 27.
258 P.A.P. Júnior and R.D. Penteado

Fig. 4. Distribution of the published Fig. 5. Distribution of the published studies on

studies on conference proceedings. journals and books.

Fig. 6. Distribution of the studies regarding to research institutions.

5 Threats to Validity

Primary Studies Selection. Aiming at ensuring an unbiased selection process, we

defined research questions and devised inclusion and exclusion criteria we believe are
detailed enough to provide an assessment of how the final set of primary studies was
obtained. However, we cannot rule out threats from a quality assessment perspective,
we simply selected studies without assigning any scores. In addition, we wanted to be
 An Overview on Aspect-Oriented Requirements Engineering Area 259

Fig. 7. Distribution of the studies regarding to the country where the research institutions are
located.

Fig. 8. Word cloud derived from the names of the authors of the studies analyzed.

as inclusive as possible, thus no limits were placed on date of publication and we

avoided imposing many restrictions on primary study selection since we wanted a
broad overview of the research area.
Missing Important Primary Studies. The search for primary studies was conducted
in two well-known search engines (Scopus and IEEE Xplorer), even though it is rather
possible we have missed some relevant primary studies. Nevertheless, this threat was
mitigated by selecting search engines which have been regarded as the most relevant
scientific sources [56] and therefore prone to contain the majority of the important
studies.
Reviewers’ Reliability. All the reviewers of this study are researchers in the software
engineering, focused on the aspect-oriented development, requirements engineering
and aspect mining. Therefore, we are not aware of any bias we may have introduced
during the analyses.
Data Extraction. Another threat for this review refers to how the data were extracted
from the digital libraries, since not all the information was obvious to answer the
questions and some data had to be interpreted. Therefore, in order to ensure the
validity, multiple sources of data were analysed, i.e. papers, technical reports, white
papers. Furthermore, in the event of a disagreement between the two primary
reviewers, a third reviewer acted as an arbitrator to ensure full agreement was reached.
260 P.A.P. Júnior and R.D. Penteado

6 Final Remarks

In this paper, we presented a SM of Aspect-Oriented Requirements Engineering

(AORE), based on the process described by Kitchenham et al., [4].
The SM presented was conducted as the planning described in Sect. 3. Through an
examination of 60 primary studies related to AORE approaches, this review has pre-
sented 38 different approaches. The steps outlined in this plan were sufficient to obtain
relevant primary studies, which generated the data needed to answer the research
questions (Table 1).
Summarizing the results observed from this SM:
1. We have identified 38 (thirty-eight) distinct AORE approaches;
2. The most of identified approaches are related to concern identification and classi-
fication, representation and composition activities; we have notice there is a lack of
studies based on conflicts detection and resolution;
3. The most studies analysed in this SM consist of presenting either new AORE
approaches or adaptation of existing approaches; this indicates that more evaluation
studies, based on these existing approaches, need to be performed to verify the real
accuracy of them;
4. We have identified five different types of techniques for concern identification and
classification, which are used by the AORE approaches: “Natural Language Pro-
cessing (NLP)”, “Probabilistic Models and Clustering”, “Manual Analysis of
Requirements Document by Software Engineers”, “Manual Analysis of Require-
ments Document by Software Engineers with the Aid of Guidelines” and “Software
Visualization”. The most used technique is the “Manual Analysis of Requirements
Document by Software Engineers with the Aid of Guidelines”. Despite being
limited to large software, it has promising benefits, such as minimizing the
dependence users’ experience during the application of the approach;
5. It was notice that most of the studies has been published in conference proceedings,
which reinforces the idea that many approaches have been proposed, but few of
them are mature enough to be published in journals; and
6. Finally, data show that the most productive research institutions on AORE in the
world are Lancaster University and University of Nova Lisboa, located in UK and
Portugal, respectively.
Researchers can use this SM as a base for advancing the field, while practitioners
can use it to identify approaches that are well-suited to their needs. With the results
obtained through this SM, it was possible to develop a set of comparison criteria for
AORE approaches, based on common features and variability of the approaches ana-
lysed in this SM [58]. Such criteria were then applied on six of the main AORE
approaches: 7, 11, 12, 15, 23 and 29. The results can serve as a guide so that users can
choose the approach that best meets their needs, and to facilitate the conduct of
research in AORE. The main future directions that emerged from this mapping are the
need for empirical, comparative evaluations and the opportunity for developing com-
bined AORE approaches.
 An Overview on Aspect-Oriented Requirements Engineering Area 261

References
1. Dijkstra, E.W.: A Discipline of Programming. Pearson Prentice Hall, Englewood Cliffs.
p. 217. ISBN: 978-0132158718, 1976
2. Sampaio, A., Greenwood, P., Garcia, A.F., Rashid, A.A.: Comparative study of
aspect-oriented requirements engineering approaches. In: 1st International Symposium on
Empirical Software Engineering and Measurement, pp. 166–175 (2007)
3. Kai, P., Robert, F., Shadid, M., Michael, M.: Systematic mapping studies in software
engineering. In: 12th International Conference on Evaluation and Assessment in Software
Engineering Proceedings, Swinton, UK, pp. 68–77 (2008)
4. Kitchenham, B., Charters, S.: Guidelines for performing systematic literature reviews in
software engineering. Technical Report. Keele University and Durham University (2007)
5. Moreira, A., Rashid, A., Araújo, J.: Multi-dimensional separation of concerns in
requirements engineering. In: Proceedings of the 13th International Conference on
Requirements Engineering (RE), pp. 285–296 (2005)
6. Chitchyan, R., Sampaio, A., Rashid, A., Rayson, P.: A tool suite for aspect-oriented
requirements engineering. In International Workshop on Early Aspects. ACM, pp. 19–26
(2006)
7. Baniassad, E., Clarke, S.: Theme: an approach for aspect-oriented analysis and design. In:
26th International Conference on Software Engineering (ICSE 2004) (2004)
8. Rashid, A., Sawyer, P., Moreira, A., Araújo, J.: Early aspects: a model for aspect-oriented
requirements engineering. In: International Conference on Requirements Engineering
(RE) (2002)
9. Yijun, Y., Leite, J.C.S.P., Mylopoulos, J.: From goals to aspects: discovering aspects from
requirements goal models. In: International Conference on Requirements Engineering
(RE) (2004)
10. Rashid, A., Moreira, A., Araújo, J.: Modularisation and composition of aspectual
requirements. In: 2nd International Conference on Aspect-Oriented Software Development
(AOSD 2003). ACM (2003)
11. Clarke, S., Baniassad, E.: Aspect-Oriented Analysis and Design: The Theme Approach.
Addison-Wesley, New York (2005)
12. Kitchenham, B., et al.: Systematic literature reviews in software engineering – a tertiary
study. Inf. Softw. Technol. 52, 792–805 (2010)
13. Wieringa, R., Maiden, N.A.M., Mead, N.R., Rolland, C.: Requirements engineering paper
classification and evaluation criteria: a proposal and a discussion. Requirements Eng. 11(1),
102–107 (2006)
14. Herrera, J., et al.: Revealing crosscutting concerns in textual requirements documents: an
exploratory study with industry systems. In: 26th Brazilian Symposium on Software
Engineering, pp. 111–120 (2012)
15. Whittle, J., Araújo, J.: Scenario modeling with aspects. IEEE Softw. 151(4), 157–172 (2004)
16. Ali, B.S., Kasirun, Z.M.D.: An approach for crosscutting concern identification at
requirements level using NLP. Int. J. Phy. Sci. 6(11), 2718–2730 (2011)
17. Ali, B.S., Kasirun, Z.M.: 3CI: a tool for crosscutting concern identification. In: Proceedings
of the International Conference on Computational Intelligence for Modeling Control and
Automation, Vienna, Austria, pp. 351–355 (2008a)
18. Ali, B.S., Kasirun, Z.M.: Developing tool for crosscutting concern identification using NLP.
In: Proceedings of the International Symposium on Information Technology, Kuala Lumpur,
Malaysia (2008b)
262 P.A.P. Júnior and R.D. Penteado

19. Duan, C., Cleland-Huang, J.: A clustering technique for early detection of dominant and
recessive crosscutting concerns. In: Proceedings of the International Conference on Software
Engineering, Minneapolis, MN (2007)
20. Lau, Y., Zhao, W., Peng, X., Tang, S.: A connector-centric approach to aspect-oriented
software evolution. In: Proceedings of the International Computer Software and
Applications Conference, Beijing, China, pp. 391–396 (2007)
21. Weston, N., Chitchyan, R., Rashid, A.: A formal approach to semantic composition of
aspect-oriented requirements. In: Proceedings of the 16th IEEE International Requirements
Engineering Conference, Catalunya, Spain, pp. 173–182 (2008)
22. Chitchyan, R., et al.: Semantic vs. syntactic compositions in aspect-oriented requirements
engineering: An empirical study. In: Proceedings of the 8th International Conference on
AOSD, Virginia, USA, pp. 149–160 (2009)
23. Agostinho, S., et al.: A Metadata-driven approach for aspect-oriented requirements analysis.
In: Proceedings of the 10th International Conference on Enterprise Information Systems,
Barcelona, Spain, pp. 129–136 (2008)
24. Medeiros, M., Silva, L., Medeiros, A.L.: A semi-automatic strategy to identify crosscutting
concerns in PL-AOVgraph requirement models. In: Proceedings of the Workshop on
Requirements Engineering, Rio de Janeiro, Rio de Janeiro, pp. 46–59 (2013)
25. Sampaio, A., Chitchyan, R., Rashid, A., Rayson, P.: EA-Miner: a tool for automating
aspect-oriented requirements identification. In: Proceedings of the International Conference
Automated Software Engineering, California, USA, pp. 353–355 (2005)
26. Monteiro, M.P., Fernandes, J.M.: Towards a catalogue of refactorings and code smells for
aspectj. In: Rashid, A., Akşit, M. (eds.) Transactions on Aspect-Oriented Software
Development I. LNCS, vol. 3880, pp. 214–258. Springer, Heidelberg (2006). Zheng, X.,
Liu, X., Liu, S.: Use case and non-functional scenario template-based approach to identify
aspects. In: Proceedings of the 2nd International Conference on Computer Engineering and
Applications, Bali Island, Indonesia, pp. 89–93 (2010).
27. Liu, X., Liu, S., Zheng, X.: Adapting the NFR framework to aspectual use-case driven
approach. In: Proceedings of the 7th International Conference on Software Engineering
Research, Management and Applications, Hainan Island, China, pp. 209–214 (2009)
28. Wehrmeister, M.A., Freitas, E.P., Pereira, C.E., Wagner, F.R.: An aspect-oriented approach
for dealing with non-functional requirements in a model-driven development of distributed
embedded real-time systems. In: Proceedings of the 10th International Symposium on
Object and Component-Oriented Real-Time Distributed Computing, Orlando, Florida, USA,
pp. 428–432 (2008)
29. Araújo, J., Zowghi, D., Moreira, A.: An evolutionary model of requirements correctness
with early aspects. In: Proceedings of the 9th International Workshop on Principles of
Software Evolution, Dubrovnik, Croatia, pp. 67–70 (2007)
30. Soeiro, E., Brito, I.S., Moreira, A.: An XML-based language for specification and
composition of aspectual concerns. In: Proceedings of the 8th International Conference on
Enterprise Information Systems, Paphos, Cyprus (2006)
31. Ribeiro, J.C., Araújo, J.: AspOrAS: a requirements agile approach based on scenarios and
aspects. In: Proceedings of the 2nd International Conference on Research Challenges in
Information Science, Marrakech, Morocco, pp. 313–323 (2008)
32. Araújo, J., Ribeiro, J.C.: Towards an aspect-oriented agile requirements approach. In:
Proceedings of the International Workshop on Principles of Software Evolution, Lisbon,
Portugal, pp. 140–143 (2005)
33. Sardinha, A., Chitchyan, R., Weston, N., Greenwood, P., Rashid, A.: EA-Analyzer:
automating conflict detection in aspect-oriented requirements. In: Proceedings of the 24th
 An Overview on Aspect-Oriented Requirements Engineering Area 263

International Conference on Automated Software Engineering, Auckland, New Zealand,

pp. 530–534 (2009)
34. Katz, S., Rashid, A.: From aspectual requirements to proof obligations for aspect-oriented
systems. In: Proceedings of the IEEE International Conference on Requirements
Engineering, Kyoto, Japan, pp. 48–57 (2004)
35. Penim, A.S., Araújo, J.: Identifying and modeling aspectual scenarios with theme and
MATA. In: Proceedings of the ACM Symposium on Applied Computing, Switzerland,
pp. 287–291 (20100
36. Marques, G., Araújo, J., Lencastre, M.: Integrating problem frames with aspects. In:
Proceedings of the 23rd Brazilian Symposium on Software Engineering, Fortaleza/CE,
pp. 196–206 (2009)
37. Mehner, K., Monga, M., Taentzer, G.: Interaction analysis in aspect-oriented models. In:
Proceedings of the 14th IEEE International Conference Requirements Engineering,
Minnesota, USA, pp. 69–78 (2006)
38. Kit, L.K., Man, C.K., Baniassad, E.: Isolating and relating concerns in requirements using
latent semantic analysis. ACM SIGPLAN Not. 41(10), 383–396 (2006)
39. Meier, S., Reinhard, T., Stoiber, R., Glinz, M.: Modeling and evolving crosscutting concerns
in ADORA. In: Proceedings of the International Conference on Software Engineering,
Minneapolis, MN (2007)
40. Marques, A., Moreira, A., Araújo, J.: Multi-dimensional composition by objective. In:
Proceedings of the International Conference on Software Engineering, Leipzig, Germany,
pp. 19–25 (2008)
41. Mussbacher, G., Amyot, D.: On modeling interactions of early aspects with goals. In:
Proceedings of the Workshop on Aspect-Oriented Requirements Engineering and
Architecture Design, Charlottesville, VA, USA, pp. 14–19 (2009)
42. Mussbacher, G., Whittle, J., Amyot, D.: Semantic-based interaction detection in
aspect-oriented scenarios. In: Proceedings of the IEEE International Conference on
Requirements Engineering, Georgia, USA, pp. 203–212 (2009)
43. Hamza, H.S., Darwish, D.: On the discovery of candidate aspects in software requirements.
In: Proceedings of the 6th International Conference on Information Technology:
New Generations, Las Vegas, USA, pp. 819–824 (2009)
44. Pinto, M., Fuentes, L., Valenzuela, J.A., Pires, P.F., Delicato, F.C.: Promoting the software
evolution in AOSD with early aspects: Architecture-oriented model-based pointcuts. In:
Proceedings of the Workshop on Aspect-Oriented Requirements Engineering and
Architecture Design, Charlottesville, VA, USA, pp. 31–37 (2009)
45. Chernak, Y.: Requirements composition table explained. In: Proceedings of the 20th IEEE
International Requirements Engineering Conference, Chicago, Illinois, USA, pp. 273–278
(2012)
46. de Paula, V., Batista, T.: Revisiting a formal framework for modeling aspects in the design
phase. In: Proceedings of the International Conference on Software Engineering,
Minneapolis, MN (2007)
47. Oliveira, A.R., Araújo, J., Amaral, V.: The VisualAORE DSL. In: Proceedings of the 5th
International Workshop on Requirements Engineering Visualization, Sydney, Australia,
pp. 11–19 (2010)
48. Alencar, F., et al.: Towards modular i* models. In: ACM Symposium on Applied
Computing, pp. 292–297 (2010)
49. Pinto, M., Gamez, N., Fuentes, L.: Towards the architectural definition of the health watcher
system with AO-ADL. In: Proceedings of the Workshop in Aspect-Oriented Requirements
Engineering and Architecture Design, Minneapolis (2007)
264 P.A.P. Júnior and R.D. Penteado

50. Mussbacher, G., Kienzle, J., Amyot, D.: Transformation of aspect-oriented requirements
specifications for reactive systems into aspect-oriented design specifications. In: Proceedings
of the Model-Driven Requirements Engineering Workshop, Trento, Italy, pp. 39–47 (2011)
51. Mussbacher, G., Amyot, D., Weiss, M.: Visualizing aspect-oriented requirements scenarios
with use case maps. In: Proceedings of the First International Workshop on Visualization in
Requirements Engineering (2007)
52. Ossher, H., et al.: C. Using tagging to identify and organize concerns during
pre-requirements analysis. In: Proceedings of the Aspect-Oriented Requirements
Engineering and Architecture Design, Charlottesville, VA, USA, pp. 25–30 (2009)
53. Brito, I., Moreira, A.: Towards a composition process for aspect-oriented requirements. In:
Proceedings of the Early Aspects Workshop at AOSD, Massachusetts, USA (2003)
54. Mussbacher, G., Amyot, D., Araújo, J., Moreira, A.: requirements modeling with the
aspect-oriented user requirements notation (AoURN): a case study. In: Katz, S., Mezini, M.,
Kienzle, J. (eds.) Transactions on Aspect-Oriented Software Development VII. LNCS, vol.
6210, pp. 23–68. Springer, Heidelberg (2010)
55. Kienzle, J., Abed, W.A., Klein, J.: Aspect-oriented multi-view modeling. In: Proceedings of
the 8th International Conference on AOSD, New York, USA, pp. 87–98 (2009)
56. Dyba, T., Dingsoyr, T., Hanssen, G.K.: Applying systematic reviews to diverse study types.
In: Proceedings of the International Symposium on Empirical Software Engineering and
Measurement, Washington, DC, USA (2007)
57. Chung, L., Leite, J.S.P.: Non-Functional Requirements in Software Engineering, p. 441.
Springer, Heidelberg (2000)
58. Parreira Júnior, P.A., Penteado, R.A.D.: Criteria for comparison of aspect-oriented
requirements engineering approaches. In: Brazilian Symposium on Software Engineering
Brasília/DF, Brazil, 2013 (in Portuguese)
 SpecQua: Towards a Framework for Requirements
Specifications with Increased Quality

Alberto Rodrigues da Silva ✉

( )

Instituto Superior Técnico, Universidade de Lisboa and INESC-ID, Lisbon, Portugal

[email protected]

Abstract. Requirements specifications describe multiple technical concerns of

a system and are used throughout the project life-cycle to help sharing a system’s
common understanding among multiple stakeholders. The interest to support the
definition and the management of system requirements specifications (SRSs) is
evident by the diversity of many generic and RE-specific tools. However, little
work has been done in what concerns the quality of SRSs. Indeed, most recom‐
mended practices are mainly focused on human-intensive tasks, mainly
dependent on domain experts, and so, these practices tend to be time-consuming,
error-prone and unproductive. This paper proposes and discusses an innovative
approach to mitigate this status, and defends that with proper tool support – such
as the SpecQua framework discussed in the paper –, we can increase the overall
quality of SRSs as well as we can increase the productivity associated to tradi‐
tional tasks of RE such as documentation and validation.

Keywords: Requirements specification · Quality of requirements specification ·

Requirements validation

1 Introduction

Requirements Engineering (RE) intends to provide a shared vision and understanding of the
system to be developed between business and technical stakeholders [1–3]. The adverse
consequences of disregarding the importance of the early activities covered by RE are
well-known [4, 5]. System requirements specification, software requirements specification
or just requirements specifications (SRS) is a document that describes multiple technical
concerns of a software system [1–3]. An SRS is used throughout different stages of the
project life-cycle to help sharing the system vision among the main stakeholders, as well
as to facilitate communication and the overall project management and system develop‐
ment processes. For achieving an effective communication, everyone should be able to
communicate by means of a common language, and natural language provides the foun‐
dations for such language. Natural language is flexible, universal, and humans are profi‐
cient at using it to communicate with each other. Natural language has minimal adoption
resistance as a requirements documentation technique [1, 3]. However, although natural
language is the most common and preferred form of requirements representation [6], it also
exhibits some intrinsic characteristics that often present themselves as the root cause of
many requirements quality problems, such as incorrectness, inconsistency, incompleteness

© Springer International Publishing Switzerland 2015

J. Cordeiro et al. (Eds.): ICEIS 2014, LNBIP 227, pp. 265–281, 2015.
DOI: 10.1007/978-3-319-22348-3_15
266 A.R. da Silva

and ambiguousness [1, 3]. From these causes, in this paper we emphasize inconsistency and
incompleteness because avoiding – or at least mitigating – them requires significant human
effort due to the large amount of information to process when combined with inadequate
tool support, namely to perform the typical requirements linguistic analysis. On the other
hand, although ambiguity and incorrectness – by definition – cannot be fixed without
human validation [7], we consider that the tasks required to minimize the effects of both
inconsistency and incompleteness (and also ambiguity at some extent) can be automated
if requirements are expressed in a suitable language, and if adequate tool support is provided.
In our recent research we consider the RSLingo approach [9–11] as a starting
point for this challenge. RSLingo is an approach for the formal specification of soft‐
ware requirements that uses lightweight Natural Language Processing (NLP) tech‐
niques [8] to translate informal requirements – originally stated in unconstrained
natural language by business stakeholders – into a formal representation provided
by a language specifically designed for RE. Unlike other RE approaches, which use
languages that typically pose some difficulties to business stakeholders (namely
graphical modeling languages such as UML or SysML, whose target audience are
engineers), RSLingo encourages business stakeholders to actively contribute to the
RE process in a collaborative manner by directly authoring requirements in natural
language. To achieve this goal, RSLingo provides (1) the RSL-PL language for
defining linguistic patterns that frequently occur in requirements specifications
written in natural language, (2) the RSL-RSL language that covers most RE
concerns, in order to enable the formal specification of requirements, and (3) an
information extraction mechanism that, based on the linguistic patterns defined in
RSL-PL, translates the captured information into a formal requirements specifica‐
tion encoded in RSL-IL [10]. However, RSLingo does not provide yet any guaran‐
tees that the RSL-IL specifications have the required quality. Therefore, the main
contribute of this research is to propose and discuss that, with proper tool support
and an intermediate representation language such as RSL-IL, we can increase the
overall quality of SRSs as well as the productivity associated to documentation and
validation tasks. To the best of our knowledge, and apart from our own research on
this issue [12], no further works have been proposed before in relation to this
complex approach and the way we support automatic validation of SRSs as well as
we have discussed the subject of modularity, dependencies and combinatorial effects
at requirements level [13].
The structure of this paper is as follows. Section 2 introduces the background under‐
lying this research. Section 3 overviews the SpecQua framework. Section 4 introduces
some more technical details of the SpecQua with the purpose to show and discuss the
practicability of the proposed approach. Section 5 discusses some experiments that are
being implemented in the context of the SpecQua, in particular related to consistency,
completeness and unambiguousness. Finally, Sect. 6 presents the conclusion and ideas
for future work.
 SpecQua: Towards a Framework for Requirements Specifications 267

2 Background

This section briefly introduces the definition for SRS’s quality attributes, overviews
requirements specification languages, introduces some considerations on requirements
validation, and briefly introduces the RSLingo approach.

2.1 SRS’s Quality Attributes

Writing good requirements is a human-intensive and error prone task. Hooks summarize
the most common problems in that respect [14]: making bad assumptions, writing
implementation (How) instead of requirements (What), describing operations instead of
writing requirements, using incorrect terms, using incorrect sentence structure or bad
grammar, missing requirements, and over-specifying. To achieve quality SRSs must
embody several characteristics. For example, the popular “IEEE Recommended Practice
for Software Requirements Specifications” states that a good-quality SRS should be [7]:
correct, unambiguous, complete, consistent, prioritized, verifiable, modifiable, and
traceable. From all of them, we briefly discuss those that are most relevant for the scope
of this paper.

Complete. A SRS is considered complete if it fulfills the following conditions: (1)

Everything that the system is supposed to do is included in the SRS; this can lead us to
a never ending cycle of requirements gathering; (2) Syntatic structures filled, e.g.: all
pages numbered; all figures and tables numbered, named, and referenced; all terms
defined; all units of measure provided; and all referenced material present; and (3) No
sections or items marked with “To Be Determined” (TBD) or equivalent sentences.
Completeness is probably the most difficult quality attribute to guarantee. In spite that
some elements are easy to detect and correct (e.g., empty sections, TBD references), but
one never knows when the actual requirements are enough to fully describe the system
under consideration. To achieve completeness, reviews of the SRS by customer or users
are essential. Prototypes also help raise awareness of new requirements and help us better
understand poorly or abstractly defined requirements.

Consistent. A SRS is consistent if no requirements described in it conflict among

themselves. Disagreements among requirements must be resolved before development
can proceed. One may not know which (if any) is consistent until some research is done.
When modifying the requirements, inconsistencies can slip in undetected if only the
specific change is reviewed and not any related requirements.

Unambiguous. A SRS is unambiguous if every requirement stated there has only one
possible interpretation. The SRS should be unambiguous both to those who create it and
to those who use it. However, these groups of users often do not have the same back‐
ground and therefore do not tend to describe software requirements the same way.
Ambiguity is a very complex phenomenon because natural language is inherently
ambiguous (a simple word can have multiple meanings) and most of the times this
ambiguity is unintentionally introduced. The most recommended solution to minimize
268 A.R. da Silva

ambiguity is the use of formal or semi-formal specification languages rather than or in

complement to natural languages. Also, the use of checklists and scenario-based reading
are common recommendations [15].

2.2 Requirements Specification Languages

Traditionally, the requirements documentation activity has consisted in creating a
natural language description of the application domain, as well as a prescription of what
the system should do and constraints on its behavior [16]. However, this form of spec‐
ification is both ambiguous and, in many cases, hard to verify because of the lack of a
standard computer-processable representation [17].
Apparently, the usage of formal methods could overcome these problems. However,
this would only address part of the problem, as we still need to take care while inter‐
preting the natural language requirements to create a formal specification, given that in
general engineers often misinterpret natural language specifications during the design
phase. The same occurs with the attempt to directly create formal requirements speci‐
fications, especially when the real requirements are not discovered and validated at first
by the business stakeholders [18]. Thus, the usage of such formal languages entails an
additional misinterpretation level due to the typically complex syntax and mathematical
background of formal method languages [17]. Given that formal methods are expensive
to apply – because they require specialized training and are time-consuming [2] –,
creating formal requirements specifications might have a negative impact.
In the attempt of getting the best from both worlds – the familiarity of natural language
and the rigorousness of formal language –, one can document requirements with controlled
natural languages, which are languages engineered to resemble natural language.
However, these languages are only able to play the role of natural language to a certain
extend: while they are easy to read, they are hard to write without specialized tools [19, 20].
Finally, there are graphical approaches, such as UML and SysML for traditional RE
modeling, and i*, KAOS and Tropos notation for Goal-Oriented RE [1]. However, these
graphical languages are less expressive than natural language and cannot be regarded
as a common language to communicate requirements, because business stakeholders
still require training to understand them. Also, despite being “easier to understand” than
formal method languages, these graphical modeling languages are regarded as less
powerful in terms of analytical capabilities because they often lack tool support to
enforce the implicit semantics of their modeling elements, or might even intentionally
leave some unspecified parts of the language itself to ease its implementation by tool
vendors, in which case they are considered as semi-formal. Some authors even argue
that the simplicity of these languages comes precisely from this lack of semantic
enforcement: it is easy to create models because “anything goes” [5].
Furthermore, the usage of graphical languages might cause another problem when
the modeler includes too much detail in the diagram, cluttering it and thus affecting its
readability. Therefore, despite the existence of such graphical approaches, textual
natural language specifications are still regarded by many as the most suitable, fast, and
preferred manner to initiate the requirements development process of the envisioned
software system.
 SpecQua: Towards a Framework for Requirements Specifications 269

2.3 Tools and Interoperability Formats

Regarding the documentation and management of requirements there are two main
approaches depending on the respective tool support. On one hand, the approach based
on generic tools (usually office applications and suites, word processors, or just text
editors) such as MS-Word, MS-Excel, OpenOffice or GoogleDocs. This is the most
popular approach and many times the solution for simple projects: these tools are ubiq‐
uitous and everyone knows how to use them; they are flexible and good enough for most
documentation needs. However, these tools do not provide specific features, particularly
in what concerns the validation and management of requirements. On the other hand,
RE-specific tools – such as Doors, RequisitePro, ProR, CaliberRM or Visure Require‐
ments –, allow to define each requirement as a set of attributes, can establish link
requirements with each other, and offer analysis and reporting features. Additionally,
some tools allow for concurrent editing and change tracking, and to establish links
between textual requirements and visual models represented in languages such as UML,
SysML or BPMN. Depending on these approaches there are different formats to support
requirements interoperability between tools. The common formats provided by generic
tools are TXT (plain text) and RTF (rich text format) while RE-specific tools provide
XML-based formats such as XMI (usual in UML CASE tools) and ReqIF (Requirements
Interchange Format).

2.4 Requirements Validation

There is not a consensus in the literature about the use of the terms “verification” and
“validation” in the context of RE. However, in this paper we adopt the term as suggested
by Pohl and Robertsons, who define requirements validation as checking requirements
with the purpose of detecting errors such as inconsistencies, ambiguities, and ignored
standards. These authors recommend the use of the term “verification” to denote the
formal (mathematical) proof of properties of a model, related to properties of concurrent
systems, such as safety or absence of deadlocks [1, 2]. Considering the premises
regarding the current practices of the requirements documentation and validation activ‐
ities – such as inspections, walkthroughs, checklists, or using scenarios and prototypes
[1–3] –, we consider that the quality of a SRS still strongly depends on the expertise of
whoever is performing this activity. Given that most RE activities are still manually
performed and involve a large amount of information, to produce a high quality require‐
ments specification one requires an experienced requirements engineer with a vast skills
set. However, to avoid large discrepancies in the results of the RE process, we advocate
that the quality of requirements specifications and the productivity of the requirements
documentation activity can be increased through the formalization of requirements. The
computer-processable requirements specifications that are obtained through such a
formalization process enable the automation of some manual validations thus relieving
requirements engineers from the burden of manually handling a large amount of infor‐
mation to identify requirements quality problems. Additionally, the degree of formali‐
zation achieved can be employed to generate complementary artefacts to better support
RE tasks, such as requirements validation.
270 A.R. da Silva

2.5 The RSLingo Approach

RSLingo is an approach for the formal specification of software requirements that uses
lightweight Natural Language Processing (NLP) techniques to (partially) translate
informal requirements – originally stated by business stakeholders in unconstrained
natural language – into a formal representation provided by a language specifically
designed for RE [9].
The name RSLingo stems from the paronomasia on “RSL” and “Lingo”. On one hand,
“RSL” (Requirements Specification Language) emphasizes the purpose of formally spec‐
ifying requirements. The language that serves this purpose is RSL-IL, in which “IL” stands
for Intermediate Language [10]. On the other hand, “Lingo” expresses that its design has
roots in natural language, which are encoded in linguistic patterns used during by the infor‐
mation extraction process [8, 21] that automates the linguistic analysis of SRSs written in
natural language. The language designed for encoding these RE-specific linguistic patterns
is RSL-PL, in which “PL” stands for Pattern Language [11]. These linguistic patterns are
used by lightweight NLP techniques and, when combined with general-purpose linguistic
resources (e.g., VerbNet (http://verbs.colorado.edu/~mpalmer/projects/verbnet.html) and
WordNet, (http://wordnet.princeton.edu)), enable the extraction of relevant information
from the textual representations of requirements. Finally, the extracted information with
these lightweight NLP techniques is formally specified in RSL-IL notation through prede‐
fined transformations from RSL-PL into RSL-IL. Upon a match of a requirement’s textual
representation with one of the RSL-PL linguistic patterns, a transformation should become
active. This transformation takes into consideration the semantic roles of each word within
the linguistic pattern, and drives the mapping between RSL-PL and RSL-IL.
RSL-IL provides several constructs that are logically arranged into viewpoints
according to the specific RE concerns they address [10]. These viewpoints are organized
according to two abstraction levels: business and system levels.
To properly understand and document the business context of the system, the busi‐
ness level of the RSL-IL supports the following business-related concerns, namely: (1)
the concepts that belong to the business jargon; (2) the people and organizations that
can influence or will be affected by the system; and (3) the objectives of business stake‐
holders regarding the value that the system will bring. Considering these concerns,
business level requirements comprise respectively the following viewpoints: Termi‐
nology, Stakeholders, and Goals.
On the other hand, at the system level, the RSL-IL supports the specification of both
static and dynamic concerns regarding the system, namely: (1) the logical decomposition
of a complex system into several system elements, each with their own capabilities and
characteristics, thus providing a suitable approach to organize and allocate requirements;
(2) the requirements that express the desired features of the system, and also the
constraints and quality attributes; (3) the data structures aligned with the business jargon,
their relations, and a logical description of their attributes; and (4) the actors, functions,
event-based state transitions, and use cases that further detail the aforementioned
requirements. Considering these concerns, the System Level comprises the following
viewpoints: Architectural; Requirements; Structural; and Behavioral, respectively.
 SpecQua: Towards a Framework for Requirements Specifications 271

3 SpecQua Framework – General Overview

Figure 1 illustrates the context diagram of the SpecQua framework and in particular
shows the actors that might interact with it, namely: users (such as requirements engi‐
neers and business analysts), and tool superusers and administrators; and third-party
tools that directly or indirectly interact with SpecQua. SpecQua provides several import
and export features to guarantee the maximum interoperability with third-party tools,
namely with commercial or open-source RE-specific tools and also general-purpose
text-based authoring tools.

uses
:SpecQua User :Third-Party RE Tool
:SpecQua ReqIF (XML) exports
manages «use»

RTF and PDF exports

:SpecQua :Third-Party Word
«use»
SuperUser Processor
(advanced) manages

:SpecQua
Administrator

Fig. 1. SpecQua context diagram.

Figure 2 shows the common usage workflow supported by the SpecQua framework.
This starts (1) by importing a set of requirements specifications produced in the context
of third-party tools, namely by importing from one of the following formats: RSL-IL
plain text, NL (natural language) plain text, or ReqIF. Then, users might (2) author the
requirements directly in the SpecQua tool and (3) run a selected number of tests to
automatically validate the quality of the involved requirements. Finally, if the require‐
ments have the intended quality, (4) the requirements can be exported to third-party tools
in formats such as ReqIF, RTF or PDF, and this process can iterate as many cycles as
needed.
The key concepts underlying the SpecQua tool are depicted in Fig. 3 in a simplified
way. There are two levels of abstraction in this domain model: project-level and generic-
level. The concepts defined at generic-level are managed by users with most responsi‐
bility (i.e., by the SpecQua SuperUsers), and involve the definition of GenericSpecifi‐
cations and Tests. GenericSpecification consists in a set of requirements (following the
ReqIF terminology) that are defined in a project-independent way, so that they can be
reused throughout different projects with minor changes. Test is a type of automatic
validation that can be defined and run against one or more requirement specifications
(see below for more information about Tests).
On the other hand, at project-level, there are common concepts to support the
management of different projects, assign them to different users, with different permis‐
sions, and so on. At this level, any user can create a project, invite other users to partic‐
ipate in that project, and then start to import from third-party tools or author requirement
272 A.R. da Silva

:SpecQua :Third-Party RE
:Third-Party RSL-IL (TXT) imports ReqIF (XML) exports
Tool
Tool «use» «use»
:Other User
NL (TXT) imports
«use»

ReqIF (XML) imports :Third-Party Word

«use» RTF and PDF exports Processor
«use»

uses

:SpecQua User

Fig. 2. SpecQua context diagram (with the common usage workflow).

+owners
User Proj ect
1..* *

+members

* *

* *

Specification TestConfiguration

* *
*

refers* *

reuses

Generic Level
0..1

GenericSpecification Test

«enumeration» «enumeration»
SpecificationCategory TestConfigurationByDefault
TestCategory
functional consistency
quality
completeness
business-level unambiguousness
other other

Fig. 3. SpecQua domain model.

Specifications directly. Users can also establish relationships between specifications and
reuse (by copy and customization) GenericSpecifications previously defined. Users may
select and configure the relevant tests (via TestConfiguration) in order to automatically
validate their specifications. If these specifications have the required quality they can be
exported to third-party tools or some kind of SRS report can be produced automatically.
 SpecQua: Towards a Framework for Requirements Specifications 273

Figure 4 suggests the general operation of the validation process with its main input,
outputs, and supporting resources. The major input is the Specifications files that
are the SRS defined in the RSL-IL concrete syntax (or the alternative formats, such as
NL or structured ReqIF). The validation outputs are the Parsing Log file and the
TestRun Reports file with the errors and warnings detected by the tool during its
execution, respectively during the Parsing and the Validation processes. Addi‐
tionally, there are some supporting resources used to better extend and support the tool
at run-time, namely: Tests, TestConfiguration, and Lexical Resources.

Fig. 4. Overview of the validation process.

Tests are directly implemented in a given programming language and have addi‐
tional metadata such as name, description and quality category. (Figure 9 gives an
example of such test directly implemented in C#.) TestConfiguration is a
resource used to support the validation in a flexible way. For example, this resource
allow requirements engineers to configure the level of completeness needed for their
purpose, in a project basis. This means that different projects may have different needs
regarding completeness for their specifications. Finally, Lexical Resources are
public domain resources (such as WordNet or VerbNet) that support some tests, mostly
those related with linguistic issues.

4 SpecQua Framework – More Details

The proposed high-level approach has to be implemented by a concrete software tool to

offer a real interest and utility. Due to that we have implemented the SpecQua Frame‐
work with the purpose to show and discuss the practicability of this approach.
274 A.R. da Silva

The SpecQua has the following objectives. First, provide SRS’s quality reports: for a
given RSL-IL specification, the system should be able to provide results for quality tests
applied to that specification. Second, easily add and configure new quality tests: it should
be easy to develop and add new quality tests in order to extend the tool; additionally it
should be easy to configure those tests. Third, provide a Web collaborative workspace: the
tool should be accessible via a Web browser and should provide the means for multiple
users to work together while specifying, analyzing and validating RSL-IL specifications.
(A preliminary prototype is available at http://specqua.apphb.com).
Figure 5 depicts the SpecQua software architecture in generic terms, with its main
blocks: WebApp, WebAPI and Core.

Fig. 5. SpecQua architecture.

SpecQua WebApp. The WebApp layer corresponds to how users interact with the
tool. In a 3-tier layer architecture this corresponds to the presentation layer, although
this is more than a simple frontend: the WebApp is a Web-based independent application
with its own logic that connects to SpecQua Core via its API.

SpecQua WebAPI. This intermediate layer serves as the Application Programming

Interface (API) that exposes all the relevant functions that Core provides to the outside.
Besides serving as a proxy between WebApp and Core, the API can still be accessed
from other clients that do not intend to use the main frontend but still want to take
advantage of SpecQua analyses or data resources.

SpecQua Core. This layer is the kernel of the SpecQua tool and this is where all the
action takes place. A key feature for the system SpecQua is the ease to add new tests
and this is implemented using two advanced programming techniques (i.e., dependency
injection and reflection) which combined make it possible to associate new quality tests
to the tool with minimal effort. At system startup, and based on referenced assemblies,
all classes that implement a specific interface become available to the system as new
 SpecQua: Towards a Framework for Requirements Specifications 275

quality tests, and are logically grouped into analyses. Currently, this grouping of testing
analysis is performed using a configuration file as exemplified in Fig. 6.

Fig. 6. Example of a SpecQuA’s test configuration file.

Each test is independent from each other and may have its own particular configu‐
ration. This configuration is defined in XML format and has no restriction on the schema
level: it is up for those who develop the test (i.e., the SpecQua superuser), to define the
schema and interpret it in the respective test. Each test may or may not have a default
configuration.
Additionally, the SpecQua Core has the Parser component which is responsible to
parse the input RSL-IL text and map it to the Domain Model (in the internal database).
This parsing is done by ANTLR (http://www.antlr.org/) using a grammar specific for
the RSL-IL language. With this grammar, ANTLR generates a parser and a lexer that
can be used to validate the syntax and the semantic of a given input. If the input does
not have errors, then the result is a representation of the domain model entities that can
be tested. This grammar has some similarities with the BNF notation as expressed in
Fig. 7. (Other parsers to different import formats might be developed in the future.)

5 Discussion

Despite having a simple user-interface, a lot is done in the background when carrying
out the analysis of a specification. In this case, when the user wants to validate the
276 A.R. da Silva

Fig. 7. An excerpt from the RSL-IL grammar for ANTLR.

Fig. 8. Excerpt of a report test for the AIDSPortal example.

specification, it is parsed with the specific ANTLR grammar. If any syntactic or semantic
error is detected in the specification, the user is alerted and the process stops. However,
if the specification is successfully parsed, the tests can be selected and configured and
are run against the specification. Finally, a report is shown to the user such as the example
shown in Fig. 8. In the following subsections we introduce and discuss some tests that
are being implemented in the context of the SpecQua framework.

5.1 Consistency Validation

The consistency validation enforces that the information model underlying the RSL-IL
specification is well-formed, or consistent in accordance with the RSL-IL metamodel,
which involves, for example, the following concrete validations.

Consistent Attribute Values. SpecQua verifies whether the value assigned to a given
attribute is valid based on the semantics of its RSL-IL construct. For instance, SpecQua
 SpecQua: Towards a Framework for Requirements Specifications 277

can systematically enforce that every id attribute follows the predefined prefix of each
RSL-IL construct. Also, SpecQua can also provide a short mnemonic for the id attribute
based on the word attribute of the related Term, which is more meaningful than just a
numeric sequence.
The combination of the specific prefix with this mnemonic allows one to better
understand when two RSL-IL constructs of different types refer to the same underlying
concept (e.g., they describe different concerns about the same business notion).

Consistent Numeric Sequences. There are several attributes in RSL-IL constructs that
follow a certain numeric sequence. For instance, SpecQua checks the order attribute of
each Sentence within a given Requirement. Also, SpecQua verifies the values assigned
to the label attributes of a given UseCase’s Steps. In all these cases, SpecQua must
ensure that each construct was assigned a unique numeric value of that sequence, and
that all the assigned numbers follow a monotonic increasing sequence without gaps.

Referential Integrity. SpecQua must check and enforce that those relationships
between different RSL-IL constructs are properly defined in terms of the values (i.e.,
references) assigned to the attributes that support the establishment of such relationships.
The most obvious case is given by the strong dependency of most RSL-IL constructs
on a Term that unambiguously defines the semantics of the underlying concept. Thus,
SpecQua must check whether all RSL-IL constructs that depend on the Terminology
viewpoint effectively provide a valid reference for a previously defined Term through
its id. That is the example of the test shown in Fig. 9 that checks if all Stakeholders are
referenced as Terms defined in the glossary.
Another important aspect of this validation is also to support for the resolution of
Term references based on their acronym or word values, instead of only relying on the
value of its id attribute.
Although we are illustrating this sort of validations mostly based on the Terminology
viewpoint, there are similar cases in other RSL-IL viewpoints. For instance, this problem
is similar to the validation performed regarding the source attribute of a given Goal,
which should be resolved to a well-defined Stakeholder.

5.2 Completeness Validation

The completeness validation is based on the test’s configuration resource that

enables the definition of the level of completeness required for each RSL-IL speci‐
fication. This level of completeness varies on a project basis or even, for the same
project, along the timeline according the needs of the project team. We consider
three levels of completeness:

Completeness at Model Level. At the macro level, one can define which viewpoints
are required for considering a concrete RSL-IL specification to be complete. For
example, during the initial stage of a project lifecycle, one may only require the Termi‐
nology, Stakeholders, and Goals viewpoints to consider the specification as being
complete. On the other hand, if the project is running in a more advanced stage
278 A.R. da Silva

Fig. 9. An excerpt of a concrete Test.

(for instance, after a couple of iterations), the remaining System Level viewpoints should
be also considered in order to provide a complete requirements specification.

Completeness at Viewpoint Level. For each viewpoint one can define which
constructs are mandatory or optional. For example, for the Behavioral viewpoint one
might only consider as being relevant the existence of the Function construct (and not
of Event) in order to consider that viewpoint as being complete.

Completeness at Construct Level. For each construct (e.g., Term, Stakeholder, Goal,
Entity, Use Case) one can define which attributes are mandatory or optional. For
example, for the Goal construct one can define the criticality as a mandatory attribute
and the source (a reference for the Stakeholder responsible for that goal) as an optional
attribute. Still at construct level, we can enforce that the names of some of these
constructs (e.g., the names of actors and entities) should be defined as a unique term in
the Terminology viewpoint.
 SpecQua: Towards a Framework for Requirements Specifications 279

5.3 Unambiguousness Validation

While in a formal specification (such as in RSL-IL) inconsistencies and incompleteness
can be automatically detected, ambiguities deal directly with the meaning of those
specifications, thus they are hard to be detected by automatic processes. Consequently,
ambiguity tends to be detected mostly by human intervention, for example through
analysis and inspection of the specification and through the use of prototypes. However,
regarding this semantic level, still some automatic validation can be applied to reduce
ambiguity, such as those discussed below.

Semantic Analysis. First, based on general-purpose linguistic resources that encode

world knowledge, SpecQua can further verify the semantic validity of relations estab‐
lished between RSL-IL constructs, especially those strongly related with the natural
language representation of concepts. For instance, an advanced validation feature
consists in using WordNet to check whether the value of the word attribute of synonym
Terms are indeed synonyms of the word attribute’s value of the primary Term to which
they are associated. Second, the information encoded within WordNet can be used to
cross-check whether the Term associated with a given Stakeholder (through its role
attribute) is aligned with the classification provided by the StakeholderType enumera‐
tion based on the lexname attribute of the WordNet synset referred by the synset attrib‐
ute’s value of that Term. Third, and still regarding the relations between different Stake‐
holders, SpecQua must verify the semantics of the hierarchical composition of these
RSL-IL constructs. For instance, it does not make sense to specify that a Stakeholder
whose type is “group.organizational” is MemberOf of another Stakeholder whose type
is “individual.person”. This means that the hierarchical Stakeholders composition must
follow the implicitly semantics entailed in the values of the StakeholderType enumer‐
ation, which are ordered from broader groups to more specific entities. Fourth, SpecQua
can determine whether the relation between a given RSL-IL construct and a Term is
semantically valid based on the pos attribute of that Term and the semantics of the other
RSL-IL construct. For instance, it does not make sense to associate an Entity with a
Term whose part-of-speech (provided by either its pos or synset attributes) classifies the
Term as a verb, instead of a noun. Fifth, another example consists in checking whether
nouns associated with the agent thematic relation (e.g., the subject of natural language
sentences in the active voice) are defined as Actors and, if so, whether they can be traced
back to the respective Stakeholders via a shared Term.

Terminology Normalization. The RSL-IL glossary (i.e., its Terminology viewpoint)

formally defined the terms associated with the main concepts used throughout the
requirements specification. There are different types of relations that can be established
between terms, i.e. relations of type synonym, antonym, and hyponym. One motivation
for using these relations is to reduce the number of redundant Terms employed within
the RSL-IL specification, by providing a unique Term for each concept. So, it is impor‐
tant to avoid the definition of two or more synonym Terms by clearly stating which one
of them should be classified as the primary Term, and the other(s) as secondary Term(s).
Based on this information, SpecQua can perform a systematic normalization of Terms
280 A.R. da Silva

through a common find and replace process and, consequently, reduce the requirements
specification’s ambiguity.

6 Conclusions

RE comprises several tasks including requirements elicitation, analysis and negotiation,

documentation and validation. We recognize that natural language is the most common
and popular form to document SRSs. However, natural language exhibits several limi‐
tations, in particular those related with requirements specification quality such as incor‐
rectness, inconsistency, incompleteness and ambiguousness.
This research extends the RSLingo approach by considering that requirements are
represented in RSL-IL, automatically extracted from natural language specifications or
authored directly by their users. This paper proposes a generic approach to automatically
validate these specifications and describes the SpecQua framework that shows the prac‐
ticability and utility of this proposal. The flexibility of SpecQua and the initial cases
studies developed so far allows us to preliminary conclude that this approach helps to
mitigate some of the mentioned limitations, in particular in what respect inconsistency,
incompleteness and ambiguousness.
For future work we plan to implement SpecQua in a more extensive way, in particular
with features related with the support of a collaborative environment, allowing end-users
to author and validate directly their requirements [22], eventually with different repre‐
sentations beyond natural language, RSL-IL and ReqIF. Several import and export
features would also be relevant in order to promote tools interoperability as discussed
in the paper. Finally, we still intend to explore the integration of requirements specifi‐
cations with testing [23] and model-driven engineering approaches [24–26] to still
increase the quality and productivity of Software Engineering in general terms.

Acknowledgements. This work was partially supported by national funds through FCT –
Fundação para a Ciência e a Tecnologia, under the project PEst-OE/EEI/LA0021/2013 and
DataStorm Research Line of Excellency funding (EXCL/EEI-ESS/0257/2012). Particular thanks
to my students David Ferreira and Joao Marques for their strong participation and involvement
in this research.

References

1. Pohl, K.: Requirements Engineering Fundamentals, Principles, and Techniques, 1st edn.
Springer, Heidelberg (2010)
2. Sommerville, I., Sawyer, P.: Requirements Engineering: A Good Practice Guide. Wiley, New
York (1997)
3. Robertson, S., Robertson, J.: Mastering the Requirements Process, 2nd edn. Addison-Wesley,
Boston (2006)
4. Emam, K., Koru, A.: A replicated survey of IT software project failures. IEEE Softw. 25(5),
84–90 (2008)
5. Davis, A.M.: Just Enough Requirements Management: Where Software Development Meets
Marketing, 1st edn. Dorset House Publishing, New York (2005)
 SpecQua: Towards a Framework for Requirements Specifications 281

6. Kovitz, B.: Practical Software Requirements: Manual of Content and Style. Manning,
Greenwich (1998)
7. IEEE Computer Society, 1998. IEEE Recommended Practice for Software Requirements
Specifications. IEEE Std 830-1998
8. Bird, S., Klein, E., Loper, E.: Natural Language Processing with Python. O’Reilly Media,
Sebastapol (2009)
9. Ferreira, D., Silva, A.R.: RSLingo: an information extraction approach toward formal
requirements specifications. In: Proceedings of the 2nd Int. Workshop on Model-Driven
Requirements Engineering (MoDRE 2012), IEEE CS (2012)
10. Ferreira, D., Silva, A.R.: RSL-IL: An interlingua for formally documenting requirements. In:
Proceedings of the of Third IEEE International Workshop on Model-Driven Requirements
Engineering (MoDRE 2013), IEEE CS (2013)
11. Ferreira, D., Silva, A.R.: RSL-PL: A linguistic pattern language for documenting software
requirements. In: Proceedings of the of Third International Workshop on Requirements
Patterns (RePa 2013), IEEE CS (2013)
12. Silva, A.R.: Quality of requirements specifications: a framework for automatic validation of
requirements In: Proceedings of ICEIS 2014 Conference, 2014, SCITEPRESS (2014)
13. Silva, A.R., et al.: Towards a system requirements specification template that minimizes
combinatorial effects. In: Proceedings of QUATIC 2014 Conference, IEEE CS (2014)
14. Hooks I.: Writing good requirements. In: Proceedings of the Third International Symposium
of the INCOSE, vol. 2 (1993)
15. Kamsties, E., Berry, D.M., Paech, B.: Detecting ambiguities in requirements documents using
inspections. In: Proceedings of the First Workshop on Inspection in Software Engineering
(2001)
16. van Lamsweerde, A.: From Worlds to Machines. In: Nuseibeh, B., Zave, P. (eds.) A Tribute
to Michael Jackson. Lulu Press (2009)
17. Foster, H., Krolnik, A., Lacey, D.: Assertion-based Design. Springer, Heidelberg (2004)
18. Young, R., 2003. The Requirements Engineering Handbook. Artech Print on Demand
19. Fuchs, N.E., Kaljurand, K., Kuhn, T.: Attempto controlled english for knowledge
representation. In: Baroglio, C., Bonatti, P.A., Małuszyński, J., Marchiori, M., Polleres, A.,
Schaffert, S. (eds.) Reasoning Web. LNCS, vol. 5224, pp. 104–124. Springer, Heidelberg
(2008)
20. Kuhn, T.: Controlled English for Knowledge Representation. Ph.D. thesis, Faculty of
Economics, Business Administration and Information Technology of the University of Zurich
(2010)
21. Cunningham, H.: Information extraction, automatic. Encyclopedia of Language and
Linguistics, 2nd edn. Elsevier, Amsterdam (2006)
22. Ferreira, D., Silva, A.R.: Wiki supported collaborative requirements engineering. In:
Proceedings of the 4th International Symposium on Wikis. ACM (2008)
23. Moreira, R., Paiva, A.C.R., Memon, A.: A pattern-based approach for GUI modeling and
testing. In: IEEE 24th International Symposium on Software Reliability Engineering
(ISSRE), IEEE CS (2013)
24. Silva, A.R., et al.: Integration of RE and MDE paradigms: the ProjectIT approach and tools.
IET Softw. J. 1(6), 294–314 (2007)
25. Savic, D., et al.: Use case specification at different levels of abstraction. In: Proceedings of
QUATIC 2012 Conference, 2012, IEEE CS (2012)
26. Ribeiro, A., Silva, A.R.: XIS-Mobile: a DSL for mobile applications. In: Proceedings of SAC
2014 Conference, ACM (2014)
Software Agents and Internet
Computing
 Analyzing Virtual Organizations’ Formation
Risk in P2P SON Environments

Rafael Giordano Vieira, Omir Correia Alves Junior, and Adriano Fiorese(B)

Department of Computer Science, Santa Catarina State University, Joinville, Brazil

{rafaelgiordano12,omalves,adriano.fiorese}@gmail.com

Abstract. Virtual Organizations (VOs) have emerged as a powerful

enterprising strategy to leverage the activities of Small and Medium
sized Enterprises, by providing a way of sharing their costs and bene-
fits when attending to particular demands. Along with their numerous
advantages, VOs also pose several challenges, including additional risks
that need to be considered to ensure its correct operation. Proper risk
analysis provides more solid means for managers to evaluate and further
decide about the most suitable VO composition for a given business.
Therefore, this paper aims to develop a mathematical method that ana-
lyzes and measures the risk in a set of partners, synthesized in the form of
Service Providers, which will compose a VO in a P2P SON environment.
By means of validation of the proposed method, a set of simulations were
carried out and the results are presented in this paper.

Keywords: Virtual organization · Risk analysis · P2P SON

1 Introduction
The fundamental developments in network technologies, particularly the advent
of Peer-to-Peer Service Overlay Networks (P2P SON) [1,2], provided an advan-
tageous environment for companies make their services available to the user
community at large. The joining of the SON and P2P fields offers a high poten-
tial for handling services, by creating dynamic and adaptive value chain networks
across multiple Service Providers (SPs). Moreover, a wide range of services can
be made available, as well as an environment where price and quality can be
competitive differentials [3].
The P2P SON concept applies to a broad range of network architectures.
This paper deals particularly with the Virtual Organization (VO) type of net-
work. A VO is a temporary and dynamic strategic alliance of autonomous, het-
erogeneous and usually geographically dispersed companies created to attend
very particular business opportunities [4,5]. In this sense, the P2P SON acts as
infrastructure that provides an environment for VO formation and, additionally,
enhances benefits to SPs, e.g., sharing costs, bandwidth and others [6].
Although the mentioned advantages of using P2P SON can improve the
VO formation process, the natural VO networked structure faces additional
c Springer International Publishing Switzerland 2015
J. Cordeiro et al. (Eds.): ICEIS 2014, LNBIP 227, pp. 285–301, 2015.
DOI: 10.1007/978-3-319-22348-3 16
286 R.G. Vieira et al.

risks than other general forms of organization. These additional risks come,
in part, from the increasing sharing of responsibilities among companies and
their dynamic nature of relationships [7]. The problem is that there is a lack
of more systematic and integrated methods to handle the several dimensions of
risk, which includes both intra-organizational and inter-organizational aspects
of the VO. Since the analysis of these risks is key to ensure the VO’s proper
operation, it should be complemented with the risk analysis support methods
that can provide more agility and transparency when creating new VOs [8].
In a previous work, it was designed a three-layer architecture for services
management in P2P SONs, named OMAN [6]. The OMAN offers an efficient
search and selection process of most suitable SPs in a multi-provider environ-
ment. Authors also presented results of SP selection by using a geographical
location criteria [2]. However, the VO risk aspects in the context of P2P SONs
were not addressed.
This paper presents an exploratory work, which complements the proposals of
[2] and [8], and looks for answering how SPs can be properly selected when con-
sidering risks. This work consists in adding an additional risk management level
in the search and selection process and conceiving a new risk analysis method,
named MARTP (Multi criteria Risk Analysis Method applied to P2P Service
Overlay Networks). In the proposed method, the SPs are two-stage evaluated,
both individually and collectively. The goal of the method is to measure the
level of risk and identify which SPs are most risky for the VO formation. This
will allow decision-makers to decide wisely about which SPs should be effec-
tively discarded for a given business collaboration opportunity, and additionally,
the identified risks can be managed and hence mitigated throughout the VO
formation process.
The remainder of this paper is organized as follows: Sect. 2 addresses the
problem of SPs search and selection in P2P SONs and contextualizes it within
the VO risk analysis proposal. Section 3 describes the proposed method for VO
risk analysis. Section 4 presents the set of experiments conducted to evaluate the
proposed method and also presents the final results. Finally, Sect. 5 concludes
the paper and outlines some future work.

2 General Background
2.1 Service Provider Integration

As cited in Sect. 1, different SPs can be grouped in a given VO in order to accom-

plish a mutual goal, the so-called collaboration opportunity. These SPs might
range from non-governmental organizations to autonomous software entities, by
sharing costs, benefits and risks, acting as they were one single enterprise [5].
Regarding to the classical main phases of a VO life cycle (creation, operation,
evolution and dissolution phases) [9], this paper focuses on the creation (or for-
mation) phase, which is seen in Fig. 1. Within the creation phase, this analysis
is carried out during the Partner’s Search and Selection step (left circle).
Analyzing Virtual Organizations’ Formation Risk in P2P SON Environments 287

Fig. 1. Framework for the VO formation process. extended from [9].

Fig. 2. BPSS model [2].

The process of collaboration among the SPs in a VO is accomplished through

interactions between their business processes, which are usually supported by a
network infrastructure. Particularly, this work addresses the use of P2P SON
to organize all the SPs committed with the eventual VO formation. A P2P
SON is an infrastructure designed to provide services and, in the context of this
work, it can be seen as a breeding environment for the creation of VOs [10]. It
is also considered that the SP’s search and selection procedures are performed
by a service management architecture developed in [6], particularly through its
BPSS module, which aims to select one SP from the set of SPs that fulfill a
required service according a particular metric.
Figure 2 details the BPSS module. P2P SON, shown as the elliptic curve,
is created covering domains (clouds in Fig. 2) that contain SPs. Every peer in
the P2P SON runs service(s) from the corresponding SPs. The AgS is created in
288 R.G. Vieira et al.

a higher level inside the P2P SON, where each AgS peer maintains an aggregation
of services published by the SON peers (providers at the P2P SON level). In
order to select a SP (peer), the BPSS sends a service request to the AgS, which
forwards the request to the peers in the aggregation overlay. In the context of
this work, this means the begin of a new Collaboration Opportunity (CO) that
will trigger the formation of a new VO [9]. The result of this request is a list of all
SPs that fulfill a required service according a particular, or a set of, application
metrics.

2.2 VO Formation Risk Analysis

The problem in choosing the most appropriate SPs to compose a VO is critical.
The concept of risk can be handled at a number of perspectives. [11] provides
an overview of risk definition, as a variation in the distribution of possible out-
comes, their probabilities, and their subjective values. [12] associates risk with
the likelihood of an unfavorable outcome. When applied on this research context,
the risk can then be viewed as a composition of three basic elements: the general
environment where it can happens; its occurrence probability; and the scope of
its impact in the case of its occurrence [13]. Specifically, a risk is characterized
by the potential of an SP – that in principle is able to be a member of a VO
– to do not perform correctly its assigned task regarding the associated CO’s
requirements and hence hazarding the VO success.
In the literature review, a number of risk analysis methods has been iden-
tified as potentially suitable for VOs, namely FMEA (Failure mode and effects
analysis), FTA (Fault Tree Analysis), AHP/ANP (Analytic Hierarchy/Network
Process), ETA (Event Tree Analysis), Bayesian Networks, CNEA (Causal Net-
work Event Analysis) and Ishikawa Diagram [13–17]. Some requirements can
be pointed out for the tackled problem: events can be treated as independent
from each other; the deterministic relation between events can be known; events
analysis can be both qualitative and quantitative; a risk can be globally quanti-
fied after a succession of events. Regarding these requirements, ETA and AHP
techniques were selected to be used and combined in the proposed method.
In that same reasoning line, there is a number of works related VO’s sources
of risk and its analysis. In [7,18], thirteen KPIs were identified as general risk
sources in VOs, further identifying the importance of each one. In [19], the prob-
lem of risk mitigation in VO was discussed, and four processes were identified
to improve the level of VOs performance reliability. In [20] two sources of risks
were specified (external and internal), and risk occurrence likelihood in the life
span of a VO was calculated based on them. [21,22] considered the fuzzy char-
acteristics and the project organization mode of VOs to propose Multi Strategy
Multi Choice (MSMC) risk programming models.
In addition, [23] presented a competence model to support efficiently the
process of partner’s selection, which works in the context of Service-Oriented
Virtual Organization Breeding Environments. When dealing with SP as the type
of partners, the same authors have proposed a hybrid DEA-Fuzzy method for
Analyzing Virtual Organizations’ Formation Risk in P2P SON Environments 289

analyzing the risk in the process of VO partners search and selection [24]. The
method is strongly based in the concept of SPs importance and efficiency within
the VO, where aspects like SPs historical performance and relative efficiency are
considered in the analysis and therefore in the calculation performed.
All the reviewed works in the literature have proposed contributions to iso-
lated elements of the whole tackled problem of this research. Nevertheless, none
have somehow formalized how the proposed KPIs should be used nor provided
means to quantity VO partners risks before the VO formation. Moreover, it
was not identified proposals that specify a method or procedure that aims to
systematize the process of risk qualification/quantification involved in the SPs’
Search and Selection for the VO formation, while focusing in the SP type of
partner. Therefore, this paper presents as a contribution a way to specify KPIs
together with a mathematical method that enable measuring the risk in the VO
formation.
The way the risk is represented should be aligned with the organization goals
so that the most important ones can be determined for further and more proper
management. Identifying risk sources is the first and most important step in risk
management [13]. Therefore, there are four main sources of risks regarding VOs
that were considered the most critical ones: trust, communication, collaboration
and commitment [7]. In this work they are modeled as KPIs and their values are
calculated and provided by the methodology developed in [8]:

– Trust. SPs who are going to compose a VO do not necessarily have prior
knowledge about each other before starting collaborating. Thus, trust is cru-
cial to bear in mind, which in turn involves commitment in doing the planned
tasks. When trust among providers is not enough established there is a hesi-
tation to share risks and so the VO can be jeopardized;
– Communication. Communication among VO’s SPs is a key factor for its
proper operation. They should provide correct information about parts, prod-
ucts and services, collaborating in solving conflicts, sharing practices, etc.
However, this can be complicated by the fact SPs are heterogeneous, inde-
pendent, geographically dispersed and usually have distinct working cultures.
The insufficient communication can put a VO on risk;
– Collaboration. Collaboration is characterized when the sharing of risks,
costs and benefits of doing business are agreed and fairly distributed among
partners. However, when a collaboration agreement is not clearly defined, i.e.,
when there is no clear definition of its main objectives, the VO risk increases;
– Commitment. Commitment is related to the attitude of VO members with
each other, i.e., it considers the contributions and agreements made by and
among them for a business. This is important as partners have complementary
skills and so it is important they feed the whole environment with the right
and timely information. The VO risk gets higher when partners fail in that
attitude.
290 R.G. Vieira et al.

3 The Proposed Method

3.1 MARTP Overview

The devised method for risk analysis is generally presented in Fig. 3. It starts hav-
ing as input a pre-selected and ranked list of most adequate SPs (through BPSS
simulation) registered in a P2P SON environment. The main goal of the proposed
risk analysis method is to add another support dimension for decision-making,
identifying and measuring how risky is each of those SP candidates involved in
the VO formation process. In this work, considering VO reference theoretical
foundations [5], the so-called VO Manager is seen as the main decision-maker.

Fig. 3. Overview of MARTP.

The method splits the problem into two stages. In the first stage, it starts
measuring the risks individually, for each possible SP, and after and based on
that, collectively, for the entire SP team for the given VO. In this context, the
VO manager has the following role: to quantify the level of importance of each
SP in the VOs before creating it. There is also a risk specialist, who is in charge
of auditing the SPs historical KPI metrics. The risk techniques and criteria are
applied to assess the risk according to the VO manager guidelines.

3.2 MARTP Architecture

The MARTP method itself is illustrated in Fig. 4. It divides the problem into
two phases: the first phase does the individual risk analysis applying the Event
Tree Analysis (ETA) method for that. The second phase does the risk analysis
taking the group of SPs as a whole into account, applying the Analytic Hierarchy
Process (AHP) method [14,15].
Individual Risk Analysis. In the first phase of MARTP, it is performed an
individual risk analysis for pre-selected SPs. ETA is particularly suitable for
risk analysis of systems where there are interactions between several types of
Analyzing Virtual Organizations’ Formation Risk in P2P SON Environments 291

Fig. 4. MARTP architecture.

probabilistic events, whether dependent or independent [14]. It uses a visual

representation based on a logical binary tree structure, known as Event Tree
(ET), as shown in Stage 1 of Fig. 4.
An ET is a probability tree, which provides two possible conditions: suc-
cess and failure. It also has three basic components: initial event; intermediary
events; and outcomes. The initial event begins the ET creation process. In this
work, it corresponds to one pre-selected SP, and the assigned probability (PIE )
is always 1 (or 100 %) in the beginning [14]. Next step consists in specifying
the (four) intermediary events, which are represented by the (four) KPIs: trust,
communication, collaboration and commitment.
These events are used to quantify the effectiveness of a particular SP, i.e., if it
is able or not to compose a VO, and to generate an ET by assigning success and
failure probabilities to each of them as shown in Stage 1 of Fig. 4. The criterion
to assign the KPI success probability to each SP takes the historical values
292 R.G. Vieira et al.

analysis of the KPI that were assigned to it in past VOs participations [25,26].
This analysis is fundamentally based on statistical inferences by quantifying both
the central trend and variability of historical values.
The central trend analysis is performed by calculating an exponentially
weighted average index (EWA) for each set of historical KPI values of a given
SP. The EWA is currently used in financial risk analysis and supply chain man-
agement being popular in practice due to its simplicity, computational efficiency
and reasonable accuracy (giving more importance for the most recent values in
an exponential factor) [27]. The EWA for a KPI k of a SP p is formally defined by
Eq. 1:
n
xi wi
X̄k (p) = i=1n (1)
i=1 wi

where x = {x1 , x2 , ..., xn } means a non-empty set of historical KPI values and w
represents a normalized exponential decay constant (note that this paper aims to
calculate a success probability by KPI historical analysis; the determination of
optimal values for central trend analysis is not within its scope). After calculating
the EWA for each SP, the Maximum Quality Index (MQI) value is assigned as
the higher value among all the results obtained with the EWA results of a given
Kk for different SPs (that is, for p = 1, 2, ...n). The MQI is calculated for each
KPI and used as a performance reference for all others SPs that will be assessed.
In this sense, considering k the number of used KPIs (in this case four ) and p
the number of SPs associated for each KPI, Eq. 2 shows the MQI calculation
procedure:
 
M QIk = maxk X̄k (p) ∀p ∈ SP (2)

For instance, Fig. 5 shows a graph with hypothetical KPI values about trust
(intermediate event KP I1 according to Stage 1 of Fig. 4) associated to a SP.
The value of the MQI (left circle in Fig. 5) assigned for this KPI would have
been set up as 6.7 (this value is the highest EWA value calculated for SPs using
the KPI trust). Nevertheless, it is obvious that, when taking into account only
the highest MQI value, a few KPIs will reach an acceptable success probability.
For this reason, a variability metric is well-suited in this scope. The metric used
is the standard deviation (SD) of MQI. Therefore, the acceptable interval will
range not only values above 6.7, but also includes the SD interval, which are 2.4
(right circle in Fig. 5). So, the acceptable range turn to 6.7 − 2.4 = 4.3.
The values assigned to each KPI can vary from 0 to 10 and are associated
with a probability success rate which varies from 0 to 1, respectively. Assuming
that each SP has participated in nP A past VOs and since that nR represents the
number of SP’s previous participation in VOs where its KPIs values are higher
than M QIk − SDk (those values with an * in Fig. 5), Eq. 3 calculates the KPI
success probability for the current participation.
nR
P r (Kk ) = (3)
nP A
Analyzing Virtual Organizations’ Formation Risk in P2P SON Environments 293

10
Considered Standard Deviation interval
8.8(*) Maximum Quality Index (MQI)
8 7.6(*)
7.0(*)
MQI = 6.7

KPI value
6
5.3(*)

4 3.6 SD = 2.4
2.5
2

0
VO1 VO2 VO3 VO4 VO5 ... VOn
Previous participations in VOs

Fig. 5. Trust KPI historical values for a given SP.

 
The failure rate for a given KPI is represented as P r K̄k by the following
equation:  
P r K̄k = 1 − P r (Kk ) (4)

According to Fig. 4, the success and failure probability rates are calculated
for all KPIs that compose the ET of an SP, which are presented by the four inter-
mediate (and independent from each other) events K1:4 that populate the ET.
Event K2 , for instance, would be related to KPI  communication, with success
and failure values of P r(K2 ) = 0.90 and P r K̄2 = 0.10, respectively.
After assigning all probabilities for all ET branches, it is necessary to identify
if the SPs are minimally qualified to compose a VO. For this, a calculation is
performed to obtain the final probabilities for all event combinations composing
the ET. They are determined for each of the 2|K| branches of ET and are got
by multiplying the probabilities of events that compose each path. Following, it
is applied the Harmonic Weighted Average (HWA) calculation over the set of
results obtained in the ET, in order to obtain the level of risk for each SP.
The presented concepts can be formalized as follows:
Let SP = {SP1 , SP2 , ..., SPn } be a set of n SPs previously selected, where
each element in this set is associated with a different type of service activity
that is being requested in a CO. Let K = {K1 , K2 , ..., Km } be a set of m KPIs
associated to a SPn , and P r (Kk ) the probability function associated with each
event Kk (as defined in Eq. 3). ETA events occur independently, i.e., where the
occurrence of an event does not affect the occurrence of the other.
Now consider P = {P1 , P2 , ..., P2|K| } as a set of all possible outcomes from the
2|K| ET events combinations. The procedure for obtaining this set was performed
using a Binary Search Tree (BST) [28], which travels 2|K| different paths and
assigns a value to each element of P , as shown in Eq. 5:
|K|
⎡ ⎤
2 |K|

P = ⎣PIE ∗ ω(i, j, k, l)⎦ (5)
k=1 l=1
294 R.G. Vieira et al.

where PIE is the initial probability of the SP. The function ω, as shown in
Eq. 6, performs a binary search in the tree, returning a path element from each
iteration. Values i and j correspond, respectively, to the beginning and ending of
the search, and have i = 0 and j = 2|K| as initial values. The value k corresponds
to the index of the sought element (an element of P ) and l, the current level of
the tree. The sequence of events can be viewed in Stage 1 of Fig. 4.

P r (Kl ) ; j = c, k ≤ c
ω (i, j, k, l) = (6)
1 − P r (Kl ) ; i = c, k > c

where c = (i + j)/2. After all the possible outputs of P for a SP p are defined,
the method calculates the value Sp , which represents its quantitative risk level,
as formalized in Eq. 7. The procedure for calculating the Sp takes into account
performing a Harmonic Weighted Average (HWA) over all elements of P . The
HWA comprises a particular type of average where the weights follow a sequence
of harmonic numbers (i.e., the sequence harmonically converges to zero at each
new number).
k 
 1
Sp = Pj (7)
j=1
j

Therefore, from the set of Sp obtained results (i.e., the risk of each partner),
the proposed method will be able to measure and analyze the SP’s risk collec-
tively.

Collective Risk Analysis. The second phase of the MARTP method aggre-
gates the results provided by the first phase (that is, the risk level of each pre-
selected SPs) to calculate the VO success probability as a whole (if the VO
formation can succeed or not). To perform this, it is used the Analytic Hierar-
chy Process (AHP) [15], as seen in Stage 2 of Fig. 4. The AHP method arranges
the problems resolution in a hierarchy, starting from the more general element
(usually the goal) to the most specific elements (often alternatives). In this paper,
the problem to be solved using AHP is specified by two components: the goal and
the alternatives. The goal of AHP is to determine the overall VO risk. The alter-
natives consist in the individual risk levels for each SP (Sp ) obtained through
the individual risk analysis (Stage 1 of Fig. 4).
The first step to perform the AHP evaluation consists in defining a normalized
level of risk SpN for each SP p. The normalization procedure is necessary as a
way to enable the collective risk analysis step. The level of risk of each SP (Sp )
is normalized in a scale that varies from 0 to 1 as seen in Eq. 8:
Sp
SpN = (8)
|S|max(S)

where max(S) represent the highest value of S and |S| the number of elements
in S.
Analyzing Virtual Organizations’ Formation Risk in P2P SON Environments 295

After normalizing the risk level of each SP, it is applied a correspondent

weight Wp , which determines the degree of importance of each SPp regarding
the VO. In this work, the degree of importance of each SP is determined by
an external entity named VO Manager [5], which is seen as the main decision
maker. Therefore, the VO Manager plays a key role in the process of evaluating
the VOs since he will inform which SPs have greater or lesser importance, so
prioritizing some specific SPs in relation to the other ones.
For example, given a VO in the formation process, composed by three SPs (as
illustrated in Fig. 4), each SP will have a level of importance (weight). The impor-
tance of each SP can be classified as follows: [0.0; 0.25]: very low; [0.25; 0.50]: rel-
atively low; [0.50; 0.75]: relatively high; [0.75; 1.00]: very high. In this sense, the
VO manager can change the weights Wp according to the degree of importance
that is assigned to each SPp . This feature increase the robustness of the method
when compared to other techniques, by determining collectively the influence
that each SP has within the VO and the level of risk of each one will impact
overall VO risk level.
Accordingly, let be W1 , W2 , ..., Wn the weight of each alternative
S1N , S2N , · · · , SpN associated the goal. The overall goal, i.e., to measure the
VO risk level, is represented by RV O , whose simplified calculation procedure is
shown in Eq. 9:
p

RV O = 1 − Wi SiN (9)
i=1

From the calculation presented in Eq. 9, is obtained the overall level of risk
in the VO formation, considering the importance of each SP in the process.

4 Evaluation Framework
This section presents results of the MARTP method evaluation. A computa-
tional simulation is conducted based on the preliminary results of [2] and [24]
researches, in order to add the risk analysis context and also analyze the impact
in the process of selecting SPs. Next subsections present the results obtained.

4.1 Computational Prototype

The developed computational prototype was split into two modules: BPSS (Best
Peer Selection Service) [2] and DFRA (Decision Framework for Risk Analysis).
The first module implements the BPSS model developed by [2,6] (view Sect. 2.1)
using the PeerFactSim.KOM discrete event simulator [29] to support the creation
of the P2P SON infrastructure and additionally make available the process for
SPs search and selection. On the other hand, the DFRA module focuses specif-
ically on the risk analysis method simulation. This model was integrated with
BPSS in order to group the pre-selected SPs into a new potential VOs and to
perform a MARTP evaluation (see Fig. 3).
296 R.G. Vieira et al.

Regarding technical system specifications, the prototype was built and the
tests were developed in a computer Intel Core i5 3.1GHz, 4.0GB of RAM and
Linux Mint 14.1 64-bit distribution.

4.2 Simulations Setup

The initial configuration for the risk scenario follows the same rules used for the
SP’s selection. The data was taken from the CAIDA project and MaxMind GeoIP
database [30]. The SPs are represented by a set of pre-selected SON peers whose
identifiers (IPs addresses) belong to five geographical domains, corresponding to
the five countries (Portugal, Spain, France, Italy and Germany). They are also
equally distributed between the five domains.
Taking into account the data setup for the risk analysis, the KPIs values
assigned to each SP follows a linear distribution (varying from 0 to 1) during
the simulation. The linear distribution strategy for generating the KPIs values is
primarily used because companies are often very variable and the implementation
of the four chosen KPIs (trust, communication, collaboration and commitment)
in real scenarios depends on the culture and working methods currently applied
by the involved organizations. In the same way, it is also considered that each
SP has participated at 10 previous VOs (in average) when it was selected.

4.3 Results
The results presented in this section aims to present a comparative analysis of
the proposed MARTP method that analyses risk of SPs to form a VO with:
(a) the selection process proposed by [2]; and (b) the DEA-Fuzzy risk analysis
method proposed by [24]. The main goal of this comparison is to verify which of
the three procedures present a more critical analysis regarding the risk level of
sets of “preformed VOs” (i.e., a set of grouped SPs). The overall procedures for
obtaining the selection and risk results are divided into two different phases as
follows:

– The first phase basically performs the process of SP’s search and selection
through the BPSS model [2]. In this paper, the process for VO formation will
take into account a set of three distinct SPs that will provide the following
services: VPN (SP1 ), Billing (SP2 ) and Video-Streaming (SP3 ). For this rea-
son, the BPSS model should be used three-times in order to provide the three
different SPs, each of them providing its particular service.
– The second phase take emphasis on the risk analysis process (MARTP). Thus,
this phase uses as input the three SPs acquired at the first phase (SP1 , SP2
and SP3 ) to group them into a consortia to measure the risk of their collab-
oration in composing a new VO. It is worth to mention that the MARTP
and the DEA-Fuzzy risk analysis methods uses the same criteria (KPIs), thus
enabling the comparison of the results.
Analyzing Virtual Organizations’ Formation Risk in P2P SON Environments 297

The results comprise four test cases, where each one present a different dis-
tribution in the absolute importance of SP1 , SP2 and SP3 (that compose a
potential VO), as specified in Table 1. It means that there will be evaluated
from cases where both the three SPs have little importance in the VO to cases
where both the three SPs have great importance at all.

Table 1. Level of importance assigned for each SP in the four test cases.

SP1 SP2 SP3

Test case 1 (Fig. 6(a)) 0.30 0.30 0.30
Test case 2 (Fig. 6(b)) 0.50 0.50 0.50
Test case 3 (Fig. 6(c)) 0.70 0.70 0.70
Test case 4 (Fig. 6(d)) 0.30 0.50 0.70

The process of comparison between the level of risk of preformed VOs without
risk analysis (i.e., only grouping the three SPs acquired in the first phase into a
consortia) and with risk analysis (analyzing the risk of the previous formed con-
sortia with the MARTP and the DEA-Fuzzy methods) is depicted in Figs. 6(a) –
(d). The simulation comprises 11 sets of individual scenarios divided into clusters
that range [50, 300] SPs. For each of the eleven scenarios (50 SPs, 75 SPs, · · · ,
300 SPs), the first and second phase early mentioned are performed 100 times,
which will result in eleven 3-bar clusters, which one varying from 0 to 100. This
scale is represented by the vertical axis and shows, in percentage, the average
level of risk of the simulated VOs when: not using any method of risk analysis
(just selection), using the MARTP method and using the DEA-Fuzzy method.
It is worth mentioning that the selected SPs in the first phase (i.e., without
risk analysis), will always form VOs since there is no criteria to prevent their
formation beyond those for selection. In this case, the risk level regarding the
first-bar cluster in the results figures will be always assigned to zero. On the
other hand, there is a significant increase on the risk percentage in the analyzed
VOs under the methods that consider the risk analysis as the formation decision
criterion (MARTP and DEA-Fuzzy) for all the four simulated test cases. In this
sense, it can be pointed out that both methods behave like a VO formation filter,
regardless of whether pre-selected SPs have been rated as the best or the worst
according to a particular selection criteria.
Considering average results (and based on a confidence interval of 95 %),
the greatest increase in the level of risk of the analyzed VOs when using the
MARTP method was 77.54 %, and for the DEA-Fuzzy method, 98.81 %. On the
other hand, the smallest increase when adding the MARTP method was 41.22 %,
and for the DEA-Fuzzy method, 8.03 %.
Taking into account the comparison between the two methods of risk analysis,
one can notice a reasonable difference in their VO’s level of risk distribution (in
average), which can be seen in the Figs. 6(a) and (c). Although analyzing the
298 R.G. Vieira et al.

Not considering Risk Not considering Risk

With MARTP method With MARTP method
With DEA−Fuzzy method With DEA−Fuzzy method
100 100
Level of Risk in Average (%)

Level of Risk in Average (%)

80 80

60 60

40 40

20 20

0 0
50 75 100 125 150 175 200 225 250 275 300 50 75 100 125 150 175 200 225 250 275 300
Number of SPs Number of SPs

(a) Test case 1 (b) Test case 2

Not considering Risk Not considering Risk

With MARTP method With MARTP method
With DEA−Fuzzy method With DEA−Fuzzy method
100 100
Level of Risk in Average (%)

Level of Risk in Average (%)

80 80

60 60

40 40

20 20

0 0
50 75 100 125 150 175 200 225 250 275 300 50 75 100 125 150 175 200 225 250 275 300
Number of SPs Number of SPs

(c) Test case 3 (d) Test case 4

Fig. 6. Distribution in the number of formed VOs when not using risk analysis (just
selection), MARTP method and the DEA-Fuzzy method [24].

risk in a similar way (i.e., the more importance have the SPs, the more risky
is the VO) the DEA-Fuzzy method presented a very higher and lower VO’s
level of risk distribution when compared with the MARTP method (41.22 % and
77.54 % for MARTP, versus 8.03 % and 98.81 % for DEA-Fuzzy method) while
in the Figs. 6(b) and (d) there were a similar occurrence of these distributions
(59.76 % and 62.69 % versus 49.04 % and 50.54 %).
The main factor that contributes to this result is related to the bicriterial
facet of the proposed method (MARTP), which takes into account a more crit-
ical and balanced analysis regarding its stages of individual and collective risk
analysis. This means that MARTP avoids the small oscillations in the calcula-
tion of each step that generates large variations in the distribution of VO’s level
of risk in [24]. It means that the performance of a given SP has more equal –
or similar – importance in relation to their relative importance within the VO,
what is desirable to ensure the consistency of the proposed method in different
kinds of scenarios. In the work presented in [24], these minor variations in the
degree of importance of each SP led to huge variation in the distribution of risk
(standing near from 0 % or near from 100 %), as shown in Figs. 6(a) and (c).
Analyzing Virtual Organizations’ Formation Risk in P2P SON Environments 299

In this sense, one can conclude that the risk analysis compared methods
may favour a greater or lesser critical evaluation, where both the individual
SP’s performance and its importance in the VO plays a decisive role in the
final outcome to calculate the final VO level of risk. Therefore, it should be
taken great prevention and control by the VO Manager and risk specialists when
evaluating the importance as well as the competence of each SP in order to
provide confidence in the future operation of the VO, which is the next step
further the VO formation.

5 Conclusions
This paper addressed some issues related to VO risk identification and mea-
surement. Overall, risk analysis has become a key element in VO planning since
small errors can lead them to impairment as a whole. Therefore, it is proposed
a new method to perform a risk analysis in a set of Service Providers (SPs) that
are going to compose a Virtual Organization (VO).
The presented method, named MARTP, is composed of two stages. The first
stage performs an individual risk analysis for all pre-selected SPs, by basing it on
ETA analysis. Having as input the results from the first stage, the second stage
calculates and analyzes the global risk considering all SPs together. It applies
AHP method to accomplish that.
Most of the works in the literature review have approached the problem of
selecting partners via an analysis focused on members competences and capa-
bilities. This work extended this vision incorporating an additional dimension
of decision, which is risk. Therefore, besides considering these two dimensions,
this work qualifies and quantifies the risk of each possible VO, suggesting to
decision-makers a measurable rank of the less risky compositions.
The risk measurement is also based on four chosen KPIs (trust, commu-
nication, collaboration and commitment) that seem appropriate regarding the
technical literature. Moreover, these indicators are combined with real geograph-
ical data in a simulation environment. The performed simulations involved sets
of pre-selected SPs (taken in [2]) in order to explore the comparison between the
proposed method and the method previously proposed by [24], evaluating which
of the two methods presents more criticality in the process of analyzing the risk
in VOs.
The achieved results seem promising about the suitability of the method
regarding its purpose. There were found that the proposed method may be more
or less critical for the assessment of an SP to be part of a VO, where this analysis
is equally dependent on both the individual SPs’ performance and their impor-
tance in the VO, unlike [24]. This particular feature enhances the consistency
of the proposed method when dealing with different kind of scenarios, which is
very common in VOs. The VO Manager plays a key role in the evaluation since
it informs which SPs have greater or lesser importance in the VO.
Likewise, the presented method contributes to a more concrete way to
express, measure, assess and deal with the risks in VO forming, both individ-
ually and collectively, while focusing only on SPs. Nevertheless, the use of the
300 R.G. Vieira et al.

method in the process of risk analysis provides an evaluation with a lower level
of subjectivity, discarding SPs or not, before composing a VO, according to the
established criteria. Future work includes testing the method in near-real scenar-
ios as well as extending the evaluation to an expert panel, in order to improve
its quality.

References
1. Duan, Z., Zhang, Z.L., Hou, Y.T.: Service overlay networks: SLAs, QoS, and band-
width provisioning. IEEE/ACM Trans. Netw. 11, 870–883 (2003)
2. Fiorese, A., Simões, P., Boavida, F.: Peer selection in P2P service overlays using
geographical location criteria. In: Murgante, B., Gervasi, O., Misra, S., Nedjah, N.,
Rocha, A.M.A.C., Taniar, D., Apduhan, B.O. (eds.) ICCSA 2012, Part II. LNCS,
vol. 7334, pp. 234–248. Springer, Heidelberg (2012)
3. Zhou, S., Hogan, M., Ardon, S., Portman, M., Hu, T., Wongrujira, K., Seneviratne,
A.: Alasa: when service overlay networks meet peer-to-peer networks. In: Proceed-
ings of the 11th Asia-Pacific Conference on Communications (APCC 2005), pp.
1053–1057, Perth, Australia (2005)
4. Mowshowitz, A.: Virtual organization. Commun. ACM 40, 30–37 (1997)
5. Camarinha-Matos, L.M., Afsarmanesh, H.: On reference models for collaborative
networked organizations. Int. J. Prod. Res. 46, 2453–2469 (2008)
6. Fiorese, A., Simões, P., Boavida, F.: OMAN – a management architecture for P2P
service overlay networks. In: Stiller, B., De Turck, F. (eds.) AIMS 2010. LNCS,
vol. 6155, pp. 14–25. Springer, Heidelberg (2010)
7. Alawamleh, M., Popplewell, K.: Risk sources identification in virtual organisation.
In: Popplewell, K., Harding, J., Raul, P., Chalmeta, R. (eds.) Enterprise Interop-
erability IV, pp. 265–277. Springer, London (2010)
8. Junior, O.C.A., Rabelo, R.J.: A KPI model for logistics partners’ search and sug-
gestion to create virtual organisations. Int. J. Netw. Virtual Organ. 12, 149–177
(2013)
9. Camarinha-Matos, L.M., Afsarmanesh, H.: Collaborative networks: a new scientific
discipline. J. Intell. Manuf. 16, 439–452 (2005)
10. Afsarmanesh, H., Camarinha-Matos, L.M.: A framework for management of virtual
organization breeding environments. In: Proceedings of the 6th Working Confer-
ence on Virtual Enterprises (PRO-VE 2005), pp. 35–48, Valencia, Spain (2005)
11. March, J.G., Shapira, Z.: Managerial perspectives on risk and risk taking. Manage.
Sci. 33, 1404–1418 (1987)
12. Moskowitz, H., Bunn, D.: Decision and risk analysis. Eur. J. Oper. Res. 28, 247–260
(1987)
13. Vose, D.: Risk Analysis: A Quantitative Guide. John Wiley & Sons, New Jersey
(2008)
14. Ericson, C.A.: Hazard Analysis Techniques for System Safety. John Wiley & Sons,
New York (2005)
15. Saaty, T.L.: Decision making - the analytic hierarchy and network processes
(AHP/ANP). J. Syst. Sci. Syst. Eng. 13, 1–35 (2004)
16. Rychlik, I.: Probability and Risk Analysis: An Introduction for Engineers. Springer-
Verlag, Berlin (2006)
17. Rhee, S.J., Ishii, K.: Using cost based fmea to enhance reliability and serviceability.
Adv. Eng. Inf. 17, 179–188 (2003)
Analyzing Virtual Organizations’ Formation Risk in P2P SON Environments 301

18. Alawamleh, M., Popplewell, K.: Analysing virtual organisation risk sources: an
analytical network process approach. Int. J. Netw. Virtual Organ. 10, 18–39 (2012)
19. Grabowski, M., Roberts, K.H.: Risk mitigation in virtual organizations. J. Comput.
Mediate. Commun. 3, 704–721 (1998)
20. Li, Y., Liao, X.: Decision support for risk analysis on dynamic alliance. Decis.
Support Syst. 42, 2043–2059 (2007)
21. Min, H., Xue-Jing, W., Lu, F., Xing-Wei, W.: Multi-strategies risk programming
for virtual enterprise based on ant colony algorithm. In: Proceedings of the 1st
International Conference on Industrial Engineering and Engineering Management,
pp. 407–411, Singapore (2007)
22. Fei, L., Zhixue, L.: A fuzzy comprehensive evaluation for risk of virtual enterprise.
In: Proceedings of the 10th International Conference on Internet Technology and
Applications, pp. 1–4, Corfu, Greece (2010)
23. Paszkiewicz, Z., Picard, W.: Modelling competences for partner selection in service-
oriented virtual organization breeding environments (2011). CoRR abs/1111.5502
24. Lemos, F.S.B.d., Vieira, R G., Fiorese, A., Junior, O.C.A.: A hybrid dea-fuzzy
method for risk assessment in virtual organizations. In: Proceedings of the 11th
International FLINS Conference on Decision Making and Soft Computing (2014)
25. Pidduck, A.B.: Issues in supplier partner selection. J. Enterp. Inf. Manage. 19,
262–276 (2006)
26. Goranson, H.T.: The Agile Virtual Enterprise Cases, Metrics, Tools. Quorum
Books, Westport (1999)
27. Montgomery, D.C., Runger, G.C.: Applied Statistics and Probability for Engineers.
John Wiley & Sons, New Jersey (2011)
28. Bentley, J.L.: Multidimensional binary search trees used for associative searching.
Commun. ACM 18, 509–517 (1975)
29. Stingl, D., Gross, C., Ruckert, J., Nobach, L., Kovacevic, A., Steinmetz, R.: Peer-
factSim.KOM: a simulation framework for peer-to-peer systems. In: Proceedings of
the 13th International Conference on High Performance Computing and Simulation
(HPCS 2011), pp. 577–584, Istanbul, Turkey (2011)
30. Caida: Macroscopic topology project (2013). http://www.caida.org/analysis/
topology/macroscopic/
 CRAWLER-LD: A Multilevel Metadata
Focused Crawler Framework for Linked Data

Raphael do Vale A. Gomes1(&), Marco A. Casanova1,

Giseli Rabello Lopes1, and Luiz André P. Paes Leme2
1
Departamento de Informática, PUC-Rio, Rio de Janeiro, Brazil
{rgomes,casanova,grlopes}@inf.puc-rio.br
2
Instituto de Computação, UFF, Niterói, Brazil
[email protected]

Abstract. The Linked Data best practices recommend to publish a new tripleset
using well-known ontologies and to interlink the new tripleset with other tri-
plesets. However, both are difficult tasks. This paper describes CRAWLER-LD,
a metadata crawler that helps selecting ontologies and triplesets to be used,
respectively, in the publication and the interlinking processes. The publisher of
the new tripleset first selects a set T of terms that describe the application domain
of interest. Then, he submits T to CRAWLER-LD, which searches for triplesets
whose vocabularies include terms direct or transitively related to those in
T. CRAWLER-LD returns a list of ontologies to be used for publishing the new
tripleset, as well as a list of triplesets that the new tripleset can be interlinked
with. CRAWLER-LD focuses on specific metadata properties, including sub-
class of, and returns only metadata, hence the classification “metadata focused
crawler”.

Keywords: Focused crawler
Tripleset recommendation
Linked data

1 Introduction

The Linked Data best practices [1] recommend publishers of triplesets to use
well-known ontologies in the triplification process and to interlink the new tripleset
with other triplesets. However, despite the fact that extensive lists of open ontologies
and triplesets are available, such as DataHub,1 most publishers typically do not adopt
ontologies already in use and link their triplesets only with popular ones, such as
DBpedia2 and Geonames.3 Indeed, according to [2–4], linkage to popular triplesets is
favored for two main reasons: the difficulty of finding related open triplesets; and the
strenuous task of discovering instance mappings between different triplesets.
This paper describes CRAWLER-LD, a crawler that addresses the problem of
recommending vocabulary terms and triplesets to assist publishers in the publication
and the interlinking processes. Unlike typical Linked Data crawlers, the proposed
crawler then focuses on metadata with specific purposes, illustrated in what follows.
1
http://datahub.io.
2
http://dbpedia.org.
3
http://www.geonames.org.

© Springer International Publishing Switzerland 2015

J. Cordeiro et al. (Eds.): ICEIS 2014, LNBIP 227, pp. 302–319, 2015.
DOI: 10.1007/978-3-319-22348-3_17
CRAWLER-LD: A Multilevel Metadata Focused Crawler Framework for Linked Data 303

In a typical scenario, the publisher of a tripleset first selects a set T of terms that
describe an application domain. Alternatively, he could use a database summarization
technique [5] to automatically extract T from a set of triplesets. Then, the publisher
submits T to CRAWLER-LD, which will search for triplesets whose vocabularies
include terms direct or transitively related to those in T. The crawler returns a list of
vocabulary terms, as well as provenance data indicating how the output was generated.
For example, if the publisher selects the term “Music” from WordNet, the crawler
might return “Hit music” from BBC Music. Lastly, the publisher inspects the list of
triplesets and terms returned, with respect to his tripleset, to select the most relevant
vocabularies for triplification and the best triplesets to use in the interlinking process,
possibly with the help of recommender tools. We stress that the crawler was designed
to help recommender tools for Linked Data, not to replace them.
This paper is organized as follows. Section 2 presents related work. Section 3
summarizes background information about the technology used. Section 4 briefly
explains how the crawler works with the help of an example. Section 5 details the
crawling process. Section 6 describes experiments that assess the usefulness of the
crawler. Finally, Sect. 7 presents the conclusions and future work.

2 Related Work

We first compare CRAWLER-LD with Linked Data crawlers. Fionda et al. [6] present
a language, called NAUTILOD, which allows browsing through nodes of a Linked
Data graph. They introduced a tool, called swget (semantic web get), which evaluates
expressions of the language. An example would be: “find me information about Rome,
starting with its definition in DBpedia and looking in DBpedia, Freebase and the New
York Times databases”.
swget < dbp:Rome > (< owl:sameAs >)* -saveGraph
-domains {dbpedia.org,rdf.freebase.com,data.nytimes.com}
LDSpider [7] is another example of a Linked Data crawler. Similarly to the crawler
proposed in this paper, LDSpider starts with a set of URIs as a guide to parse Linked
Data. Ding et al. [8] present the tool created by Swoogle to discover new triplesets. The
authors describe a way of ranking Web objects in three granularities: Web documents
(Web pages with embedded RDF data), terms and RDF Graphs (triplesets). Each of
these objects has a specific ranking strategy.
The Linked Data crawlers just described have some degree of relationship with the
proposed crawler, though none has exactly the same goals. As explained in the
introduction, the proposed crawler focuses on finding metadata that are useful to design
new triplesets. Furthermore, rather than just dereferencing URIs, it also adopts
crawling queries to improve recall, as explained in Sect. 5.1.
We now comment on how the proposed crawler relates to recommender tools for
Linked Data. Some generic recommender tools use keywords as input. Nikolov et al.
[2, 3] use keywords to search for relevant resources, using the label property of the
resources. Indeed, a label is a property used to provide a human-readable version of the
304 R. do Vale A. Gomes et al.

name of the resource.4 A label value may be inaccurate, in another language or simply be
a synonymous of the desired word. There is no compromise with the schema and its
relationships. Therefore, the risk of finding an irrelevant resource is high. Martínez-
Romero et al. [4] propose an approach for the automatic recommendation of ontologies
based on three points: (1) how well the ontology matches the set of keywords; (2) the
semantic density of the ontology found; and (3) the popularity of the tripleset on the Web
2.0. They also match a set of keywords to resource label values, in a complex process.
CRAWLER-LD may be used as a component of a recommender tool, such as those
just described, to locate: (1) appropriate ontologies during the triplification of a database;
(2) triplesets to interlink with a given tripleset. We stress that the crawler was not
designed to be a full recommender tool, but rather to be a component of one such system.

3 Background

The Linked Data principles advocate the use of RDF [9], RDF Schema [10] and other
technologies to standardize resource description.
RDF describes resources and their relationships through triples of the form (s, p, o),
where: s is the subject of the triple, which is an RDF URI reference or a blank node;
p is the predicate or property of the triple, which is an RDF URI reference and specifies
how s and o are related; and o is the object, which is an RDF URI reference, a literal or
a blank node. A triple (s, p, o) may also be denoted as “<s > <p > <o>”.
A tripleset is just a set of triples. In this paper will use dataset and tripleset
interchangeably.
RDF Schema is a semantic extension of RDF to cover the description of classes and
properties of resources. OWL [11] in turn extends RDF Schema to allow richer
descriptions of schemas and ontologies, including cardinality and other features.
RDF Schema and OWL define the following predicates that we will use in the rest
of the paper:
• rdfs:subClassOf indicates that the subject of the triple defines a subclass of
the class defined by the object of the triple
• owl:sameAs indicates that the subject denotes the same concept as the object
• owl:equivalentClass indicates that both the subject and the object are
classes and denote the same concept
• rdf:type indicates that the subject is an instance of the object
For example, the triple
<dbpedia:Sweden > < rdf:type > < dbpedia:Country > .
indicates that the resource Sweden is an instance of the class Country.
Triplesets are typically available on the Web as SPARQL endpoints [12] or as file
dumps (large files containing all the data from a tripleset, or small files containing only
the relevant data for a defined term). A third option is through URL dereferencing,

4
http://www.w3.org/TR/rdf-schema/#ch_label
CRAWLER-LD: A Multilevel Metadata Focused Crawler Framework for Linked Data 305

which means that the resource contains description data about itself so it is possible to
discover more data simply by reading the resource content.
More than just a query language similar to SQL, SPARQL is a protocol: it defines
the query interface (HTTP), how requests should be made (POST or GET) and how the
data should be returned (via a standard XML). Thus, an agent can perform queries on a
dataset and acquire knowledge to create new queries and so on.
Finally, VoID [13] is an ontology used to define metadata about triplesets. A VoID
document is a good source of information about a tripleset, such as the classes and
properties it uses, the size of the tripleset, etc.
Let d be a tripleset and V be a set of VoID metadata descriptions. The classes and
properties used in d can be extracted from tripleset partitions defined by the properties
void:classPartition and void:propertyPartition that occur in V. Class
partitions describe sets of triples related to subjects of a particular class. Property
partitions describe sets of triples that use a particular predicate. These partitions are
described by the properties void:class and void:property respectively. The set
of vocabulary terms used in d can be generated by the union of all values of the
properties void:class and void:property. In some cases, the VoID description
of a tripleset does not define partitions, but it specifies a list of namespaces of the
vocabularies used by the tripleset with the void:vocabulary predicate. One can
enrich the set of vocabulary terms used in d with such a list.

4 Use Case

Consider a user who wants to publish as Linked Data a relational database d storing
music data (artists, records, songs, etc.). The crawler proposed in this paper will help
the user to publish d as follows.
First, the user has to define an initial set T of terms to describe the application
domain of d. Suppose that he selects just one term dbpedia:Music, taken from
DBpedia.
The user will then invoke the crawler, passing T as input. The crawler will query
the DataHub catalogue of Linked Data triplesets to crawl triplesets searching for new
terms that are direct or transitively related to dbpedia:Music. The crawler focuses
on finding new terms that are defined as subclasses of the class dbpedia:Music, or
that are related to dbpedia:Music by owl:sameAs or owl:equivalent-
Class properties. The crawler will also count the number of instances of the classes
found.
The crawler will return: (1) the list of terms found, indicating their provenance -
how the terms are direct or transitively related to dbpedia:Music and in which
triplesets they were found; and (2) for each class found, an estimation of the number of
instances in each tripleset visited.
The user may take advantage of the results the crawler returned in two ways. He
may manually analyze the data and decide: (1) which of the probed ontologies found he
306 R. do Vale A. Gomes et al.

will adopt to triplify the relational database; and (2) to which triplesets the crawler
located he will link the tripleset he is constructing. Alternatively, he may submit the
results of the crawler to separate tools that will automatically recommend ontologies to
be adopted in the triplification process, as well as triplesets to be used in the linkage
process [14, 15].
For example, suppose that the crawler finds two subclasses, opencyc:Love_
Song and opencyc:Hit_Song, of wordnet:synset-music-noun-1 in the
ontology opencyc:Music. Suppose also that the crawler finds large numbers of
instances of these subclasses in two triplesets, musicBrainz and bbcMusic. The
user might then decide that opencyc:Music is a good choice to adopt in the
triplification process and that musicBrainz and bbcMusic are good choices for
the linkage process.

5 A Metadata Focused Crawler

5.1 Processors
The crawler works with catalogues that use the CKAN framework (such DataHub) to
identify SPARQL endpoints and RDF dumps, and the user can also manually add new
datasets. It receives as input a set of terms T, called the initial crawling terms. Such
terms are typically selected from generic ontologies, such as WordNet,5 DBpedia and
Schema.org,6 albeit this is not a requirement for the crawling process. Given T, the
crawler uses a list C of processors, in successive stages (see Sect. 5.2), to extract new
terms from the triplesets listed in the catalogues.
The tool is engineered as a framework, whose pseudo-code is listed in Appendix 1.
It passes each crawling term t to each processor in C. The processor annotates the
provenance of its crawled data and returns a list of terms to be crawled in the next
stage, after filtering, based on parameters specified by the user (see Sect. 5.2). Cur-
rently, the crawler includes three processors, described in the rest of this section.

Dereference Processor. The first processor is responsible for extracting information of

the resource itself. It tries to find new resources using the properties owl:sameAs,
owl:equivalentClass and rdfs:subclassOf (see Sect. 3). For each such
property, the processor applies a SPARQL query to extract new information. The
following template illustrates how each query works, where p is one of the above
properties and t is the crawling term itself; the values assigned to the variable ?item
are resources that to be crawled in a next stage.
SELECT distinct ?item
WHERE {< t > p ?item}

5
http://wordnet.princeton.edu/.
6
http://schema.org.
CRAWLER-LD: A Multilevel Metadata Focused Crawler Framework for Linked Data 307

Given that owl:sameAs and owl:equivalentClass are reflexive, the pro-

cessor also applies SPARQL queries generated by a new code template, with the
subject and object inverted:

SELECT distinct ?item

WHERE {?item p < t>}

Property Processor. This processor is responsible for crawling other datasets. It uses a
special SPARQL query, which runs over each dataset discovered in DataHub and
manually as described Sect. 5.1. The motivation is to extract information that is not
directly related to the resources already processed. Given the crawling term R that will
be processed by CRAWLER-LD and a dataset D that uses R to describe a fraction of its
data. While a conventional crawling algorithm is not able to find D since R does not
have any reference to D, CRAWLER-LD, on the other hand, traverses all datasets
available and is able to find the relationship between D and R.

The processor uses the following SPARQL template, where t is the resourced
being crawled:

SELECT distinct ?property ?item

WHERE {
{?item owl:sameAs < t>.}
UNION { < t > owl:sameAs ?item.}
UNION {?item owl:equivalentClass < t>.}
UNION { < t > owl:equivalentClass ?item.}
UNION {?item rdfs:subClassOf < t>.}
?item ?property < t>. }

Note that, for each term t to be crawled, the template inverts the role of t (for the
details, see lines 7 and 9 of the code in Appendix (1), when the predicate is owl:
sameAs and owl:equivalentClass, since these predicates are reflexive.
However, the crawler does not invert the role of t, when the predicate is rdfs:
subClassOf, since this predicate is not reflexive.
For example, in the specific case of the crawling property rdfs:subClassOf,
suppose that C and C’ are classes defined in triplesets S and S’, respectively, and
assume that C’ is declared as a subclass of C through a triple of the form

(C’, rdfs:subClassOf, C)

Triples such as this are more likely to be included in the tripleset where the more
specific class C’ is defined than in the tripleset where the more generic class C is
defined. Hence, after finding a class C, the crawler has to search for subclasses of C in
all triplesets it has access using the template above.
308 R. do Vale A. Gomes et al.

Another case occurs when the relationship between C and C’ is defined in a third
ontology S”. Similarly to the previous example, we need a subclass query over S” to
discover that C’ is a subclass of C. S’’ is obtained by dereferencing the URI of C’. In
most cases the returned tripleset is the complete ontology where C’ is defined, while in
some other cases only a fragment of the ontology where C’ is defined is returned.
Instance Counter Processor. The last processor extracts information about the
quantity of instances available in each dataset for each crawling term. It runs queries
over all datasets, using the same principle as the property processor. To reduce the
bandwidth, the processor uses grouping functions to query datasets:

SELECT distinct (count(?instance) AS ?item)

WHERE {?instance rdf:type < %s > . }

Unfortunately, grouping functions are only available in SPARQL version 1.1 [16]
and above. Therefore, the processor also crawls the remaining datasets using the fol-
lowing query, which spend more bandwidth:

SELECT distinct ?item

WHERE {?item rdf:type < %s > . }

5.2 Crawling Stages

The crawler simulates a breath-first search for new terms. Stage 0 contains the initial set
of terms. The set of terms of each new stage is computed from those of the previous
stage, as described in Sect. 5.1.
The crawling frontier is the set of terms found which have not yet been processed.
To avoid circular references, we used a hash map that indicates which terms have
already been processed.
Since the number of terms may grow exponentially from one stage to the next, we
prune the search by limiting:
• The number of stages of the breath-first search
• The maximum number of terms probed
• The maximum number of terms probed in each tripleset, for each term in the
crawling frontier
• The maximum number of terms probed for each term in the crawling frontier
For each new term found, the processors create a list that indicates the provenance
of the term: how the term is direct or transitively related to an initial term and in which
tripleset(s) it was found. That is, the crawler identifies the sequence of relationships it
traversed to reach a term, such as in the following example:

wordnet:synset-music-noun-1 - > owl:sameAs ->

opencyc:Music - > rdfs:subClassOf ->
opencyc:LoveSong - > instance ->
500 instances.
CRAWLER-LD: A Multilevel Metadata Focused Crawler Framework for Linked Data 309

6 Tests and Results

6.1 Organization of the Experiments
We evaluated the crawler over triplesets described in DataHub. The tool was able to
recover 1,042 triplesets with SPARQL endpoints. However, despite this number, it
could run queries on just over 35 % of the triplesets due to errors in the query parser or
simply because the servers were not available.
To create the initial crawling terms, we used three generic ontologies, WordNet,
DBpedia and Schema.org, as well as ontologies specific to the application domain in
question.

Table 1. Namespace abbreviation.

Abbreviation Namespace
akt http://www.aktors.org/ontology/portal#
bbcMusic http://linkeddata.uriburner.com/about/id/entity/http://www.bbc.co.uk/
music/
dbpedia http://dbpedia.org/resource/ or http://dbpedia.org/ontology/
dbtune http://dbtune.org/
freebase http://freebase.com/
freedesktop http://freedesktop.org/standards/xesam/1.0/core#
lastfm http://linkeddata.uriburner.com/about/id/entity/
http://www.last.fm/music/
mo http://purl.org/ontology/mo/
musicBrainz http://dbtune.org/musicbrainz/
nerdeurocom http://nerd.eurecom.fr/ontology#
opencyc http://sw.opencyc.org/2009/04/07/concept/en/
schema http://schema.org/
twitter http://linkeddata.uriburner.com/about/id/entity/http://twitter.com/
umbel http://umbel.org/
wordnet http://wordnet.rkbexplorer.com/id/
yago http://www.yago-knowledge/resource/

WordNet is a lexical database that presents different meanings for the same word.
For example, the word “music” is given two different meanings, denoted by two distinct
terms: wordnet:synset-music-noun-1 means “an artistic form of auditory
communication incorporating instrumental or vocal tones in a structured and continuous
manner”, while wordnet:synset-music-noun-2 means “any agreeable (pleas-
ing and harmonious) sounds; “he fell asleep to the music of the wind chimes””.
DBpedia is the triplified version of the Wikipedia database. The triplification
process is automatically accomplished and the current English version already has 2.5
million classified items.
310 R. do Vale A. Gomes et al.

Schema.org is the most recent ontology of all three. It focuses on HTML semantics
and was created by Google, Bing and Yahoo. Therefore, Schema.org is now used by
many triplesets.7 Schema.org is also developing other ways to increase the search
results by creating a mapping with other ontologies, such as DBpedia and WordNet.
In the examples that follow, we use the abbreviations shown in Table 1.

6.2 Results
The experiments involved two domains, Music and Publications, and used the fol-
lowing parameters:
• Number of stages: 2
• Maximum number of terms probed: 40
• Maximum number of terms probed for each term in the crawling frontier: 20
• Maximum number of terms probed in each tripleset, for each term in the crawling
frontier: 10

Music Domain. We chose Music as the first domain to evaluate the crawler and
elected three ontologies, DBpedia, WordNet and Music Ontology,8 to select the initial
crawling terms. The Music Ontology is a widely accepted ontology that describes
music, albums, artists, shows and some specific subjects.
The initial crawling terms were:
mo:MusicArtist dbpedia:Album
mo:MusicalWork dbpedia:MusicalArtist
mo:Composition dbpedia:Single
dbpedia:MusicalWork wordnet:synset-music-noun-1
dbpedia:Song
In what follows, we will first comment on the results obtained in Stage 1, for each
initial term. Then, we will proceed to discuss how the new terms obtained in Stage 1
were processed in Stage 2.
Table 2(a) shows the results of Stage 1 for mo:MusicalArtist. On Stage 2, for
each of the terms mo:MusicGroup and mo:SoloMusicArtist, the crawler
obtained different results: while mo:MusicGroup recovered over 1.5 million
instances over three datasets, mo:SoloMusicArtist did not find any new result.
Table 2(b) shows the results of Stage 1 for mo:MusicalWork. Note that the
crawler found a variety of instances from multiple databases. On Stage 2, when pro-
cessing mo:Movement, the crawler did not find any new instance or class.
Table 2(c) shows the results of Stage 1 for the first DBpedia term, dbpedia:
MusicalWork. The crawler found 5 subclasses from DBpedia and over a million
instances in 13 datasets, with 8 being DBpedia in different languages (such as French,
Japanese, Greek and others), which was only possible because it focused on metadata.

7
http://schema.rdfs.org/mappings.html.
8
http://musicontology.com/
CRAWLER-LD: A Multilevel Metadata Focused Crawler Framework for Linked Data 311

Table 2. Related terms.

Related terms
Query type Description
(a) Related terms for mo:MusicArtist
subclass mo:MusicGroup, mo:SoloMusicArtist
instance 2,647,957 instances from over four datasets
(b) Related terms for mo:MusicalWork
subclass mo:Movement
instance 1,166,365 instances found in multiple databases
(c) Related terms for dbpedia:MusicalWork
subclass dbpedia:Album, dbpedia:Song, dbpedia:Single, dbpedia:
Opera, dbpedia:ArtistDiscography
instance 939,480 instances from 13 datasets
(d) Related terms for dbpedia:Song
equivalentclass schema:MusicRecording
subclass dbpedia:EurovisionSongContestEntry
instance 35,702 instances from 9 datasets
(e) Related terms for dbpedia:Album
equivalentclass schema:MusicAlbum
instance 871,348 instances from 13 datasets
(f) Related terms for dbpedia:MusicalArtist
instance 424,152 instances from 19 datasets
(g) Related terms for dbpedia:Single
instance 305,041 instances from 10 datasets

Crawlers that use text fields [2] can only retrieve data in the same language as that of
initial terms.
The first three terms, dbpedia:Album, dbpedia:Song and dbpedia:
Single, will be analyzed in the next paragraphs since they are also in the initial set of
terms.
On Stage 2, the processing of dbpedia:Opera returned no results and the
processing of dbpedia:ArtistDiscography returned 48,784 instances, but no
new term.
Table 2(d) shows the results of Stage 1 for dbpedia:Song. The crawler was able
to find a relationship with other generic dataset (Schema.org) and also found a variety
of resources from DBpedia in different languages.
On Stage 2, when processing dbpedia:EurovisionSongContestEntry,
the crawler found 7,807 instances from 7 datasets. The other resource probed on the
Stage 2 was schema:MusicRecording, which returned 38,464 instances and no
new crawling terms.
Table 2(e) shows the results of Stage 1 for dbpedia:Album. The processing of
this term also found schema:MusicAlbum and a large number of instances. On
312 R. do Vale A. Gomes et al.

Stage 2, the tool was able to find 662,409 instances of schema:MusicAlbum, but no
new resource.
Table 2(f) shows the results of Stage 1 for dbpedia:MusicalArtist. The tool
was not able to find any new related resource, but it found a large number of datasets
that have instances of this class.
Table 2(g) shows the results of Stage 1 for dbpedia:Single. The tool found
more than 300 thousand instances from triplesets in many languages.
The last term probed in Stage 1 was wordnet:synset-music-noun-1. The
crawler found a sameAs relationship with an analogue term from another publisher:
http://www.w3.org/2006/03/wn/wn20/instances/synset-music-noun-1.
Finally, we remark that, when we selected the terms to evaluate, we expected to find
relationships between DBpedia and the Music Ontology, which did not happen. In
addition, we found much better results using terms from DBpedia than from the Music
Ontology, which is specific to the domain in question. The definition of links between
the Music Ontology and DBpedia could increase the popularity of the former. For
example, if the term mo:MusicArtist were related to the term dbpedia:Musi-
calArtist, crawlers such as ours would be able to identify the relationship. Also,
matching or recommendation tools would benefit from such relationship.
Publications Domain. For the second domain, we focused on two ontologies,
Schema.org and Aktors,9 which is commonly used by publications databases. We
selected the following terms:

schema:TechArticle
schema:ScholarlyArticle
akt:Article-Reference
akt:Article-In-A-Composite-Publication
akt:Book, akt:Thesis-Reference akt:Periodical-Publication
akt:Lecturer-In-Academia
akt:Journal

The results were quite simple. Both ontologies (Schema.org and Aktors) returned a
small number of instances, but with no complex structure. A quick analysis showed
that almost all triplesets were obtained from popular publications databases (such as
DBLP, IEEE and ACM) by the same provider (RKBExplorer), which used the Aktors
ontology. In addition, the Aktors ontology is not linked to other ontologies, which lead
to an almost independent cluster in the Linked Data cloud.
Processing Times. Table 3 shows the processing time for each experiment. In general,
the time spent to process each term was direct related to the number of terms found
(some exceptions apply due to bandwidth issues). The experiment was performed on a
virtual machine hosted by Microsoft Azure10 with 56 GB and two AMD Opteron™
4171 processors.

9
http://www.aktors.org
10
http://azure.microsoft.com/
CRAWLER-LD: A Multilevel Metadata Focused Crawler Framework for Linked Data 313

Table 3 shows that the minimum time was 4 min, when no new terms were found,
but the maximum time depended on the number of new terms in the crawling frontier
and how the network (and the endpoints) responded.
Finally, we observe that the processing time can be optimized, provided that:
(1) the endpoints queries have lower latency; (2) the available bandwidth is stable
across the entire test; (3) cache features are used.

Table 3. Performance evaluation.

Term Proc. time (minutes)
Music domain
mo:MusicArtist 11
mo:MusicalWork 8
mo:Composition 4
dbpedia:MusicalWork 22
dbpedia:Song 11
dbpedia:Album 8
dbpedia:MusicalArtist 4
dbpedia:Single 4
wordnet:synset-music-noun-1 11
Publications domain
schema:TechArticle 4
schema:ScholarlyArticle 4
akt:Article-Reference 4
akt:Article-In-A-Composite-Publication 8
akt:Book 5
akt:Thesis-Reference 5
akt:Periodical-Publication 4
akt:Lecturer-In-Academia 5
akt:Journal 4

Table 4. Number of terms found using swget.

Term Subclass SameAs Equivalentclass Type
mo:MusicArtist 6 0 0 0
mo:MusicalWork 8 0 0 0
dbpedia:MusicalWork 21 0 0 0
dbpedia:Song 7 0 1 0
dbpedia:Album 6 0 1 0
dbpedia:MusicalArtist 10 0 0 0
dbpedia:Single 6 0 0 0
314 R. do Vale A. Gomes et al.

6.3 A Comparison with Swget

We opted for a direct comparison between CRAWLER-LD and swget for three reasons.
First, there is no benchmark available to test Linked Data crawlers such as ours and it is
nearly impossible to manually produce one such (extensive) benchmark. Second, swget
is the most recent crawler available online. Third, it was fairly simple to setup an
experiment for swget similar to that described in Sect. 6.2 for the Music domain.
Briefly, the experiment with swget was executed as follows. Based on the examples
available at the swget Web site, we created the following template to run queries (where
t’ is the term to be probed and q’ the current crawling property):
t’ -p < q’ > < 2-2>
The above query means “given a term t’, find all resources related to it using the
predicate q’, expanding two levels recursively.
Then, we collected all terms swget found from the same initial terms of the Music
domain used in Sect. 6.2, specifying which crawled property swget should follow.
Table 4 shows the number of terms swget found, for each term and crawling property.
Based on the experiments with swget and CRAWLER-LD, we compiled a list of
terms shown in Appendix 2. Then, we manually inspected the terms and marked those
that pertain to the Music domain and those that swget and CRAWLER-LD found.
The results can be summarized by computing the precision, recall and balanced
F-measure (F1) [17] obtained by swget and CRAWLER-LD for the list of terms thus
obtained:
• swget: Precision = 29.17 % Recall = 100 % F1 = 45.16 %
• CRAWLER-LD: Precision = 100 % Recall = 78.57 % F1 = 88.00 %
These results should be interpreted as follows. Swget achieved a much lower
precision since it finds more generic and more specific terms at the same time, while
CRAWLER-LD only searches for the more specific terms. This feature creates unde-
sirable results for the purposes of focusing on an application domain. For example,
using rdfs:subClassOf as predicate and dbpedia:MusicalWork as object,
swget returned dbpedia:Work, a superclass at the first level. At the next stage, swget
then found resources such as dbpedia:Software and dbpedia:Film, each of
them subclasses of dbpedia:Work, but unrelated to the Music domain.
Swget achieved a larger recall value. However, Swget found only two related
resources that CRAWLER-LD could not find. The basic reason is that the crawling
resources were subclasses of domain resources, which implied that CRAWLER-LD
could not reach the resources, while swget was able to.
Analyzing the overall quality of the crawlers using balanced F-measure,
CRAWLER-LD outperformed swget, obtaining an F1 result almost twice as large as
that of swget. Thus, in this experiment, CRAWLER-LD was able to find a better
balance between recall and precision values than swget.
Swget was also unable to identify the same number of instances as CRAWLER-LD.
While CRAWLER-LD searched a large number of datasets, swget tried to obtain data
CRAWLER-LD: A Multilevel Metadata Focused Crawler Framework for Linked Data 315

only by reading the resource’s content, which does not describes instances from any
other dataset. This behavior should not be regarded as defect of swget, though, but a
consequence of working with a general-purpose crawler, rather than a metadata focused
crawler, such as CRAWLER-LD.

6.4 Lessons Learned

In this section, we highlight the main lessons learned from the results of our experi-
ments. We first enumerate some aspects that may influence the crawling results, such as
the settings of the parameters and the availability of sufficient information about the
crawled triplesets.
Parameter Setting. Since, in our crawler, the set of terms of each new stage is computed
from that of the previous stage, the number of terms may grow exponentially. We defined
some parameters to prune the search. Hence, the user must adequately set such parameters
to obtain results in reasonable time, without losing essential information.
Choosing the Initial Crawling Terms. In the Music domain experiments, we started
with terms from three different triplesets, DBpedia, WordNet and Music Ontology, the
first two being more generic than the last one. It seems that the resources defined in the
Music Ontology are not interlinked (directly or indirectly) with the more popular
triplesets. This limitation is related to the fact that some triplesets do not adequately
follow the Linked Data principles, in the sense that they do not interlink their resources
with resources defined in other relevant triplesets.
Ontologies Describing the Domain of Interest. Our crawler proved to return more
useful when there are relationships among the metadata. In the experiments using the
publications domain, our crawler returned a simplified result because all triplesets
related to the initial crawling terms used the same ontology to describe their resources.
In general, the larger the number of triplesets in the domain, the more useful the results
of our crawler will be.
Reducing the Number of Request. Our crawler demands a high number of requests for
each dataset, and creating ways to reduce this number would improve the performance.
Our approach, primarily implemented on property processor, combines all queries
inside one using the UNION clause and processing the result set locally.
We now highlight some improvements obtained by our metadata focused crawler,
when compared to traditional crawlers.
Discovering Relationships Between Resources of Two Triplesets Described in a Third
One. Using our crawler, we can find cases in which a relationship between two resources
r and r’, respectively defined in triplesets d and d’, was described in another tripleset d”.
This happens, for example, when the ontologies used by d and d’ are only stored in a
different dataset d”. In these cases, it is necessary to crawl all triplesets, other than d and
d’, to find the relationship between r and r’. A traditional crawler following links from
d would not find any link between r and r’ because it is only declared in d”.
316 R. do Vale A. Gomes et al.

Crawling with SPARQL Queries. Our crawler returns richer metadata than a traditional
crawler since it uses SPARQL queries, executed over all triplesets. In particular, our
crawler discovers not only the links between resources, but also the number of
instances related to the crawling terms.

Identifying Resources in Different Languages and Alphabets. Our crawler was able to
identify resources in different languages, even in different alphabets, through the
sameAs and seeAlso queries.

Performing Simple Deductions. Using the provenance lists the crawler generates, one
may perform simple deductions, using the transitivity of the subclass property, perhaps
combined with the sameAs relationship. For example, suppose that the crawler dis-
covered that opencyc:Hit_music is a subclass of opencyc:Music, which in
turn has a sameAs relationship with wordnet:synset-music-noun-1. Then,
one may deduce that opencyc:Hit_music is a subclass of wordnet:
synset-music-noun-1.

7 Conclusions and Future Work

This paper presented CRAWLER-LD, a metadata focused crawler for Linked Data. The
crawler works in stages, starting with a small set T0 of generic RDF terms and creating a new
set of terms Ti+1 by enriching Ti with related terms. The tool is engineered as a framework
that currently includes three processors, implementing different crawling strategies.
In general, the metadata focused crawler introduced in this paper helps simplify the
triplification and the interlinking processes, thereby contributing to the dissemination of
Linked Data. Indeed, the results of the crawler may be used: to recommend ontologies
to be adopted in the triplification process; and to recommend triplesets to be used in the
linkage process.
Finally, the overall crawling process is open to several improvements. For example,
we may use summarization techniques to automatically select the initial set of terms.
We could also use some sort of caching system to improve overall performance.

Acknowledgements. A This work was partly funded by CNPq, under grants 160326/2012-5,
303332/2013-1 and 57128/2009-9, by FAPERJ, under grants E-26/170028/2008,
E-26/103.070/2011 and E-26/101.382/2014, and by CAPES. Microsoft Azure for Research
program was also valuable providing capable machines for experimentations.
CRAWLER-LD: A Multilevel Metadata Focused Crawler Framework for Linked Data 317

Appendix 1: Framework Pseudo-Code

CRAWLER-LD(maxLevels, maxTerms, maxFromTerm, maxFromSet; T, C, PR; Q, P,

D)

CRAWLER-LD(maxLevels, maxTerms, maxFromTerm, maxFromSet; T, C, PR; Q, P, D)

Parameters: maxLevels - maximum number of levels of the breath-first search
maxTerms - maximum number of terms probed
maxFromTerm - maximum number of new terms probed from each term
maxFromSet - maximum number of terms probed from a tripleset, for each term
input: T - a set of input terms
C - a list of catalogues of triplesets
PR - a list of processors
output: Q - a queue with the terms that were crawled
P - a provenance list for the terms in Q
D - a provenance list of the triplesets with terms in Q
begin Q, P, D := empty;
#levels, #terms := 0;
nextLevel := T;
while #levels < maxLevels and #terms < maxTerms do
begin
#levels := #levels + 1;
currentLevel := nextLevel; /* currentLevel and nextLevel are queues of terms */
nextLevel := empty;
for each t from currentLevel do
begin

terms += terms;
if ( #terms > maxTerms ) then exit;
add t to Q;
resourcesForEachDataset := (dataset,resourceList) := empty
for each p from PR do
begin
/* use t on the processor p and save the results for each dataset */
call (dataset,resultList) := p(t,P,D)
add p to resourceForEachDataset
end
/* limiting results phase */
resourcesFromTerm := empty
for each dataset d from resourcesForEachDataset
begin
resultList := results from dataset D on term t;
truncate resultList to contain just the first maxFromSet terms;
resourcesFromTerm := concatenate(resultList, resourcesFromTerm);
end
truncate resourcesFromTerm to contain just the first maxFromTerm terms;
nextLevel := concatenate(resourcesFromTerm, nextLevel);
end /* t loop */
end /* level loop */
return Q, P, D;
end /* algorithm */
318 R. do Vale A. Gomes et al.

Appendix 2: A Comparison Between Swget and CRAWLER-LD

for the Music Domain
CRAWLER-LD: A Multilevel Metadata Focused Crawler Framework for Linked Data 319

References
1. Bizer, C., Heath, T., Berners-Lee, T.: Linked data - the story so far, lnt’l. J. Seman. Web
Info. Sys. 5(3), 1–22 (2009)
2. Nikolov, A., d’Aquin, M.: Identifying relevant sources for data linking using a semantic web
index. In: Proceedings Workshop on Linked Data on the Web. vol. 813 of CEUR Workshop
Proceedings, CEUR-WS.org (2011)
3. Nikolov, A., d’Aquin, M., Motta, E.: What should I link to? identifying relevant sources and
classes for data linking. In: Proceedings Joint Int’l Semantic Technology Conference,
pp. 284–299 (2012)
4. Romero, M.M., Vázquez -Naya, J.M., Munteanu, C.R., Pereira, J., Pazos, A.: An approach
for the automatic recommendation of ontologies using collaborative knowledge. In: Setchi,
R., Jordanov, I., Howlett, R.J., Jain, L.C. (eds.) KES 2010, Part II. LNCS, vol. 6277, pp. 74–
81. Springer, Heidelberg (2010)
5. Saint-Paul, R., Raschia, G., Mouaddib, N.: General purpose database summarization. In:
Proceedings 31st Int’l Conference on Very Large Data Bases. VLDB Endowment, pp. 733–
744 (2005)
6. Fionda, V., Gutierrez, C., Pirró, G.: Semantic navigation on the web of data: specification of
routes, web fragments and actions. In: Proceedings of the 21st Int’l Conference on World
Wide Web, pp. 281–290 (2012)
7. Isele, R., Harth, A., Umbrich, J., Bizer, C.: LDspider: an open-source crawling framework
for the web of linked data. In: Proceedings Int’l Semantic Web Conference (Posters),
Shanghai, China (2010)
8. Ding, L., Pan, R., Finin, T.W., Joshi, A., Peng, Y., Kolari, P.: Finding and ranking
knowledge on the semantic web. In: Gil, Y., Motta, E., Benjamins, V., Musen, M.A. (eds.)
ISWC 2005. LNCS, vol. 3729, pp. 156–170. Springer, Heidelberg (2005)
9. Manola, F., Miller, E.: RDF Primer, W3C Recommendation 10 February 2014 (2004)
10. Brickley, D., Guha, R.V. (eds.): RDF vocabulary description language 1.0: RDF schema. In:
W3C Recommendation 10 February 2004 (2004)
11. W3C OWL Working Group: OWL 2 Web Ontology Language Document Overview
(Second Edition). W3C Recommendation 11 December 2012
12. Prud’hommeaux, E., Seaborne, A.: SPARQL Query Language for RDF, W3C
Recommendation 15 January 2009
13. Alexander, K., Cyganiak, R., Hausenblas, M., Zhao, J.: Describing linked datasets - on the
design and usage of void, the ‘vocabulary of interlinked datasets’. In: Proceedings
Workshop on Linked Da-ta on the Web (LDOW 2009), Madrid, Spain (2009)
14. Leme, L.A.P., Lopes, G.R., Nunes, B.P., Casanova, M.A., Dietze, S.: Identifying candidate
datasets for data interlinking. In: Daniel, F., Dolog, P., Li, Q. (eds.) ICWE 2013. LNCS, vol.
7977, pp. 354–366. Springer, Heidelberg (2013)
15. Lopes, G.R., Leme, L.A.P.P., Nunes, B.P., Casanova, M.A., Dietze, S.: Recommending
tripleset interlinking through a social network approach. In: Proceedings 14th International
Conference on Web Information System Engineering, Nanjing, China (October 13–15,
2013), pp. 149–161 (2013)
16. Garlik, S.H., Seaborne, A., Prud’hommeaux, E.: SPARQL 1.1 query language. World Wide
Web Consortium (2013)
17. Manning, C.D., Raghavan, P., Schütze, H.: Introduction to Information Retrieval.
Cambridge University Press, New York (2008)
 A Framework Reactive and Proactive
for Pervasive Homecare Environments

Alencar Machado1,2(&), Daniel Lichtnow2, Ana Marilza Pernas4,

Amel Bouzeghoub5, Iara Augustin3, Leandro Krug Wives1,
and José Palazzo Moreira de Oliveira1
1
Instituto de Informática, Universidade Federal do Rio Grande do Sul,
Porto Alegre, Brazil
[email protected], {wives,palazzo}@inf.ufrgs.br
2
Colégio Politécnico, Universidade Federal de Santa Maria, Santa Maria, Brazil
[email protected]
3
PPGI, Universidade Federal de Santa Maria, Santa Maria, Brazil
[email protected]
4
Centro de Desenvolvimento Tecnológico, Universidade Federal de Pelotas,
Pelotas, Brazil
[email protected]
5
Télécom SudParis, UMR CNRS 5157 SAMOVAR, Evry, France
[email protected]

Abstract. Smart environments are spaces that interact with users taking into
account their needs and preferences. Systems that manage these environments
need to manipulate the context for interacting in a suitable way with users. One
big challenge for creating a smart environment is to deal with context dynam-
icity. Ideally, these computer-supported environments must detect relevant
events to forecast future situations and to act proactively to mitigate or eliminate
situations related to specific user’s needs. This paper proposes a framework for
providing extensible, reactive and proactive behavior in Smart Environment
systems. The focus of the present work is related to a specific class of Smart
Environments: an Ambient Assisted Living. In this sense we define a framework
and illustrate practical aspects of the use of framework by describing a
home-care scenario in which the system observes the behavior of the user, as the
time goes by, and detects relevant situations and acts reactively and proactively
for preserving user health condition.

Keywords: Proactive
Reactive
Situation
Context
Event
Ambient

intelligence

1 Introduction

Ambient Intelligence (AmI) aims to support people in their needs and to prevent user’s
behaviors [1]. Fields such as Ambient Assisted Living (AAL) [2] and Smart Homes [3]
are emerging as Smart Environment focused on specific characteristics of the user.
These systems must (a) manage heterogeneous sources (sensors and appliances) to
provide high level information such as situations; (b) process events for detecting

© Springer International Publishing Switzerland 2015

J. Cordeiro et al. (Eds.): ICEIS 2014, LNBIP 227, pp. 320–338, 2015.
DOI: 10.1007/978-3-319-22348-3_18
 A Framework Reactive and Proactive for Pervasive Homecare Environments 321

situations in the environment; (c) make predictions of unwanted situations and to react
in advance; (d) determine the policy of actions to consume appropriate services for
adapting the environment ahead the situation envisaged; (e) have expansive capacities
to manipulate different situations. However, an Ambient Intelligence system does not
incorporate reactive and proactive behavior in an adequate way.
Thus, taking into account these challenges, this paper focuses on how to process
events to detect and predict future unwanted situations. In this sense, we argue that for
fulfilling user needs, AmI systems should be reactive and proactive. Therefore, these
systems must be aware of the user’s current situation and foresee future situations. The
system must make decisions in advance, taking into account evidences that demon-
strate the possibility of an unwanted situation happening in the future.
In our approach, the user and his/her actions are monitored through sensors that
capture environmental data. This data is used to characterize the user context, using
entities for obtaining a semantic characterization that determines the state of the
environment. In the proposed approach, the state of the environment is called “situa-
tion”. Thus, when a situation is detected, if necessary, it is possible to act reactively and
proactively on the environment, using capabilities (services) provided by electronically
controlled devices, seeking to adapt automatically the environment according to the
detected situation.
In our approach the actions of the system are achieved by using functionalities
implemented by Web services embedded in physical objects such as mobile phones,
televisions and radios. The use case of this work focuses on unwanted situations
involving the lives of elderly people at home. In this context, systems that foresee and
handle unwanted situations proactively can assist caregivers of users who do not have
physical or cognitive conditions for managing their own health, because it is necessary
to act before an unwanted situation occurs.
As first concept proof, the approach was applied to a specific case-based application
for managing medicines. In this case, citizens self-manage their health, and an appli-
cation assists this activity. Over time, the citizens are affected by cognitive decline,
when they become unable to manager their medications and the system must adapt
itself to assist caregivers. The aim is to analyze the user behavior for predicting the
need of some early actions to aid them to take their medications at the right time,
preventing a decrease in their health condition.
This paper is organized as follows. Section 2 discusses background and related
work. Section 3 presents the proposed reactive and proactive approach. In Sects. 4 and
5, presents the case study developed. Finally, in Sect. 6, we present and discuss our
conclusions and future works.

2 Background and Related Work

Ambient intelligence systems need to know the world around users they monitor and,
in order to perform actions, they need to interact with users through the devices that
surround them [4]. Therefore, intelligence systems must be context-aware and proac-
tive, automatically adapted to changes in the environment and considering user needs,
without requiring their personal attention.
322 A. Machado et al.

Regarding to context-aware systems, the concept of context is not new. Among the
large number of existing definition in the literature, we adopt the definition of Ye,
Dobson and McKeever [5], in which context is seen as “the environment in which the
system operates”. It’s different from the most adopted definition of Dey [6]: “context is
the situation of an entity in an environment”. In the present work, the context of an
environment is thus represented by a set of entities that surround or interact with the
user, and their semantic relations.
As the time goes by, different entities or interactions may be active. In this sense,
we need to verify the current contextual state of the user and act upon it or on its
changes. For that, we define the concept of situation, since it is used by us to char-
acterize the state of the user environment. For instance, Ye, Stevenson and Dobson [7]
define situation as “the abstraction of the events occurring in the real world that are
derived from the context and hypotheses about how the observed context relates to
factors of interest”.
In this sense, applications that deal with situations are called situation-aware.
Awareness implies vigilance in observing or alertness in drawing inference from a
previous experience, so something is aware only if it is able to observe some objects
and design conclusions through previous observations [8]. Observations could be made
by services provided through devices, such as sensors. Therefore, by these observa-
tions, it is possible to detected events that change the state of the environment, char-
acterizing thus a situation.
Another important concept is the notion of event. According to Etzion and Niblet
[9], “an event is an occurrence within a particular system or domain, it is something
that has happened, or is contemplated as having happened in that domain”. Events can
be modeled as raw or derived. Derived events are higher-level events in the semantic
hierarchy. It normally corresponds to a pattern of observation. Raw events are produced
by some entity of context (e.g., sensors). Events can change the state of the environ-
ment, therefore producing new situations.
Works such as SOPRANO [10], PERSONA [11], among others, aim at modeling
context, events and situations in middleware systems to provide a platform of health
services in Ambient Assisted Living (AAL). They propose conceptual models to
transform homes into AAL environments, modeling their context and services [12].
SOPRANO, for instance, has the intention of recognizing facts, objects, and people
surrounding users allowing systems to act more appropriately and providing support to
daily activities. However, in these works, the system is reactive, since it only acts after
the evidence that an unwanted situation has occurred. They are not able to provide
proactive actions to mitigate or eliminate undesired situations in advance.
The term proactive computing was first described by Tennehouse [13], who pro-
posed the following principles for proactive systems: “they should be closely connected
with their surrounding world; they should also deliver results to humans before the
user action; and they must operate autonomously”.
The characteristics proposed by Tennehouse turn systems essentially reactive. In
this sense, proactive computing overlaps with the term of autonomic computing [14].
An autonomic system is one that reacts to events that already happened. Our approach
follows the proactivity definition of Engel and Etzion [15] who describe proactively in
computing systems as “the ability to mitigate or eliminate undesired future situations,
 A Framework Reactive and Proactive for Pervasive Homecare Environments 323

identifying and taking advantage of future opportunities by applying prediction and

automated decision making technologies”. Thus, the aim of the system actions is to
prevent future unwanted situations. One example is the work of Fu and Xu [16] where
event correlations are used for predicting future failures in networked computing
systems.
The current vision of proactive behavior is listed as the next phase in the evolution
of complex event processing [9]. Thus, Engel and Etzion [15] present the Proactive
Event-Driven Computing. They propose an extension of the event processing con-
ceptual model and include more two types of agents to the architecture of proactive
event-driven applications: predictive and proactive.
Proactive systems apply prediction methods for predicting future information and
decision making. Boytsov and Zaslavsky [17], for instance, analyze and compare
prediction methods in order to identify their benefits and shortcomings. Among these
methods, they describe Bayesian networks as an appropriate approach for predictive
models. Similarly, Nazerfard and Cook [18] present a sequence-based activity pre-
diction approach that uses Bayesian networks in a two-step process to predict both
activities and their corresponding features.
Lotfi et al. [19] seek to make prediction of abnormal behavior of elderly with partial
dementia. They use sensor data for identifying anomalous sensor behaviors to predict
the future values of the possible data for each sensor. The predicted values are used to
inform a caregiver in the case of anomalous behavior of sensors in the near future.
Analyzing the facts presented above, we identify that works related to smart
environments propose strictly reactive systems. We have seen some researches
describing proactive behavior to anticipate user’s actions, but reacting only after a
situation has happened, such as, systems for handling situations of agitations for elderly
patients who take actions to anticipate actions of the caregivers. They, however, do not
seek to identify this situation in advance to avoid it happening. Besides, in general
these proposals do not address or include extensibility technologies, i.e., are not able to
handle different situations in the course of time.

3 A Reactive and Proactive Approach

Our approach differs from other works because we present a new reactive and proactive
approach that is more appropriate to attend the proper demands of AAL systems.
Besides, it provides extensibility for residential smart environments. The approach is
explained taking into account the recent history of self-management of a citizen’s
health, where the system triggers reactive and proactive actions.
The extensibility aspect of our approach is related to the concept of pervasive
applications, and is based in a work named Situation as a Service (SIaaS) [20], which is
described in the next section. In the present work, we have added temporal aspects,
prediction and decision making techniques in order to prevent the existence of
unwanted situations in future (Sect. 3.2).
324 A. Machado et al.

3.1 Situation as a Service

In our approach, pervasive applications are installed in a middleware named SItuation
as a Service (SIaaS). Pervasive applications (appPerv) are software applications
developed by companies specialized in specific fields, such as health, surveillance,
energy. This feature allows to attend the requirement ‘ready to use’, plug-and-play
style. Designers of appPerv must implement in a conceptual model that corresponds to
specific situations of the environment that are relevant to the appPerv. They also must
inform the appPerv context of interest, generating instances of a particular type (e.g.,
Patient, Sensor) and make the linking semantics among them (as hasSensor).
Therefore, the pervasive application informs the SIaaS middleware about the sit-
uations that are important and should be managed for the detection and prediction of
situations, as well as a set of contextual information necessary for decision making. In
this work, we are interested in a specific domain: home-care health support. Thus, users
could extend the capabilities of the middleware buying a complete solution for man-
aging chronic diseases or only one pervasive application for managing the schedules of
their medicines. For example, the pervasive applications described by Machado et al.
[20] perform reactive actions (consumption of services) when a patient’s agitation
situation becomes true.
The SIaaS manages the environment and provides the context of interest, as well as
the situation of interest for pervasive applications, so it is possible to trigger the more
appropriate action when a situation is detected. It is presented in Fig. 1.

Fig. 1. Levels of the architecture.

 A Framework Reactive and Proactive for Pervasive Homecare Environments 325

The SIaaS is a Service-Oriented Architecture (SOA) with complex event processing

that is implemented in three levels: Lower, Intermediate, and Upper. The Lower Level
(3) corresponds to the physical environment (i.e., sensors and appliances).
The Intermediate Level (2) contains the middleware for providing a stable and
secure environment for pervasive applications. The Applications Manager Subsystem
installs applications and obtains the context and situation of interest. After this,
Applications Manager Subsystem subscribes the situation of interest in Subsystem to
Manager Prediction and Inference, also subscribes the context of interest of the
installed application in Subsystem to Manager Context.
The Subsystem to Manager Context has two modules: the Monitor Module and the
Complex Event Processing Module. The Monitor Module generates instances of events
and manages raw data from sensors. Each time an event arises, Complex Event Pro-
cessing Module sends the event to the Complex Event Processing Module. After the
Complex Event Processing Module uses the pattern informed in the event description,
it processes the flow of events to determine if some received events are relevant. If it is
the case, the Complex Event Processing Module notifies the Prediction and Inference
Managing Subsystem. The Prediction and Inference Managing Subsystem is composed
of two models that are activated at the same time: Inference and Prediction. The
Module of Inference processes the rules to detect the current situation, and the Module
of Prediction processes the influence model (Sect. 3.4) to determine the probability of a
situation happening in the future. If a situation happens or could happen in the envi-
ronment, the application assigned to the situation is started. At this moment, the per-
vasive application chooses the most relevant action to manipulate the situation and
sends a request to the Manager Action Subsystem to execute the action. The Manager
Action Subsystem needs to select the services with the most appropriate functionalities
of appliances, according to the detected situation.
The Upper Level (1) offers a stable computing environment for pervasive
applications.

3.2 Proactive Model

A pervasive application has interest in specific situations that involve users and what
happens in their living environment. The pervasive application contains the knowledge
of what kind of reactive and proactive actions (two business processes) should be taken
when an unwanted situation happens. The SIaaS middleware monitors the events that
could generate a situation of interest for any pervasive application, and notifies it
whenever such situation becomes true.
These events occur in time windows and are related to situations. These situations
are described by pervasive applications that specify their situations of interest and
subscribe their situations in the middleware. Taking into account examples from the
health domain, we consider a scenario where the aim is to monitor medicine admin-
istration. Thus, the pervasive application declares to the middleware its interest in the
event “medicine X is not administrated”, and also provides how this event can be
detected through of model (Sect. 3.4). For instance, if it is detected during consecutive
days, it may result in health problems for the patient, i.e., an unwanted situation.
326 A. Machado et al.

Fig. 2. Environmental states.

The SIaaS must avoid such unwanted situations (e.g., the patient forgot to take his
medicine and became sicker). For this purpose, the SIaaS initially learns a predictive
model using data provided by the pervasive application. The pervasive application data
consists of patient’s behavior patterns (proactive model and data) that will be managed
by the SIaaS. After the learning stage, the SIaaS can predict situations through events
detected in real-time and is able to perform proactive actions, avoiding the occurrence
of unwanted situation. Consequently, the reactive actions requested by the pervasive
applications will not be executed, because the events (e.g., patient forgot to take his
medicine) that would determine the situation will not occur. In this sense, the SIaaS
acts proactively to prevent an unwanted situation.
As depicted in Fig. 2, the environment may be in two states: controlled or
uncontrolled. An event stream, predefined as normal (pattern that is not interest of the
middleware), characterizes a controlled environment where reactive actions are helpful.
However, if the events are being detected outside this predefined state (interest pattern
of the middleware), it could characterize that the environment will become uncon-
trolled, thus increasing the dependency on proactive actions being performed by the
SIaaS.
In the environment, event streams are constantly monitored through data made
available by sensors. Still, evaluating the events flow of Fig. 2, at t+1 a pattern of
events (provided by the pervasive application) is detected (1) by the Context Manager.
This subsystem uses a prediction algorithm to determinate the probability of an
unwanted situation becoming true in the user living environment. The Context Man-
ager has t+2 times to make the prediction, having enough time to take corrective
actions. After identifying the probability of the occurrence of a situation in t+2 (2), the
 A Framework Reactive and Proactive for Pervasive Homecare Environments 327

Fig. 3. Sliding Windows

Inference Manager processes the respective rules to determinate if the rate of proba-
bility is relevant. If it is positive, the Proactive Actions Manager must be activated in t
+3 (3) to trigger proactive actions. This module is responsible for choosing the most
appropriate policy to consume appropriate services corresponding to environmental
devices and health care providers.
All these actions are taken in order to conduct the environment towards a normal
state and to avoid future unwanted situations. The goal is to return to the initial streams
of events, thereby characterizing the consistency of the environment.

3.3 Sliding Windows

In the present approach, we consider that it is necessary to analyze a sequence of events
for learning and detecting patterns aiming to predict situations. Besides, we consider
that this sequence of events occurs in a period of time, referenced as Sliding Window,
i.e., a valid space of time where situations are predicted or detected and the relevant
decisions are taken. The Sliding Windows model used in our approach is adapted from
Salfner, Lenk and Malek [21]. In this model, a sliding can be sized or timed. A sized
sliding window (SSW) has a specific size that corresponds to the number of events of a
given pattern of interest. For instance, it is possible to use a SSW with the last hundred
events that match specific selection criteria. A timed sliding window (TSW) has a finite
time frame where events of interest are monitored. In this work, only Δtd is modeled as
these two possibilities, the remainders are windows of time, because they are used to
model the window in future for an unwanted situation.
Figure 3 presents a sliding window associated with a real time proactive behavior.
At time t (current time), the possibility of occurrence of a situation can be predicted
with some time in advance. This period of time is called prediction time (Δtpr) and is
based on the events currently detected in the environment. Situations are predicted in
Δtpr, which uses a size or time sliding window (Δtd) that corresponds to the event
stream monitored by the system. These timed or sized windows (Δtd) are used to
perform prediction. We assume that proactive actions are valid for some period of time,
named period of proactivity (Δtp). In this time window, triggered actions can change
predicted situations, which are expected to occur in the reaction time (Δtr). If Δtpr >
Δtp then there will not be enough time for all proactive actions to be triggered before
the predicted situation (i.e., an unwanted situation) becoming true. Thus, Δtr is the
maximum time the system has to react, since Δtr is the time estimated to the situation to
occur. Then, Δtr is the period where reactive actions, related to a specific unwanted
situation, are triggered.
328 A. Machado et al.

Fig. 4. Conceptual model.

3.4 Event and Situation Model

In this work, the semantic relations {R} that form the context are represented by triples
‹Es, p, Eo› where the subject Es and the object Eo represent instances of entities of the
environment, which could belong to the same domain or not. Similarly, as formalized
by Ye, Stevenson and Dobson [7], p represents a context predicate that encapsulates
two entities of context in a relation. For instance, in the relation ‹John, hasSensor,
RFID› John and RFID represent entities. Subjects and objects can also be represented
by variables in reasoning rules. For example, in the following triple, ‹x, hasSensor, y›,
x represents any entity instantiated in the user domain and y represents any entity
instantiated in the sensor domain. In this example, any pair of values of User and
Sensor, related by the relationship hasSensor, can validate this context predicate.
Figure 4 depicts the conceptual model of our approach. The Activity entity rep-
resents daily activities performed by the citizen in his home, like breakfasting,
watching television, taking medicine or doing exercises. The activities are made up of
human actions, for instance the activity to take medicine is composed by picking up a
glass with water and taking the drug, represented by the semantic relations presented
below (1).

Besides human actions, we also take into account automated actions taken by the
system. For instance, the system may notify the citizen through an audible warning
using some device (we consider that automated actions are actions performed by
devices). Thus, actions are carried out by an agent (human or device) in order to
achieve a goal. When an action either achieves or not its goal, it can generate events.
In our approach, we adopt internal and external events, which are defined in the
theory of Situation Calculus [22]. An external event is generated externally by human
 A Framework Reactive and Proactive for Pervasive Homecare Environments 329

actions or by interacting with pervasive applications. In addition, external events can be

generated by changing a user device or by changing the network connection used by a
device. Internal events are generated internally by the system and are represented by
assertions that can be given by an axiom. An internal event could be generated by an
automated action, the detection of successive state variable changes, or by the modi-
fication of one specific state variable. Thus, as more human actions are transferred to
automated actions, more extensible the system becomes. Events are represented by the
following syntax (2).

Event : ðname; type; time; fRg; pÞ ð2Þ

An event has a name and is characterized by a type (internal or external), a time

(timestamp) within Δtd windows and a set of contextual semantic relations {R}. When
the event is not produced by a single entity (e.g., raw data sensor), it may also have a
detection pattern (p). Events can be linked to one or more contexts; for instance, a
pattern that defines that an event must be detected if a specific sequence of events
happens within a given sliding window of time or size involving the “user” in his/her
living room. In this work, events can determine the evidence of the beginning and the
ending of a situation. Thus, events change the state of the environment and characterize
a new situation. The current situation is represented by the following syntax (3).

Cs : ðname; Ie; fag; FeÞ ð3Þ

As shown in (2), the current situation (Cs) has a name and a set of events that
characterize its beginning (Ie) and ending (Fe), and the time attribute of these events
that characterize the valid time window of this situation, which will always be in Δtr. In
addition, the current situation has a set of triggered reactive actions {a} that were
detected during a valid time for handling the current situation. For instance, below we
present how to represent an event (Ie) that initiates the “unmedicated” situation (4), its
corresponding final event (Fe) and the actions to be performed in this situation.

The event evaluation can lead the system to find out that an unwanted situation has
a probability of happening in the future. In (5) we show that a Predictive Situation (Ps)
is characterized by a set of events; a set of patterns (p), which describes some form of
correlation among events that shape this situation, the probability value (pr) of its
330 A. Machado et al.

occurrence in a context in the future; and a timestamp (time) during which it may occur
within Δtr.

Ps : ðname; feventg; fpg; pr; timeÞ ð5Þ

Figure 5 presents the influence model. The model is a set of events modeled by a
pattern that determine a correlation model among the events. For example, the detec-
tion of the event “E3” influences event “E1”, and the detection of the event “E4”
influences “E1”, and “E1” influences the arise of predictive situation. In this model, a
situation is always a left node and their parents are always events.

4 Case Study Scenarios

This case study aims to demonstrate the use of our approach in a scenario where the
necessity of a mechanism that acts in proactive way is emphasized. The scenario is
related to the aforementioned self-administration of medicines by patients in their
homes. The aim is to identify when patients are no longer able to control their own
medication.
In this sense, we are considering that patients, in general, will take their medicine in
a stipulated period of time or after an action system (e.g., using functionalities (ser-
vices) of devices of the environment to remember the patient about the medication).
However, some patients can have cognitive decline over time, compromising the
self-management of their health condition, thus needing help to take the medication at
the correct time. In this scenario, the situation known as “unmedicated” becomes
habitual, and reactive/proactive actions are necessary to control this situation, con-
sidering the recent history of self-management of citizen’s health, thus assisting the
patient to take his medicine in the best possible way. We considered a pervasive
application deployed in the SIaaS to help patients in the described scenario. For the
implementation, we have used the monitoring component Esper [23], the Bayesian tool
Netica [24], which provided an API that we used to create the Bayesian network that
was inserted into the Context Manager. In addition Jess [25] was used to build Java
software process able to perform inference rules in the Inference Manager module.
For the case study, the following fictitious scenario was considered for describing
the approach supported by the decision making process implemented by the pervasive
application. Imagine ‘Ms. Smith’, a 70 years old citizen who has some aging associated

Fig. 5. Influence model.

 A Framework Reactive and Proactive for Pervasive Homecare Environments 331

diseases such as diabetes, hypertension and lightweight dementia. Ms. Smith’s home is
an intelligent environment managed by a SIaaS middleware, where a number of per-
vasive applications are installed. An example is the pervasive application for managing
medications (appPervMed).
Ms. Smith initially controls the medication herself. However, as any ordinary
person, sometimes, to be involved in some particular activity, she forgets to take or
takes her medicines late, which puts her in an “unmedicated” situation. In these cases,
the appPervMed requests to the SIaaS middleware to trigger an audible or visual
warning through devices located near Ms. Smith. Whenever a warning reaches her
attention, she can interact with the system through a smart phone or smart TV to report
explicitly that she took her medication. After that interaction, the system will close
(finish) the “unmedicated” situation.
After some stipulated time, if Ms. Smith does not take her medication, the system
sends a warning to her caregiver. It warns her caregiver that Ms. Smith had not taken
her medicine, thus placing the responsibility of interacting with the SIaaS on the
caregiver. Once the caregiver gives her the medicine, and informs the system about
that, the appPervMed will know that Ms. Smith took the medicine and will determine
the end of the “unmedicated” situation.
Eventually, the caregiver himself may forget or may be not close to any device that
could warn him about the moment that Ms. Smith must be medicated. Thus, the event
“medicine X not taken” is detected, corresponding again to the beginning of the sit-
uation of “Ms. Smith is unmedicated”. Audible and visual warnings are generated in
different moments in the environment, and, after some parameterized time, the care-
giver is warned. The appPervMed waits for a notification that Ms. Smith took the
medication by the caregiver. If it is not notified in a specific period, the appPervMed
triggers a warning directly to the healthcare provider (consuming a specific Web ser-
vice), placing the responsibility on the healthcare provider to make Ms. Smith taking
her medicine, and, once taken, it ends the situation.
As explained, alerting the caregiver is an exception. However, after some time, if Ms.
Smith takes her medicine only after a system warning to her caregiver and if this behavior
becomes more usual, this behavior may indicate a cognitive decline of Ms. Smith.
Thus, it is necessary to identify (in a proactive way) when the cognitive impairment
happens, because if this identification does not happens fast, the treatment will be
harmed by the administration of drugs in wrong times. This moment may characterize
the end of the patient’s ability to medicate her-self, requiring the caregiver to this
function. Therefore, the system must adapt itself and assist the caregiver in his task of
assisting Ms. Smith.

5 Applying the Approach to the Case Study

The scenario described before shows that a pervasive application must react to an event
(Ms. Smith did not take her medication) that characterize the unmedicated situation and
also must forecast some situation (e.g., Ms. Smith will not take her medication without
her caregiver help) and be proactive.
332 A. Machado et al.

In this paper, the pervasive application has two business processes for taking
decision (reactive and proactive). Next, following our approach, we show how a
pervasive application should work either reactively and proactively.

5.1 Reactive Behavior

The events related to this case study that are relevant to the appPervMed are described
here, as proposed in Sect. 3.4. For performing this task we used the Semantic Web Rule
Language (SWRL), using the already defined semantic relations {R} and triplets in the
form ‹Es, p, Eo›, as presented in Sect. 3.4.
The Fig. 6 presents the reactive decision making process where the appPervMed is
selected when Ms. Smith does not took her medicine and always it is initiate because
event “ea1” is detected. The patterns to detect events were modeled as Esper state-
ments. The variables were replaced to the values used in the scenario to provide an
easier interpretation.

Event name: ea1; description: Not took the medicine; typed: Internal

In this rule, timer:at is an expression of a specific time that turns true. The
syntax is timer:at (minutes, hours, days of month, months, days of week, seconds) [23].
Event name: ea2; description: Took medicine; typed: External

Event name: ea3; description: Took medicine after some action; typed: External
 A Framework Reactive and Proactive for Pervasive Homecare Environments 333

Fig. 6. Proactive and reactive network.

Event name: a1; description: Audible warning after 10 min of ea1; typed: Internal

The events “a2”, “a3” and “a4” are modeled the same way and are not describe
here. As presented above, if event ea2 is detected, Ms. Smith is in the situation of
“medicated”. Therefore, the appPervMed is not started because the situation
“unmedicated” did not happen. The internal and external events and relevant actions
that determine each situation are presented below:

If Ms. Smith does not take the medicine (event ea1 was detected) the “unmedi-
cated” situation is initiated and appPervMed chooses the reactive action to be triggered.
After the detection of the event ea1, appPervMed waits for 10 min and, if an event ea2
was not detected, it will trigger an audible warning (a1). After the audible warning, the
application needs to wait for a feedback. If this feedback is not received, appPervMed
triggers a2 to produce some visual warning. Thus, appPervMed terminate its execution
when the resulting feedback is ea3 (i.e., took medicine after some action).

5.2 Proactive Behavior

This section presents the proactive behavior of the SIaaS, showing how it would
prevent Ms. Smith from entering in an “unmedicated” situation. Initially, it makes a
historical analysis of the situations generated by the events detected that are relevant for
appPervMed.
Below, we present a historical situation (HS) that has happened and the reactive
actions triggered for manipulating this situation.
334 A. Machado et al.

This representation is based in a notation proposed by Wasserkrug, Gal and Etzion

[26], which shows the initial and final time of the situation. In this case, curly brackets
represent the actions triggered during the period of time when the situation was valid.
For instance, the first event ea1 was detected at 10:15 (1), and the situation was
finalized when the event ea3 was detected at 10:23 and the action a1 was triggered. In
(2), Ms. Smith took the medicine on the right time (10:00), so, no reactive action was
triggered.
This historical data of behavioral management of medicines by Ms. Smith shows
the sequence of actions that were needed to handle the unwanted “unmedicated” sit-
uation. This HS is used to generate the Conditional Probability Table (CPT) for each
node event of the Bayesian network. In this case study, the generation of the CPT must
be sensible to a cognitive decline, so the Bayesian network cannot be established with
all the stored history of that situation. In order to identify a cognitive decline, the
system needs to build a valid sliding window with an event stream of Δtd (which was
described in Sect. 3.3) that corresponds to the current behavior of Ms. Smith. There-
fore, the appPervMed will register in the Module Monitor (Esper) of the Context
Manager only the patterns that correspond to the sliding window used to calculate the
network CPT represented in (7).

In the previous pattern, which was deployed in Esper, we have defined a sliding
window of Δtd corresponding to the last 45 times in which Ms. Smith had not taken her
medicine (event ea1), as well as the actions that were detected after that situation had
happened. In this pattern, if Ms. Smith should take a medication once a day, this
window would correspond to 45 days. Thus, the value of the probabilistic predictive
situation always will be set to a percentage that corresponds to 45 days. In this sense,
we avoid network scalability problems related to the excessive number of events (since
they are modeled as nodes in the network), and detect cognitive declines with periods
less than 45 days. This pattern generates the sequence of events that constantly updates
the Bayesian network and the values of the probability of the event that determines the
beginning of unwanted situations of this kind.
 A Framework Reactive and Proactive for Pervasive Homecare Environments 335

As Fig. 7 shows, event ea1 is the cause of the “unmedicated” situation. This event,
as well as events a1, a2, a3, a4, also influences event ea3. Besides, ea2 and ea3
influence the fact of Ms. Smith being medicated or not. The probability of Ms. Smith
not taking her medicine at the correct time is 63.4%, according to the simulated
behavior. Based on this data, the SIaaS will update the probability attribute of the
corresponding predictive situation for this case, as shown below:

Fig. 7. appPervMed Bayesian network.

Predictive Situation name = unmedicated;

{event} = {ea1, ea3, a1, a2, a3, a4};
{p} = ea1; pr = 63.4%; time = ea1.drug.timeTakeDrug;
The value of time is extracted by navigating through the event ea1 and following to
the entity of context drug and attribute time to take the medicine (timeTakeDrug) this
entity of context. The appPervMed thus registers the following rule (8) that shows the
relevance of the prediction value.

This rule demonstrates that the “unmedicated” situation could happen the next time
that Ms. Smith needs to take the medicine, activating the proactive decision making
process of the appPervMed.
Figure 8 presents how the pervasive application uses the Bayesian network to
identify, among the triggered actions, which one had the greatest influence for ea3 to be
detected and ending the “unmedicated” situation of Ms. Smith. It gives more relevance
to the actions that were followed by ea3 (i.e., took medicine after some action). If there
is a sequence of actions triggered after the detection of an unwanted situation (corre-
sponding by ea1), the appPervMed will chose the last action as being responsible for
the ending of the unwanted situation (i.e., ea3 detected).
Figure 7 shows that the action a3 (notify the caregiver) was the most successful
action at this moment, thus the appPervMed request for middleware trigger the action
a3, thus the polity will change from “notify caregiver after 45 min of ea1” to “notify
caregiver BEFORE 5 min of ps.time (ea1.drug.timeTakeDrug) ”. Thus, the SIaaS will
336 A. Machado et al.

Fig. 8. Proactive decision process.

warn the caregiver that he/she has to ensure that Ms. Smith will take the medicine at the
right time. In this case, the appPervMed will not be triggered to perform a reactive
action, because event ea1 will not happen. This behavior causes more events of type
ea2 to be detected since Ms. Smith starts taking her medicine at the right time, so
consequently the situation “unmedicated” will not happen and the probability of the
Bayesian Network for being medicated (ea2) increases.
We assume that there is a parameterized value with the criteria of policy selection
actions to update the triggering order of proactive actions after this rule being updated.
This will avoid that an action, after being identified as most effective, be always chosen
as a proactive action that should be executed forever. This calibration is necessary
because, in this case study, we are monitoring the behavior of a person, and this
behavior may change over time. In the example given, Ms. Smith could not respond to
the warnings of the SIaaS for some period of time because she was unmotivated with
the treatment, so the warnings to the caregiver would effectively make her taking her
medicine. However, if she did not show cognitive decline, she could return to her
self-health management without requiring the notification of the caregiver. Therefore,
there is a need for policies that trigger proactive actions to be updated. So, the SIaaS
will again generate warnings to Ms. Smith.

6 Conclusions

Most of the research efforts in situation awareness for AAL are generally related to the
detection of situations and the immediate reaction for these situations. In this sense, we
have demonstrated the necessity of mechanisms to act reactivity in order to avoid
unwanted situations in AAL.
 A Framework Reactive and Proactive for Pervasive Homecare Environments 337

In addition, we consider that for an Ambient Intelligence application to act pro-

actively it must have learning capabilities. Therefore, these applications must under-
stand and learn from the events that happened, predicting situations of interest and
making decisions in advance related to the user needs. Thus, we consider the use of a
Bayesian Network for identifying when it is necessary to act in a changed way.
In this paper, we presented an approach for enabling smart environments with
extensible, reactive and proactive characteristics, more specifically in AAL. The
extensible principles are related to an enrichment of pervasive applications, where
different situations and contexts can be managed by the middleware over the time.
We introduce reactive and proactive principals related to a situation. Reactive
actions are selected if the situation is occurring at the present moment and proactive
actions are triggered if a situation is predicted in the nearby future.
The main contributions of our approach are: (i) a method for supporting extensi-
bility in systems to Ambient Assisted Living by including experts experience while
modeling pervasive applications; (ii) an approach for handling reactive and proactive
behaviors; and (iii) a model of sliding windows for modeling time in complex event
processing.
The next steps of this research include; (i) testing the situation prediction over a real
world automated environment; (ii) improving aspects related to the prediction model;
and (iii) adapting the predictive model for taking decisions in a dynamic Bayesian
network.

References
1. Sadri, F.: Ambient intelligence: a survey. ACM Comput. 43(4), 1–66 (2011)
2. Jara, A.J., Zamora, M.A., Skarmeta, A.F.G.: An internet of things based personal device for
diabetes therapy management in ambient assisted living (AAL). Pers. Ubiquit. Comput. 15
(4), 431–440 (2011)
3. Chan, M., Esteve, D., Escriba, C., Campo, E.: A review of smart homes—present state and
future challenges. Comp. Comput. Methods Programs Biomed. 91(1), 55–81 (2008)
4. Augusto, J.C., Nakashima, H., Aghajan, H.: Ambient intelligence and smart environments: a
state of the art. In: Handbook of Ambient Intelligence and Smart Environments: Part 1.
Springer, New York, pp. 3–31 (2009)
5. Ye, J., Dobson, S., McKeever, S.: Situation identification techniques in pervasive
computing: a review. Pervasive Mobile Comput. 8(1), 36–66 (2011)
6. Dey, A., Abowd, G.: The context toolkit: aiding the development of context-enabled
applications. In: Proceedings of the SIGCHI Conference on Human Factors in Computing
Systems, Pittsburgh, Pennsylvania, US, pp. 434–441 (1999)
7. Ye, J., Stevenson, G., Dobson, S.: A top-level ontology for smart environments. Pervasive
Mobile Comput. 7(3), 359–378 (2011)
8. Kokar, M.M., Matheus, C.J., Baclawski, K.: Ontology-based situation awareness. J. Inf.
Fusion 10(1), 83–98 (2009)
9. Etzion, O., Niblett, P.: Event Processing in Action. Manning Publications Co, Stamford
(2010)
338 A. Machado et al.

10. Sixsmith, A., Meuller, S., Lull, F., Klein, M., Bierhoff, I., Delaney, S., Savage, R.:
SOPRANO – an ambient assisted living system for supporting older people at home. In:
Mokhtari, M., Khalil, I., Bauchet, J., Zhang, D., Nugent, C. (eds.) ICOST 2009. LNCS, vol.
5597, pp. 233–236. Springer, Heidelberg (2009)
11. Tazari, M.R., Furfari, F., Lazaro, R.J.P., Ferro, E.: The persona service platform for AAL
spaces. In: Nakashima, H., et al. (eds.) Handbook of Ambient Intelligence and Smart
Environments, pp. 1171–1199. Springer, Heidelberg (2010)
12. Paganelli, F., Giuli, D.: An ontology-based system for context-aware and configurable
services to support home-based continuous care. IEEE Trans. Inf Technol. Biomed. 15(2),
324–333 (2011)
13. Tennenhouse, D.L.: Proactive computing. Commun. ACM 43(5), 43–50 (2000)
14. Want, R., Pering, T., Tennenhouse, D.L.: Comparing autonomic and proactive computing.
IBM Syst. J. (IBMSJ) 42(1), 129–135 (2003)
15. Engel, Y., Etzion, O.: Towards proactive event-driven computing. In: Proceedings of the 5th
ACM International Conference on Distributed Event-based System, pp. 125–136 (2011)
16. Fu, S., Xu, C.: Quantifying event correlations for proactive failure management. Networked
Comput. Syst. J. Parallel Distrib. Comput. 70(11), 1100–1109 (2010)
17. Boytsov, A., Zaslavsky, A.: Context prediction in pervasive computing systems:
achievements and challenges. In: Supporting Real Time Decision Making, ser. Annals of
Information Systems, Springer, vol. 13, pp. 35–63 (2011)
18. Nazerfard, E., Cook, D.J.: Bayesian networks structure learning for activity prediction in
smart homes. In: 8th International Conference Onintelligent Environments (IE), pp. 50–56
(2012)
19. Lotfi, A., Langensiepen, C., Mahmoud, S.M., Akhlaghinia, M.J.: Smart homes for the
elderly dementia sufferers: identification and prediction of abnormal behavior. J. Ambient
Intell. Humanized Comput. 3(3), 205–218 (2012)
20. Machado, A., Pernas, A.M., Augustin, I., Thom, L.H., Krug, L., Palazzo, M., de Oliveira:
Situation-awareness as a key for proactive actions in ambient assisted living. In: Proceedings
of the 15th International Conference on Enterprise Information, pp. 418–426 (2013)
21. Salfner, F., Lenk, M., Malek, M.: A survey of online failure prediction methods. ACM
Comput. Surv. 42(3), 10–42 (2010)
22. Mccarthy, J.: Actions and other events in situation calculus. In: Proceedings of the 8th
International Conference on Principles of Knowledge Representation and Reasoning.
Morgan Kaufmann Publishers, Toulouse, pp. 615–628 (2002)
23. Esper: Complex Event Processing. Espertech event stream intelligence (2010)
24. Netica (2013). https://www.norsys.com/netica.html
25. Friedman-Hill, E.: Jess in Action: Java Rule-based Systems. Manning Publications
Company, Greenwich (2003)
26. Wasserkrug, S., Gal, A., Etzion, O.: A model for reasoning with uncertain rules in event
composition systems. In: Proceedings of the 21st Annual Conference on Uncertainty in
Artificial Intelligence, pp. 599–606 (2005)
Human-Computer Interaction
 Designing an IDTV Application in a Situated Scenario:
A Participatory Approach Based on Patterns

Samuel B. Buchdid ✉ , Roberto Pereira, and Maria Cecília Calani Baranauskas

( )

Institute of Computing, University of Campinas (UNICAMP), Av. Albert Einstein N1251,

Campinas, SP 13083-852, Brazil
{buchdid,rpereira,cecilia}@ic.unicamp.br

Abstract. Interactive Digital TV (iDTV) applications suffer with problems that

are inherent to the technology and their social aspects. In addition, theoretical and
practical references for designing iDTV applications are not easily found. In this
sense, designing for iDTV has become an increasingly complex activity. In this
paper, we propose three participatory practices for supporting a situated design
and evaluation of iDTV applications. The practices were articulated in a situated
design process conducted inside a Brazilian TV company. The approach favored
the participation of important stakeholders, supporting different design activities:
from design ideas to the creation and evaluation of an interactive prototype. The
results suggest the practices’ usefulness for supporting design activities, indicate
the benefits of a situated and participatory approach for iDTV applications, and
may inspire researchers and designers in other contexts.

Keywords: Socially aware computing · Organizational semiotics · Design

patterns · Participatory design · HCI · iDTV

1 Introduction

In the last years, the amount and diversity of technical devices have increased both inside
and outside people’s homes (e.g., tools, mobiles, cars, airports), being increasingly
interconnected (e.g., through bluetooth, wireless LAN, 4G) [15]. Systems are not
working in isolation, but in plural environments, bringing different people together as
citizens and members of global communities [29]. As Bannon [2] suggests, in this
scenario, there are problems that go beyond the relationship between users and tech‐
nologies, requiring more than a man-machine approach and ergonomic fixes to make
useful and meaningful design.
Therefore, designing interactive systems is becoming a more complex task, not only
in the technical sense, but also in the social one [15]. However, Winograd [37] highlights
that the majority of techniques, concepts, methods and skills to make design for a new
and complex scenario are foreign of the computer science mainstream. In this sense, it
is necessary to look at the technology comprehensively within the situated context in
which it is embedded, incorporating knowledge of several stakeholders, areas, subjects
and theories [19].

© Springer International Publishing Switzerland 2015

J. Cordeiro et al. (Eds.): ICEIS 2014, LNBIP 227, pp. 341–360, 2015.
DOI: 10.1007/978-3-319-22348-3_19
342 S.B. Buchdid et al.

Within this scenario, the emergency of the Interactive Digital TV (iDTV) (which
includes digital transmission, receiver processing capability and interactivity channel)
opens up a variety of possibilities for new services for TV [28]. However, as Bernhaupt
et al. [6] argue, with new devices connected to TV, watching it has become an increas‐
ingly complicated activity.
In fact, the iDTV has technical issues as well as social characteristics that influence
directly their use and acceptance. For instance: the interaction limited by the remote
control, the lack of custom of people to interact with television, the high amount and diver‐
sity of users, the usual presence of other viewers in the same physical space, to cite a few
[20]. As Cesar et al. [11] assert, the TV is a highly social and pervasive technology —
characteristics that make it a challenging and interesting field to investigate, but that
usually are not receiving attention from current works. Furthermore, Kunert [20] high‐
lights that every emergent technology suffers from a lack of references, processes and arti‐
facts for supporting its design. Therefore, new simple techniques and artifacts that fit the
broadcasters’ production chain and explore the challenge of designing applications within
the broadcasters’ context are welcome.
Despite not abundant, some literature has proposed ways to support the design of
iDTV applications. Chorianopoulos [12] analyzed works on media and studies about
television and everyday life, proposing design principles to support user interactivity
during leisure pursuits in domestic settings. Piccolo et al. [27] proposed recommenda‐
tions to help designers with accessibility issues for iDTV applications. Kunert [20]
proposed a collection of pattern for the iDTV focused in usability issues. Solano et al.
[30] presented a set of guidelines that should be considered in iDTV applications for
preventing frequent usability problems.
Focused on the users’ aspects, Rice and Alm [28] proposed methodologies and
interactive practices influenced by the Participatory Design (PD) to design solutions for
supporting elderly people to interact with iDTV. Bernhaupt et al. [6], in turn, used the
Cultural Probes Method to conduct ethnographic studies in order to understand users’
media behavior and expectations, indicating trends concerned with personalization,
privacy, security and communication.
Focusing on the broadcaster company’s aspects, some works have adapted tradi‐
tional methodologies for software development [18] and Agile Methods [35] to the
companies’ production chain. The adapted methodologies encompass the entire soft‐
ware development process (e.g., requirement analysis, project, implementation, testing
and support); although robust in terms of the technical process of software development,
end users are usually not considered in the process.
Shedding light on this scenario, we draw on Socially Aware Computing (SAC) [3, 5],
Organizational Semiotics theory [22], Participatory Design [24], and Design Patterns for
iDTV applications [20] to propose three situated and participatory practices for supporting
designers to create and evaluate iDTV applications: (i) the Participatory Pattern Cards; (ii)
the Pattern-guided Braindrawing; and (iii) the Participatory Situated Evaluation.
In this paper, we present the three practices and the theories underlying them, and
discuss the results obtained from their usage in the practical context of a Brazilian
broadcasting company. The practices were planned to facilitate the participation of
professionals from the TV domain that are not familiar with iDTV applications design.
 Designing an IDTV Application in a Situated Scenario 343

A group of 9 persons, with different profiles, participated in design workshops for

creating an iDTV application for one of the company’s programs. The results suggest
both the practices’ usefulness for supporting design activities and the benefits of situated
and participatory design for iDTV applications, indicating the viability of conducting
the practices in industrial settings.
The paper is organized as follows: the Sect. 2 introduces the theories and method‐
ologies that ground our work. Section 3 describes the new practices created for
supporting a situated and participatory design of iDTV applications. Section 4 presents
the case study in which the techniques were applied, and Sect. 5 presents and discusses
the findings from the case study analysis. Finally, Sect. 6 presents our final considera‐
tions and directions for future research.

1.1 Software Development Vs TV Content Processes

TV Companies have teams with well-defined roles (e.g., director, designer, producer
and engineer) and tasks (e.g., pre-production, production and transmission) that work
in a synchronized way to produce the television content [7]. For example, while the
production teams produce television content, engineers provide infrastructure for broad‐
casting the TV content to viewers. A software development process has also well-defined
steps and roles. Pressman [16] argues that the main steps depend on different process
models (e.g., Cascade Model, Agile Methods) and should encompass from requirements
analysis and design to development and testing of the system. These steps are supported
by designers, developers, testers, and analysts among other professionals.
With iDTV and the need to produce interactive content, the production process of
television content gains new roles and new needs toward the software production.
However, the broadcasting companies are not ready for this new component into their
production chains. In addition, Veiga [35] argues that designing iDTV applications is
hardly supported by existing methodologies (e.g., Cascade Model) because it is different
from designing traditional software systems (e.g., desktop, web). There is a distance
between a production process of television content and conventional software develop‐
ment process. In this sense, the application of iDTV must be created on the optics of the
two process models, and the opportunity arises to develop a new model specific process
model for developing iDTV applications, from inspirations in both process models. In
this work we draw on the SAC approach that seeks to understand the needs of the
production chain from TV Company, and brings theoretical references from software
engineering, but also seeks to bring the end user’s needs for the artifact to be designed.

2 Theoretical and Methodological Foundation

Organizational Semiotics (OS) and Participatory Design (PD) are two disciplines which
represent the philosophical basis for the design approach considered in this work. Design
patterns for iDTV add to this theoretical basis contributing to shaping the design product.
OS proposes a comprehensive study of organizations at different levels of formali‐
zation (informal, formal, and technical), and their interdependencies. OS understands
344 S.B. Buchdid et al.

that all organized behavior is effected through the communication and interpretation of
signs by people, individually and in groups [22, 32]. In this sense, the OS supports the
understanding of the context in which the technical system is/will be inserted and the
main forces that direct or indirectly act on it. If an information system is to be built for
an organization, the understanding of organizational functions from the informal to the
technical level is essential [22].
The PD, originated in the 70’s in Norway, had the goal of giving to workers the rights
to participate in design decisions regarding the use of new technologies in the workplace
[24]. In this sense, PD proposes conditions for user participation during the design
process of software systems. PD makes use of simple practices that use fewer resources
(e.g., pen and paper), and considers that everyone involved in a design situation is
capable of contributing, regardless of his/her role, hierarchical level, and socio-
economic conditions. Two examples of participatory practices are Brainwriting [34] and
Braindrawing [24]. Both practices are examples of cyclical brainstorming conducted to
generate ideas and perspectives from various participants for the system to be built.
While Brainwriting was created to generate ideas for system features, Braindrawing was
proposed for generating graphical ideas for the User Interface (UI).
Drawing on OS and PD, the Socially Aware Computing (SAC) proposes to under‐
stand the design cycle by working on the informal, formal and technical issues in a
systematic way; moreover, it recognizes the value of participatory practices to under‐
stand the situated character of design.

2.1 Socially Aware Computing

The Socially Aware Computing (SAC) is a socially motivated approach to design [3]
that supports the understanding of the organization, the solution to be designed, and the
context in which the solution will be inserted, so that it can effectively meet the socio‐
technical needs of a particular group or organization.
Considering the Semiotic Onion (see “SAC’s Meta-Model” detail in Fig. 1), SAC
understands design as a process that must go through the informal, formal and technical
layers cyclically — see the dashed cycle. According to Baranauskas [5], the design
process should be understood as a movement that starts in the society (outside of the
semiotic onion) and progresses through the informal and formal layers in order to build
the technical system. Once (an increment of) the technical system is projected, the
movement returns impacting on formal and informal layers alike, including the people
for whom the system was designed, the environment in which it is/will be inserted, and
the society in general. SAC is an iterative and incremental process. Therefore, each
iteration favors the problem clarification, knowledge-building, and the design and eval‐
uation of the proposed solution.
For understanding the organization’s situational context and the system inside it,
SAC uses concepts and techniques inspired by PD and OS. More than the end user, SAC
considers and involves key stakeholders and heterogeneous groups of people who may
influence and/or may be influenced by the problem being discussed and/or the solution
to be designed.
The practices conducted in SAC are held throughout the design process within
Semio-participatory Workshops (SpW). According to Baranauskas [4], each SpW has
 Designing an IDTV Application in a Situated Scenario 345

Fig. 1. Design process.

well-defined goals and rules within the design process, such as: (i) socialization and
personal introductions of the participants. (ii) explanations about the SpW to be
conducted, its concepts and objectives. (iii) the role of the SpW in a whole design process
(in the cases where there are more than one SpW to be conducted). (iv) a well-defined
schedule for activities. (v) artifacts and methods created/adapted to be articulated with
the practices, and so on.
SAC has been used to support design in several different contexts, being applied in
design scenarios of high diversity of users (e.g., skills, knowledge, age, gender, special
needs, literacy, intentions, values, beliefs) and to create different design products in both
academic and industrial environments [25]. Specifically for the iDTV context, SAC has
being used to support the consideration of stakeholders’ values and culture during the
design process [26], for proposing requirements and recommendations to iDTV appli‐
cations [27], and to physical interaction devices [23].

2.2 Design Patterns for IDTV

Design patterns were originally proposed to capture the essence of successful solutions
to recurring problems of architectural projects in a given context [1]. In addition to their
use in the original field of architecture, design patterns have been used in other fields,
such as Software Engineering [17] and Human-Computer Interaction (HCI) [8], and
within different contexts, such as Ubiquitous Computing [13] and iDTV [20].
For new technologies, Kunert [20] and Chung et al. [13] argue that design patterns
present advantages: (i) they are distributed within a hierarchical structure, which makes
it easier to locate and differentiate between patterns of different granularity; (ii) they are
proposed in a simple language; and (iii) they incorporate references that may indicate
other forms of design guidance.
346 S.B. Buchdid et al.

In the iDTV field, few studies proposing HCI patterns are found in literature. For
instance, Sousa et al. [31] identified a list of usability patterns for specific interactive
iDTV tasks, and Kunert [20] proposed a pattern collection that focuses on interaction
design for iDTV applications, paying special attention to usability issues.
The pattern collection used in this work is the one proposed by Kunert [20]. The
patterns are divided into 10 groups: Group A: Page Layout — Defines the layout types to
be used in the application; Group B: Navigation — Defines what types of navigation are
to be used in the application; Group C: Remote Control Keys — Defines the main keys
of the remote control; Group D: Basic Functions — Highlights the basic functions that
should be considered in the design of interaction; Group E: Content Presentation —
Determines the basic elements that form an application; Group F: User Participation —
Describes the interaction of specific tasks; and the way how the approval for connectivity
should be handled; Group G: Text Input — Defines the multiple ways to input text, when
to use each, and how to use them in an application; Group H: Help — Defines the types
of help and how to provide them for users in an appropriate way, according to the context
of use; Group I: Accessibility & Personalization — Deals with accessibility and person‐
alization issues; and Group J: Specific User Groups — Illustrates patterns for specific user
groups (e.g., children). Each of the 10 groups describes and illustrates first-level problems
that are divided into new design problems of second and third levels. On the second level,
there are 35 interaction problems; for each one, there is a corresponding pattern.
There is not a strict order when choosing patterns, however, Kunert [20] suggests
choosing the layout and navigation patterns before the other patterns, because this initial
decision directly influences the remaining ones.

3 The Proposed Participatory Practices

Drawing on the design patterns and the participatory design techniques, we proposed
three practices for supporting design activities in a situated context: (i) Participatory
Pattern Cards; (ii) Pattern-guided Braindrawing; and (iii) Participatory Situated Evalu‐
ation. These practices were articulated with other design activities in an instantiation of
Baranauskas’ SAC design process [3, 5] in order to favor the situated and participatory
design of iDTV applications — see Fig. 1.
The “A” detail in Fig. 1 suggests that the problem domain must be clarified and a
solution proposal must be discussed in a participatory way before engaging in further
design activities. When the problem is clarified and a solution is proposed, three partic‐
ipatory practices (“1”, “2” and “3” details) support the production of the first version of
the prototype (“B” detail); one participatory practice supports the inspection of the
designed prototype (“4” detail), and one extra evaluation may be conducted with
prospective end-users (“5” detail). These activities contribute to build and evaluate a
prototype for the application, offering useful information for further iterations of the
process (e.g., the codification stage, the design of new functionalities, redesign).
The Participatory Pattern Cards (PPC) (“1” detail in Fig. 1) was conceived to
support discussions about design patterns for the iDTV, and the identification and selec‐
tion of the patterns suitable for the application being designed. For this practice, we
 Designing an IDTV Application in a Situated Scenario 347

created 34 cards based on Kunert’s Design Patterns [20] for the iDTV. Table 1 presents
a description for the practice.

Table 1. Description of the PPC practice.

Participatory Pattern Cards (PPC)

1. A set of 34 cards representing Kunert’s collection of patterns [20]: the cards are organized in 5
Materials
(input)

predefined groups (e.g., patterns for the application’s layout; patterns for the text input mode);
2. All the material produced in previous activities (e.g., a brief description of the design problem,
a general description of a solution proposal, a list of requirements).
1. Cards overview: participants are introduced to the Pattern Cards, their different types and
usage examples;
Methodology

2. Selection of patterns: for each card group, participants should individually select the cards that
would potentially be used in the application.
3. Consensus: a brainstorming section where the participants present the selected patterns and
discuss the pros and cons of each one in order to decide the ones they will adopt;
4. Justification for the choices: once a consensus was reached, participants must justify their
choices based on the project’s scope and requirements.
1. A subset of patterns that will potentially be used for the application.
(output)
Results

As byproducts, the practice: i) brings participants closer to the iDTV domain; ii) draws attention
to the limited resources and technology that will be provided for the system to be designed; and
iii) may inspire design ideas for future projects.

Figure 2 illustrates an example of a Pattern Card created for the practice. Each card
has the following sections: (i) group, reference and name of the pattern, (ii) an example
of the pattern being used in a given situation; (iii) a brief description of the problem;
(iv) forces (advantages and disadvantages) that act directly and indirectly on the problem
to be solved; and (v) the solution to the problem.

Fig. 2. Example of a pattern card created from Kunert’s collection [20] of patterns.

The PPC practice is useful to clarify the constraints and potentials of iDTV tech‐
nology and to choose design patterns in a participatory way, contributing to the construc‐
tion of a shared knowledge among the participants.
The Brainwriting (“2” detail in Fig. 1) is a silent and written generation of ideas by
a group in which participants are asked to write ideas on a paper sheet during a
348 S.B. Buchdid et al.

pre-defined time (e.g., 60 s). Once this time was elapsed, each participant gives his/her
paper sheet with ideas to other participant and receives another paper sheet to continue
the ideas written on it. This process is repeated several times until a predefined criterion
is satisfied — e.g., the fixed time has run out; each paper sheet passed by all the partic‐
ipants [36]. On the one hand, Brainwritting is a good method for producing different
ideas in a parallel way, allowing the participation of all without inhibition from other
participants. On the other hand, it focuses on the question/problem being discussed rather
than on the person discussing it [34], avoiding conflicts between the participants.
The Pattern-guided Braindrawing (PgB) (“3” detail in Fig. 1) is an adapted version
of Braindrawing that aims to generate ideas for the UI of the application being designed,
taking into account the Design Patterns for iDTV. Table 2 presents a description for the
technique.

Table 2. Description of the PgB practice.

Pattern-guided Braindrawing (PgB)

Materials
(input)

Paper sheets for drawing, colored pens, chronometer.

1. Situating: participants are arranged in a circle; the design problem and the results from the
previous activities (e.g., requirements, PPC) are briefly reviewed;
2. Generation of design elements: keeping visible the design patterns selected in the PPC and a
list of requirements for the application, participants start drawing the application’s interface
on a paper sheet. After a pre-defined time (e.g., 60 seconds), participants stop drawing, move
Methodology

the paper sheet to the colleague seated on their right side, and receive a paper sheet from a
colleague seated on their left side, continuing to draw on the received paper sheet. This step
repeats until all participants contributed with ideas to all the paper sheets at last once, i.e., a
complete cycle;
3. Synthesis of design elements: From their own paper sheets (the ones the participants initiated
the drawing), participants highlight the design elements that appeared in their draws and that
they find relevant for the application.
4. Consensus: Based on the highlighted design elements from each paper sheet, the group
synthesizes the ideas and consolidates a final proposal that may include elements from all the
participants;
1. Different UI proposals that were created in the participatory activity: each proposal presents
(output)
Results

elements drawn by different participants, differing from each other because they were started
by a different person;
2. A collaborative proposal for the application’s UI, guided by design patterns, and created
from the consolidation of the different proposals by the participants.

The PgB allies the benefits from PD and Design Patterns, being useful to materialize
ideas and proposals produced in the previous steps into prototypes for the application.
Therefore, while the participatory nature of both PPC and PgB techniques motivate
participants to generate design ideas that rely on the perspectives of different stake‐
holders, the use of Design Patterns informs these ideas and guides their materialization.
A picture of a television device and a screenshot of the TV program may be used as
background of the paper sheets used in PgB. This contributes to bring reality to the
participants during the activity, situating them according to the device’s physical limi‐
tations, the program layout and content. Example of this template is reported in Buchdid
et al. [10].
 Designing an IDTV Application in a Situated Scenario 349

The third practice created was the Participatory Situated Evaluation (PSE) (“4”
detail in Fig. 1). The PSE is an adapted version of Thinking Aloud method [21] that
aims to bring together all participants for the evaluation of an iterative application —
Table 3 presents a description for the practice. This practice is useful to promote a
collective analysis and discussion about the produced prototype; to identify shared
doubts and difficulties, as well as ideas for improving the application. It avoids the
prevalence of individual opinions, favoring the collective discussion and making sense
about the application being evaluated, and optimizing the time spent by the participants
during the activity.

Table 3. Description of the PSE practice.

Participatory Situated Evaluation (PSE)

Materials
(input)

Laptop, interactive prototype, video camera, and software to record users interacting with the
prototypes.

1. Situating: participants are arranged in a circle; the interactive prototype is introduced to the
participants and the evaluation activity is explained; participants can either conduct pre-defined
tasks (e.g., voting in a pool) or explore the application in a free way;
Methodology

2. Interacting with the prototype: a participant is invited to interact with the prototype; using the
Thinking Aloud method [21], the participant speaks aloud for the group while interacts with
the prototype, reporting his/her thoughts (e.g., general impressions about the prototype, inten-
tions, goals, difficulties, questions, reasoning). The other participants can talk to each other and
to the person who is interacting with the prototype, speaking their thoughts alike.
3. Consensus: based on the doubts, ideas, feelings and difficulties found during the activity, the
participants elaborate a list of problems and suggestions for improving the application.
(output)
Results

1. A mapping of the interaction and interface problems identified through the activity;
2. Suggestions of improvements presented in the group’s suggestion list.

User Evaluation (“5” detail in Fig. 1) proposal: the Thinking Aloud technique [21] can
be used to capture users’ impressions and opinions. The participants’ interaction, voices
and facial expressions can be recorded, and participants may be invited to answer an
evaluation questionnaire, providing their overall impressions about the prototype. The
activity and data usage should be conducted in accordance to ethical principles in
academic research.

4 The Case Study

The case study was conducted in a real context of a television broadcasting company,
named EPTV (Portuguese acronym for “Pioneer Broadcasting Television Stations”). EPTV
is affiliate of a large Brazilian broadcasting company. Currently, EPTV programming
reaches more than 10 million citizens living in a microregion of about 300 cities [14].
“Terra da Gente” (TdG, “Our Land”, in English) is one of several programs produced
by EPTV. The program explores local diversity in flora and fauna, cooking, traditional
music, and sport fishing. Currently, the program runs weekly and is structured in 4
blocks of 8 to 10 min each. It counts on a team of editors, writers, producers, designers,
350 S.B. Buchdid et al.

technicians, engineers and journalists, among other staff members. In addition to the
television program, the TdG team also produces a printed magazine and maintains a
web portal. Both the magazine and the web portal serve as complementary sources of
material for the TdG audience [33].
The activities reported in this paper were conducted from January to July, 2013, and
involved 3 researchers from Computer Science and 6 participants playing different roles
at EPTV:
• TdG Chief Editor: is the person who coordinates the production team (e.g., editors,
content producers and journalists) of the television program and the web portal.
• Designer: is the responsible for the graphic art of the TV program as well as of the
web portal, and who will be responsible for the graphic art of the iDTV application.
• Operational and Technological Development Manager: is the person who coor‐
dinates the department of new technologies for content production.
• Supervisor of Development and Projects: is the person who coordinates the staff
in the identification and implementation of new technologies for content production
and transmission.
• Engineer on Technological and Operational Development: is the engineer of
infrastructure, and content production and distribution.
• Technical on Technological and Operational Development: is the person respon‐
sible for the implementation, support and maintenance of production systems and
content distribution.
• Researchers (3 people): are researchers in Human-Computer Interaction and the
responsible for preparing and conducting the workshops. One of them is expert in
the SAC approach and other is an expert in iDTV technologies.
All the participants, except for the researchers, work in the television industry. The
participants (P1, P2…P9) collaborated in the workshops proposed to the problem clar‐
ification, problem solving, requirement prospecting, as well as the creation of prototypes
for the application and their evaluation, within a SAC approach.
Regarding the familiarity of participants with iDTV applications, from the 9 partic‐
ipants, 2 are experts; 2 are users of applications; 5 participants had already used/seen
iDTV applications. Regarding the frequency which the participants watch the TdG
program, 5 participants have been watching the TdG program, but not very often: 1
participant watches the program every week, 1 participant watches the program in
average twice a month, and 2 participants watch at least once a month.

4.1 Designing an Application for TdG

This section presents the main activities conducted to create the first prototype of an
iDTV application for the TdG program. Before these activities, participants had colla‐
borated for the problem understanding, and for the clarification, analysis and organiza‐
tion of requirements for the application to be designed — as proposed by the SAC
approach, and that are out of scope of this paper (“A” detail in Fig. 1). The materials
produced by the previous activities were used as input for the design activities presented
in this paper, and were reported in Buchdid et al. [9].
 Designing an IDTV Application in a Situated Scenario 351

Before the beginning of each activity, the results obtained from the previous activities
were presented and discussed briefly, and the techniques to be used, as well as their
methodologies and purposes were introduced to the participants. For instance, before
the PPC activity, examples of different existing iDTV applications, and the patterns from
Kunert [20], were briefly presented and discussed with the participants.
The PPC activity was the first participatory practice conducted to design the appli‐
cation prototype (“1” detail in Fig. 1). Its input were the documentation produced in the
problem clarification activities, the participant’s knowledge about the project, and
Pattern Cards based on the Kunert’s patterns [20].
Originally classified into 10 different categories (from “A” to “J”), the patterns were
grouped into 5 major groups in order to facilitate the participants’ understanding: 1.
Layout (Group A); 2. Navigation (Group B); 3. Operation (Groups C, D and G); 4.
Content presentation (Groups E and F); and 5. Help, accessibility and personalization
(Groups H and I). Patterns such as “B3 Video Multi-Screen” and “J1 Children” were
not considered because they were out of the projects’ scope.
The dynamic for this practice followed the description presented in Table 1. While
each group of pattern was presented and discussed, participants were asked to select the
ones that would potentially be used in the application — see “A” detail in Fig. 3. This
practice lasted 90 min and was important to generate discussion and ideas to the appli‐
cation; they also led to a shared knowledge about iDTV potentialities and limitations
among the participants.

Fig. 3. Participatory activities in a situated context.

Guided by the discussions and the results identified in the PPC practice, the Brain‐
writing (“2” detail in Fig. 1) was used to identify what the participants wanted in the
application and what they thought the application should have/be. The dynamic for this
activity is similar to the PgB presented in Table 2: each participant received a paper sheet
with the following sentence: “I would like that the “Terra da Gente” application had…”;
the participants should write their initial ideas and, after a pre-defined time (e.g., 60 s), they
should exchange the paper sheets and continue to write on the ideas initiated by the other
participants. After each paper had passed by all the participants and returned to the one who
started writing the idea, participants should highlight the concepts that appeared in their
352 S.B. Buchdid et al.

paper sheet, and expose them to the group for discussion. The group reached a consensus
creating a list of the main features that should appear in application — see “B” detail in
Fig. 3. This activity took 90 min.
The PgB practice was conducted based on the ideas generated during the Brain‐
writing and took into account the patterns selected in the PPC (see “3” detail in Fig. 1).
The dynamic for this activity is presented in Table 2: each participant received a template
in a paper sheet, and they were asked to explore the initial call for the application, the
layout and other specific content that they would like to see in the application. Partici‐
pants started drawing the application interface, exchanging their paper sheets periodi‐
cally and continuing to draw on the paper sheets of the other participants until they
received their paper sheet back — see “C” detail in Fig. 3. This activity generated several
ideas for the iDTV application that were consolidated by the team in a final proposal.
This activity lasted 30 min.
Based on the results obtained from these activities, the first prototype for the appli‐
cation was built (“B” detail in Fig. 1) by a researcher who has experience in the devel‐
opment of iDTV applications. The Balsamiq® tool was used to create the UI and the
CogTool® was used to model the tasks and to create an interactive prototype. The Pattern
Cards were used again in order to inspect whether the application was in accordance
with the design patterns, guiding the layout definition (e.g., font, elements size and
position, visual arrangement of these elements) and interaction mechanisms (e.g.,
remote control’s keys that were used).
The PSE was conducted in order to evaluate the produced prototype — “4” detail
in Fig. 1. The activity was conducted according to the structure presented in Table 3.
The interactive prototype was presented to the participants, and one of them explored
the application using the “Thinking Aloud” technique — see “D” detail in Fig. 3. The
other participants observed the interaction, took notes, and were able to ask, suggest and
discuss with the evaluator at any time. Both the user interaction and the group dynamic
were recorded, providing interesting information about the general perception of the
participants and possible features to be redesigned before programming the final appli‐
cation. This practice lasted 50 min and, after concluded, participants answered a ques‐
tionnaire evaluating the prototype.
Finally, a User Evaluation was conducted in order to evaluate the prototype with
prospective representatives from the target audience that did not participate in design
activities — “5” detail in Fig. 1. This activity was important to serve as a parameter to
the PSE evaluation, assessing whether the prototype made sense to a more diverse audi‐
ence. For this activity, 10 participants explored the prototype: 3 participants are 21–30
years old, 5 are 31–40 years old, 1 is 41–50 years old, and 1 participant is over 60 years
old. Regarding their formal education, from the 10 participants: 1 has high school, 3
have bachelor’s degree, 2 has specialization course, 3 have master’s degree and 1 partic‐
ipant has a doctor’s degree. None participant had previous experience using iDTV
applications; 8 participants were aware of them, but had never seen any application; and
2 participants had seen them before. Furthermore, from the 10 participants, 6 have been
watching the TdG program, but not often; 3 participants watch once a month; and 1
participant do not watch TdG.
 Designing an IDTV Application in a Situated Scenario 353

5 Results and Discussion

In this section, we present and discuss the main results from the practices we proposed
in this paper to create the interactive prototype for the TdG TV program.

5.1 Results of Design Practices

During the PPC practice, the participants selected 20 patterns that could be used in the
application design. At least one pattern from each group of patterns was considered by
the participants. Table 4 presents some of the patterns selected by the participants. The
“Groups” column presents the general group of the selected pattern; the “Patterns”
column presents the name of the pattern; the “PPC” column indicates whether the pattern
was selected during the PPC practice; the “Explanation” column explains the reason
why the pattern was selected; and the “PgB” column indicated whether the pattern was
identified in the prototype produced in the Brain-Drawing practice.

Table 4. List of Patterns used in the activities.

Groups Patterns PPC Explanation PgB

C3 Ok-key It must be the main method of interaction together with arrow keys
C4 Colour keys To be used in case of voting and multiple-choice question
C5 Number keys Would not be used due to the difficulty of use
Operation

C6 Special keys Hard to find on remote control

An unobtrusive call that does not disturb who does not want to use
D1 Initial call to action
the application
…

Must not occupy much space on the screen. It will only be used in
G3 Mobile phone keyboard
case of text input
H1 On-Screen instruction It is not necessary because the application is simple
Help and

H2 Help section Help only in the Option menu

cia

I1 Accessibility Universal Design

I2 Personalisation It is very sophisticated to this kind of application

For instance, the pattern “C3 Ok-key” was selected to be “the main interaction
method together with arrow keys” in the PPC practice, and was identified in the prototype
produced in the PgB. The pattern “C6 Special keys”, in turn, was not selected in the
PPC, but appeared in the prototype created by the participants: It can be partially
explained by the fact that the participants got more used to the patterns and may have
perceived the need/benefits of patterns they did not select during the PPC. Therefore,
this is both an indication that the PPC does not narrow the participants’ views during
the creation of prototypes, and an evidence that the PgB facilitates the revision of the
selected patterns during the creation of prototypes.
From the Brainwritting practice, 11 concepts were created to be included in the
application: 1. Gallery/Making of: pictures from the TV program and information about
the backstage; 2. Localization/Mapp: geographic coordinates of the place in which the
TV program was recorded; additional information about roads, flights, trains, etc.
354 S.B. Buchdid et al.

3. Recipe/Ingredients: it presents the ingredients of the recipe that will be prepared

during the TV program. 4. Information/Curiosity: offers information and curiosities
about the fauna and flora existing in the place where the TV program was recorded. 5.
Evaluation Pool: a pool that allows users to answer whether they liked the program
they are watching. 6. Quiz: a question-answer based-game about subjects directly
related to the TV program content. 7. Fishing Game: a ludic game intended to keep
users’ attention through a virtual fishing while they watch the TV program (e.g., a little
fish will appear on the screen and the user must select a different key to fish it).
8. Fisherman Story: a specific Quiz that allow users to answer whether a given story
is true or false. 9. Abstract: a summary of the current TV program. 10. Prospection
Pool: a pool that allows users to vote in the subjected that will be presented in the next
program. 11. Chat: asynchronous communication on the TV program.
The first 6 concepts were selected to be used in the PgB activity. In addition, the
participants were invited to explore ideas to application’s trigger (Pattern: “DI Initial
Call to Action”) in the same activity. The other concepts were not considered because
they were similar to a selected concept (e.g., Fisherman Story is similar to the Quiz),
because they were considered uninteresting (e.g., Summary), or because they would
require high attention and cognitive effort to be used (e.g., Chat).
All the six selected concepts appeared in the individual prototypes created by the
participants of the PgB practice as well as in the final prototype consolidated by the
participants. For instance, the “Gallery/Making of” concept appeared in 7 individual
prototypes (see the column “Frequency” in Table 5), and was represented in 4 different
forms (column “Forms”). The 9 individual prototypes also represented the “Localiza‐
tion/Mapp” concept in 4 different forms. Furthermore, the “Fishing Game” appeared 3
times even not being one of the chosen concepts; indicating that the activity favored the
appearance of different and diverse ideas.

Table 5. List of concepts represented in the individual prototypes.

Concept Frequency Forms

Gallery/Making Of 7 4
Localization/Mapp 9 4
Recipe/Ingredients 7 4
Information/Curiosities 7 4
Evaluation Pool 5 3
Quiz 5 3
Application’s Trigger 6 6
Fishing Game 3 3

The individual prototypes generated in the PgB were consolidated into a final proto‐
type that, in turn, was used as the basis for creating an interactive prototype for the TdG
iDTV application. The six concepts cited previously, as well as the patterns presented
in Table 4, and general ideas elaborated by the participants were reflected in the inter‐
active prototype. In fact, the design patterns selected in the PPC practice were reflected
in both the final prototype produced by the participants and the interactive prototype
created by the researcher. For instance, the patterns “C4 Colour keys” and “H2 Help
 Designing an IDTV Application in a Situated Scenario 355

section” were selected in the PPC activity and were considered in the individual proto‐
types — see Table 4, and were also considered in the interactive prototype — see details
“C4” and “H2” in Fig. 4.

Fig. 4. Patterns highlighted on the mockups from the PgB and on the final prototype.

5.2 Evaluation
Results of the PSE Practice. During the PSE we identified that users could leave the
application at any moment/any level of interaction; however, the evaluation indicated
that it could cause interaction problems, such as the user accidentally leaving the appli‐
cation while trying to see a picture from the backstage. The participants recommended
disabling the “Exit” functions when the user enters in a second level menu/function.
Furthermore, the “Help” function also should be applied only to the general application
(not in specific sub-menus), because the application is very easy to use and the button
could disturb the user in specific activities.
Other useful feedbacks were obtained from the PSE practice, such as the suggestion
to use numbers in the pool’s options in order to facilitate the selection, and not confuse
users with other application’s functions that use colors key; and the recommendation to
not deploy the “Quiz” and the “Pool” features simultaneously in the application in order
to not overload users with similar features.
The participant who explored the interactive prototype in the PSE practice was clearly
pleased for not having difficulties while using it, highlighting the simplicity and consis‐
tency of the interactive prototype. Using his words: “(…) if even me was able to under‐
stand and use the prototype, then it means the prototype is very intuitive.” [laughs] — he
had never used an iDTV application before.
The participants’ responses to the evaluation questionnaire also indicated a positive
opinion about the interactive prototype. From the 9 participants who answered the ques‐
tionnaire, 7 (78 %) responded they really liked the prototype, and 2 (22 %) answered
that they liked moderately. No indifferent or negative response was provided, indicating
that the prototype met the participants’ expectations
356 S.B. Buchdid et al.

Results of the Test with Prospective Users. The test with representatives from the audi‐
ence reinforced a favorable opinion about the interactive prototype. The 10 prospective
users were able to understand and explore the prototype, indicating its simplicity. From
their responses to the evaluation questionnaire, 5 users (50 %) answered they really liked
the prototype, 4 users (40 %) answered they liked moderately, and 1 users (10 %)
answered with indifference. Although users reported the application was intuitive and
easy to use, they had initial difficulties to interact with the prototype, in parts because
of their lack of experience with iDTV applications. After a few interactions, users got
more comfortable with the application and started to explore it, visiting its different
features. However, when some users felt lost, they tried to change the channel, turn off
the TV device, tried many remote key options or visited parts of application randomly.
The representatives liked the application because they would be able to access rele‐
vant content (curiosities, recipes, game fishing and exclusive content); to take notes of
information that is difficult to annotate during the TV show (e.g., recipe and directions);
to view part of the content they missed in the TV Show; iv) to learn more about a new
subject (e.g., curiosities). However, to continue accessing the iDTV application, users
answered that the application should be updated for each new program (e.g., extra curi‐
osities, new games, and different recipes).
Although we need to test the application with a higher number of users in order to
have data with statistical relevance, obtaining 90 % of positive responses is a good
indication given that they did not participate in design activities and had no prior contact
with iDTV applications.

Preliminary Evaluation of the Design Process. The process has shown strong aspects
in different situations. One of them, participants with different roles in EPTV showed
themselves comfortable at all times of the process to expose their ideas and desires.
While some participants want more interactivity and functionality, other participants,
with responsibilities on the product itself (the TV program content), wanted an aesthet‐
ically simple application that does not influence on the television content. Accordingly,
when conflicts emerged from the participants, the participatory activities guided by the
Organizational Semiotics artifacts were important for achieving a cohesive and uniform
decision from the whole group.
Other point that deserves consideration in this project is its situated context. The
conduction of participatory practices in a situated context contributed to understand
different forces related to the project and the organization in which it was being
conducted. In each new practice, it was possible to clarify tensions between the partic‐
ipants, the context in which the EPTV operates, the high importance of the TdG
program for EPTV organization, the relation between the affiliate and its headquarter
and, mainly, the role that the application might play in the TV program. Participants
have different views and understandings regarding the competition (for the user atten‐
tion) between the interactive application and the TV program, and different opinions
about what the application should offer to users and the way it should be offered. Such
complex context would be difficult to capture and understand in a non-situated design
approach, and such conflicts would be hard to deal with if participatory practices were
not part of the methodology.
 Designing an IDTV Application in a Situated Scenario 357

Furthermore, the four workshops conducted at EPTV took about 12 h. It means that
all the process, from the problem clarification to the prototype evaluation, took them
less than two days of work. It is clear that a great effort from the researchers was needed
in order to summarize, analyze and prepare the practices as well as to prepare the
presentations and build the interactive prototype. Indeed, this effort is expected because
a lot of work must be done in parallel to the practices organization and conduction.
Therefore, this experience shows that it is possible, viable and worth the time used to
make participatory design in a situated context.

5.3 Discussion
During the participatory practices, the constructive nature of the process allowed to see
how different viewpoints were conciliated, different proposals were consolidated, a
shared understanding about the problem domain and the application was created, and
how the discussions were materialized into a solution proposal. Ideas and concepts that
were discussed when the project started could be perceived during the practices and
were reflected in the final prototype.
The interactive prototype reflected the results from both PPC and PgB practices,
allowing the participants to interact with the prototype of the application they co-created.
The examples of existing applications presented to the participants were useful to illus‐
trate different solutions regarding the patterns, inspiring the design of the new applica‐
tion and avoiding design decisions that would not satisfy them. The PPC practice was
especially useful to: (i) present the constraints, limitations and challenges of designing
for iDTV; and (ii) introduce participants to design patterns for iDTV, which may support
their design decisions.
The PgB, in turn, was useful for supporting a pattern-guided construction of UI
proposals for the application from the material produced in the previous activities. This
practice is especially important because it favored the consideration of Design Patterns
in the prototype design, and because it allowed all the participants to expose their ideas
and to influence the prototype being designed, avoiding the dominance of a single view‐
point. For instance, the “Pool” and the “Quiz” were concepts that emerged from the
Brainwritting and were materialized during the PgB practice, but were strongly
discussed among the participants because some of them did not approve these features.
However, after listening pros and cons of keeping/removing these concepts from the
project’s scope, the participants decided to keep both concepts in the final prototype.
Regarding the prototype evaluation, the PSE was important to foment discussions
on the design decisions. Furthermore, the feedback from prospective users was impor‐
tant to verify decisions made with outsiders: people who did not participate in the design
process (e.g., how to present the recipe: only the ingredients should be included? The
preparation mode should also be displayed?).
The practices reported in this paper demonstrate that it is possible to conduct situated
and participatory design in industrial settings. There is usually a myth that these practices
are expensive and difficult to be conducted. In fact, in less than 4 h a prototype was built
from the documentation produced in the previous practices and from the discussion
between the participants — including the time spent to present examples of existing
358 S.B. Buchdid et al.

applications and the lecture for presenting the patterns. Some of the participants had a
vague idea about how to design iDTV applications and none of them had designed this
kind of application before.
The experience at EPTV also indicated that a situated and participatory design
contributes to the development of solutions that are in accordance to both the people
directly involved in design practices and the prospective end users of the designed solu‐
tion. On the one hand, the participatory evaluation indicated that the participants
approved the interactive prototype they co-designed; it was expected because of the
participatory and situated nature of the process conducted. On the other hand, the eval‐
uation with representatives from the target audience reinforced the positive results,
indicating that the application was understood and well accepted by users that were not
present in design activities and that had never experienced an iDTV application before.
These results suggest that a situated and participatory design perspective favors the
ideas of solutions that make sense to the interested parties, reflecting an understanding
about the problem domain and its complex social context.

6 Conclusions

Designing iDTV applications is a complex activity due to several factors including the
ecosystem of media that compete and cooperate with the TV. In addition, the production
chains of the broadcasters are still not prepared to the design of iDTV applications. This
paper proposed three different practices and presented activities for supporting a situated
and participatory design of iDTV applications; a case study situated in real scenario of
a TV organization illustrated the proposal in action.
The results obtained from the case study indicated the benefits of using the practices
for supporting the involved parties to understand the situated context that the iDTV
application will be inserted, and to design an application that reflects that understanding.
The results suggested that the interactive prototype designed was widely accepted by
both the participants and prospective end users, pointing out the situated and participa‐
tory process as a viable and useful perspective for designing iDTV applications.
Although the results so far are very positive, the prototype still needs to be broad‐
casted as an iDTV application in Terra da Gente TV show. Thus, further work involves
the next steps of implementing and testing the final application and releasing it for use
by the TV program viewers. We also intend to conduct further studies within the
perspective of the Socially Aware Computing, to investigate the potential impact of the
practices presented in this paper to the TV staff and iDTV end users.

Acknowledgements. This research is partially funded by CNPq (#165430/2013-3) and FAPESP

(#2013/02821-1). The authors thank the EPTV team and the participants who collaborated and
authorized the use of the documentation of the project in this paper. This paper is an extended and
revised version of a paper previously published at ICEIS 2014: “Playing Cards and Drawing with
Patterns: Situated and Participatory Practices for Designing iDTV Applications” [10]. We thank
the Editors for the invitation.
 Designing an IDTV Application in a Situated Scenario 359

References

1. Alexander, C.: The Timeless Way of Building, 1st edn. Oxford University Press, Oxford
(1979)
2. Bannon, L.: Reimagining HCI: toward a more human-centered perspective. Interactions
18(4), 50–57 (2011)
3. Baranauskas, M.C.C.: Socially aware computing. In: ICECE 2009, VI International
Conference on Engineering and Computer Education, pp. 1–5 (2009)
4. Baranauskas, M.C.C.: O Modelo semio-participativo de design. In: Baranauskas, M.C.C.;
Martins, M. C.; Valente, J. A. (Org.). Codesign De Redes Digitais - Tecnologia e Educação
a Serviço da Inclusão. 103edn., vol. 1, pp. 38–66. Penso, Porto Alegre (2013)
5. Baranauskas, M.C.C.: Social Awareness in HCI. ACM Interact. 21(4), 66–69 (2014)
6. Bernhaupt, R., Weiss, A., Pirker, M., Wilfinger, D., Tscheligi, T.: Ethnographic insights on
security, privacy, and personalization aspects of user interaction in interactive TV. In:
EuroiTV 2010, 8th international interactive conference on Interactive TV and Video, pp. 187–
196. ACM Press, New York (2010)
7. Bonasio, V.: Televisão Manual de Produção & Direção. Editora Leitura, Minas Gerais (2002)
8. Borchers, J.: A Pattern Approach to Interaction Design. Wiley, England (2001)
9. Buchdid, S.B., Pereira, R., Baranauskas, M.C.: Creating an iDTV application from inside a
TV company: a situated and participatory approach. In: Liu, K., Gulliver, S.R., Li, W., Yu,
C. (eds.) ICISO 2014. IFIP AICT, vol. 426, pp. 63–73. Springer, Heidelberg (2014)
10. Buchdid, S.B., Pereira, R., Baranauskas, M.C.C.: Playing cards and drawing with patterns:
situated and participatory practices for designing iDTV applications. In: ICEIS 2014, 16th
International Conference on Enterprise Information Systems, pp. 14–27. SciTePress, Lisboa
(2014)
11. Cesar, P., Chorianopoulos, K., Jensen, J.F.: Social television and user interaction. Mag.
Comput. Entertainment 6(1), 1–10 (2008)
12. Chorianopoulos, K.: Interactive TV design that blends seamlessly with everyday life. In:
Stephanidis, C., Pieper, M. (eds.) ERCIM Ws UI4ALL 2006. LNCS, vol. 4397, pp. 43–57.
Springer, Heidelberg (2007)
13. Chung, E.S., Hong, J.I., Lin, J., Prabaker, M.K., Landay, J.A., Liu, A.L.: Development and
evaluation of emerging design patterns for ubiquitous computing. In: DIS 2004, 5th
Conference on Designing Interactive Systems: Processes, Practices, Methods, and
Techniques, pp. 233–242. ACM Press, New York (2004)
14. EPTV Portal. http://www.viaeptv.com
15. Fallman, D.: The new good: exploring the potential of philosophy of technology to contribute
to human-computer interaction. In: CHI 2011, Annual Conference on Human Factors in
Computing Systems, pp. 1051–1060. ACM Press, New York (2011)
16. Pressman, R.S.: Software Engineering - A Practitionar’s Approach, 5th edn. McGraw-Hill
International Edition, New York (2001)
17. Gamma, E., Helm, R., Johnson, R., Vlissides, J.: Design Patterns: Elements of Reusable
Object-Oriented Software. Addison-Wesley, Boston (1995)
18. Gawlinski, M.: Interactive Television Production. Editora Focal Press, Oxford (2003)
19. Harrison, S., Tatar, D., Sengers, P.: The three paradigms of HCI. In: Alt.CHI 2007 Annual
Conference on Human Factors in Computing System, pp. 1–18. ACM Press, New York (2007)
20. Kunert, T.: User-Centered Interaction Design Patterns for Interactive Digital Television
Applications, 1st edn. Springer, New York (2009)
21. Lewis, C.H.: Using the “Thinking Aloud” Method In Cognitive Interface Design. IBM
Research Report RC-9265, Yorktown Heights, NY (1982)
360 S.B. Buchdid et al.

22. Liu, K.: Semiotics in Information Systems Engineering, 1st edn. Cambridge University Press,
Cambridge (2000)
23. Miranda, L.C., Hornung, H., Baranauskas, M.C.C.: Adjustable interactive rings for iDTV.
IEEE Trans. Consum. Electron. 56, 1988–1996 (2010)
24. Müller, M.J., Haslwanter, J.H., Dayton, T.: Participatory practices in the software lifecycle.
In: Helander, M.G., Landauer, T.K., Prabhu, P.V. (eds.) Handbook of Human-Computer
Interaction, 2nd edn, pp. 255–297. Elsevier, Amsterdam (1997)
25. Pereira, R.: Key pedagogic thinkers: Maria Cecília Calani Baranauskas. J. Pedagogic Dev.
3(1), 18–19 (2013). UK
26. Pereira, R., Buchdid, S.B., Baranauskas, M.C.C.: Keeping values in mind: artifacts for a
value-oriented and culturally informed design. In: ICEIS 2012, 14th International Conference
on Enterprise Information Systems, pp. 25–34. SciTePress, Lisboa (2012)
27. Piccolo, L.S.G., Melo, A.M., Baranauskas, M.C.C.: Accessibility and interactive TV: design
recommendations for the brazilian scenario. In: Baranauskas, C., Abascal, J., Barbosa, S.D.J.
(eds.) INTERACT 2007. LNCS, vol. 4662, pp. 361–374. Springer, Heidelberg (2007)
28. Rice, M., Alm, N.: Designing new interfaces for digital interactive television usable by
older adults. Comput. Entertainment (CIE) – Soc. Telev. User Interact. 6(1), 1–20
(2008). Article 6
29. Sellen, A., Rogers, Y., Harper, R., Rodden, T.: Reflecting human values in the digital age.
Commun. ACM 52(3), 58–66 (2009)
30. Solano, A.F., Chanchí, G.E., Collazos, C.A., Arciniegas, J.L., Rusu, C.A.: Diseñando
interfaces graficas usables de aplicaciones en entornos de televisión digital interactiva. In:
IHC + CLIHC 2011, 10th Brazilian Symposium on Human Factors in Computing Systems
and the 5th Latin American Conference on Human-Computer Interaction, pp. 366–375. ACM
Press, New York (2011)
31. Sousa, K., Mendonça, H., Furtado, E.: Applying a multi-criteria approach for the selection
of usability patterns in the development of DTV applications. In: IHC 2006, 7th Brazilian
symposium on Human factors in computing systems, pp. 91–100. ACM Press, New York
(2006)
32. Stamper, R., Liu, K., Hafkamp, M., Ades, A.: Understanding the roles of signs and norms in
organisations: a semiotic approach to information system design. J. Behav. Inf. Technol.
19(1), 15–27 (2000)
33. Terra da Gente Portal. http://www.terradagente.com.br
34. VanGundy, A.B.: Brainwriting for new product ideas: an alternative to brainstorming. J.
Consum. Mark. 1(2), 67–74 (1983)
35. Veiga, E.G.: Modelo de Processo de Desenvolvimento de Programas para TV Digital e
Interativa. 141 f. Masters’ dissertation - Computer Networks, University of Salvador (2006)
36. Wilson, C.: Brainstorming and Beyond: a User-Centered Design Method, 1st edn. Elsevier
Science, Burlington (2013)
37. Winograd, T.: The design of interaction. In: Beyond Calculation: The Next Fifty Years of
Computing, pp. 149–161. Copernicus/Springer, Secaucus/New York (1997)
 Video Stream Transmodality

Pierre-Olivier Rocher1(B) , Christophe Gravier1 , Julien Subercaze1 ,

and Marius Preda2
1
Laboratoire Télécom Claude Chappe, Télécom Saint-Etienne,
Université Jean Monnet, 42000 Saint-Etienne, France
{pierre-olivier.rocher,christophe.gravier,
julien.subercaze}@telecom-st-etienne.fr
http://satin-ppl.telecom-st-etienne.fr/
2
Département ARTEMIS/GRIN, Télécom SudParis, 91000 Evry, France
[email protected]

Abstract. Transmodality is the partitioning of an image into regions

that are expected to present a better entropy using different coding
schemes, depending on their structural density, at constant bandwidth.
In this paper we present the transmodality of video stream. Our contri-
bution is a transmoder module that includes various different optimized
video codecs and implements the concept of transmodality on a set of
video streams. We evaluate our proposal in the context of cloud gam-
ing, using an optimized remote rendering chain and several games. A per
game adaptation allows an optimal refinement of video encoding para-
meters, including both quantization and modality parameters. Our algo-
rithm shows comprehensive results by saving up to 2 % of bandwidth for
the same PSNR in comparison with the state-of-the-art video encoding
baselines.

Keywords: Video encoding · Vectorization · Modality · Cloud gaming

1 Introduction

Video communication accords for one of the highest development slope among
various Internet applications for the last few years. It is forecasted to be one
of the main bandwidth consumer with respect to future applications. Different
kinds of applications are using video compression: live TV, online newspapers,
social networks. . . Over the last few years, new applications have emerged and
one of the most important one is probably the cloud gaming. All cloud gaming
solutions like Gaikai1 or OnLive 2 are based on remote rendering; such a system
is generally based on two parts. On the server side, the game is rendered and
the current picture is compressed using a state-of-the-art video encoder like
MPEG-4 AVC (a.k.a. H.264), then streamed to the clients. Currently, for all

1
http://www.gaikai.com/.
2
http://www.onlive.com/.
c Springer International Publishing Switzerland 2015
J. Cordeiro et al. (Eds.): ICEIS 2014, LNBIP 227, pp. 361–378, 2015.
DOI: 10.1007/978-3-319-22348-3 20
362 P.-O. Rocher et al.

previously described applications, the compression engine handles each image as

an atomic element, regardless if the image contains more or less homogeneous
regions. Adaptive encoding schemes bring new features, both in size and quality
term, but the video is still processed as a pixel set. Some regions of the picture
will be encoded with different parameters, but using the same global encoding
scheme. Criteria used in the adaptative encoding scheme are most often related to
objective metrics, like network bandwidth or latency. Recent studies highlight the
importance of subjective approaches, they seemed to be an interesting alternative
for controlling the adaptation process.
The intuition behind this work is to go one step beyond by defining an
adaptive region-based encoding algorithm using fundamentally different encod-
ing schemes. The encoding scheme should be chosen independently for each
region, depending on its structure and heterogeneity. In this paper we propose
several video coding systems to encode one video stream. More, proposed encod-
ing solution uses important areas (from the user’s viewpoint) in addition to a
structure and heterogeneity analysis. Proposed modifications will ensure that
the video stream is optimally compacted.
The remainder of this paper is organized as follows: Sect. 2 presents a review
of background literature. Section 3 details the concept of transmodality. Section 4
presents our video encoding system from a global perspective to detailed com-
ponents. Section 5 is dedicated to comparative testing and Sect. 6 concludes the
paper and draws perspectives.

2 Related Work

This section investigates existing work in the broad field of video encoding. While
a complete state-of-the-art on video coding is out of scope of this paper, we aim
at focusing only on relevant results in this part. We present two major encoding
methods relevant to our approach: matrix-based and the one based on graphics
primitives. We also present their applications to image and video coding. The
initiated reader may skip this section.

2.1 Image

We can consider that there are two ways to store and manipulate a digital pic-
ture. The first one is a matrix-based representation while the second one is based
on graphics primitives. Nowadays, both solutions coexist but each representation
brings specific advantages and flaws.
The matrix-based method is widely used for storing and sharing pictures.
This image representation, also known as raster or bitmap, is based on the pixel
definition. The most common encoders are BMP, PNG, and JPEG. Under this
method, images are represented as 2D-matrices. This is a very convenient way for
storing and even compressing pictures and pixel-based compression algorithms
are proved to be quite efficient. Although, its major drawback is decrease quality
while zooming into the parts of the picture. Even if the picture is uncompressed,
the original pixel element will eventually outsize the corresponding pixel in the
 Video Stream Transmodality 363

screen space. At this point, the picture quality will strongly suffer from this
limitation.
The second competing method does not use the pixel as the atomic element
but instead makes use of graphic primitives in addition to color profiles. Each
graphic primitive is a mathematical objet: it can be a point, a line or a polygon,
and is defined by a formal definition. These graphic primitives are a kind of
skeleton for the picture on which corresponding color profiles will be applied.
This representation then uses a set of graphic primitives and their associated
color profiles to build an image. However, for complex fitting more powerfull
objects like splines, Bezier curves or even Non Uniform Rational Basis Spline
(NURBS) can be used. Each graphic primitive is positioned in a reference system
and bears specific attributes like its color, shape or thickness. Due to its nature,
the picture is first rendered and displayed in a second time. i.e. every point is
computed using the formal definition to which it belongs. The result of this
computation is what is being displayed on the users device. The very main
advantage of this representation is the independence from the rendered image
size. Performance in size term is to mitigate since its mainly related on the
picture content itself: it can be greatly reduced, or largely more important. When
a picture or a video is live recorded or streamed, the raster format is used.
Meantime, graphic primitives formats (EPS, SVG) are more and more used in
the numeric world: due to the increase of terminal heterogeneity in our daily
lives (IPTV, smartphones, tablets. . . ) multimedia content is now often included
in web pages. These vector formats will insure that whatever the screen size,
display picture quality is maximal.
Building a raster representation of a vector picture is a trivial task, but
the reverse process (often called vectorization) is not obvious [1]. References on
this subject can be found [2–4], but proposed solutions are quite basic and are
most often limited to black and white pictures processing. Some approaches to
handle color exist but they cannot bear with photo realistic pictures (natural
shots). Another important limitation to overcome is the vectorization required
process time which is mainly picture content related. In subsequent sections, we
distinguish codec types based on pixel and vector: the first one is related to a
raster representation while the second one denotes the vector representation.

2.2 Videos
This subsection presents state-of-the-art video encoders (based on the MPEG-4
AVC video encoding standard). Our main focus is on literature aiming to use
several encoders to process a single video, but also on existing adaptive solutions.
The first paragraph is a reminder on how a modern video encoder is working.
To compress a video, it is first splitted into a so called Group Of Pictures
(GOP), which is basically made of three different types of pictures called I, P
and B. An I picture is a reference picture also known as a key frame. P pictures
are predicted using a past referenced picture. B pictures are bi-predicted, using
both past and future P pictures (depending on the standard definition). Main
steps of such encoders are prediction, transformation, quantization and entropy
coding. In the encoder scheme the main step responsible of data compression is
364 P.-O. Rocher et al.

quantization. In order to control the final bitrate, the user can tune the quan-
tization factor, called Qp in most common implementations. Various strategies
can be applied to control this parameter throughout the encoding process, like
constant Qp or constant bitrate (therefore with a non-constant Qp). A little Qp
value will ensure a good video quality. Furthermore usually, when Qp = 0 the
compression is lossless. Resulting entropy coded data, and necessary decoding
information like prediction mode, motion vectors, quantization factor form the
encoded video stream.
In [5,6] some macroblocks are encoded in a parametric manner, using an
Auto Regressive (AR) process [7]. Selection of these macroblocks is done using
a particular edge detection. Some moving textures can be modeled as a spatio-
temporal AR process, which is in fact a tridimensional basis model version. In
their system, I pictures are H.264 encoded, while P pictures are encoded using
the proposed method. The process is based on 16 × 16 macroblock size (owing
to H.264). macroblocks are then categorized in two sets: with or without edges,
where macroblocks without edges are processed with the proposed solution. The
quality assessment was done using the Mean Opinion Score (MOS) system since
the reconstruction scheme is statistically built. No information within the video
like spatial resolution, frames per second or even related enconding time are
exploited. In [8] authors offer to simply delete some macroblocks during the
encoding process and to rebuild them by synthesis methods at the decoding
stage. The method chosen to regenerate missing parts is the spatio-temporal
patch searching. The system is running with an I-P-B scheme. Some tests have
been conducted with QCIF videos, at 30 FPS using different Qp. Experimental
results show a bitrate reduction with a quite similar image quality (although
no objective metrics were used). In [9] the encoding process is based on both a
traditional bloc coding and on a model coding. For such appraoch, they use a
long-term temporal redundancy. Object detection related part is based on region
of interest (ROI) algorithms, these ROIs are detected using principal component
analysis (PCA). ROI areas are then segmented using graph-cut algorithms [10].
Finally a resultant area analysis is done, aiming to regroup some of them for
whole optimization. A tracking algorithm is used for following ROIs in futures
images. Theses ROIs areas, defined by a rectangle and an angle are finally com-
pressed using an active appearance model, which is statistically based.
Jointly to this research, work on adaptive video coding is largely present.
One can discern two major trends: the first needs several versions of a same
video; the second is based on a single version, which is adapted in realtime. The
first solution is covered by the Scalable Video Coding (SVC), but also by the
upcoming MPEG-DASH standard. Several recent articles can be found about
the second adaptive way. An other wide used solution is to produce as many
videos as screen sizes, like Netflix (each video is in 120 different copies). In [11]
real time adaptation is performed on video encoder as according to the network
delay. Latency measurements are done before the launch of the system, but also
periodically during the whole process. This latency evaluation makes possible the
detection of network congestion and then the ability to take necessary actions to
adapt the stream. In [11], they use a cloud gaming application, with an aim to
 Video Stream Transmodality 365

guarantee fluidity and responsiveness of the system. The proposed solution is to

reduce the quantization parameter (among all macroblocks) and the necessary
bitrate. In [12], the assumption between importants objects to the user and their
position in the depth map is done. In other words, more the object is close to
the user, more it is important. To reduce the necessary bitrate without loosing
in quality, they refine the Qp parameter, acting at macroblock level.
This approach seems to be better suited for outside games environment, i.e.
for scenes with an important depth map amplitude. In indoor games, range of
depth map values is much smaller, in consequence take a decision on each mac-
roblock is more complex and error prone. Additionnaly, the depth map criterion
is not a so good one: a defect can be pushed in this approach, especially when
the player tries to eliminate enemies far from his position. In this case, despite
the fact that the player is actually focusing on these enemies, corresponding
macroblocks will therefore be compressed with a greater Qp value. In fact, from
the depth map perspective, these macroblocks are considered as not important.
This scenario thus degrades the quality in an important area of the image.
In the previous sections, various ways to optimize the compactness of a video
stream were reviewed. To adjust the video compression, some are based on net-
work related metrics whereas others use empirical considerations. The inter-
est around cloud gaming is still increasing, many researchers are interested in
improving the different parts of the system, mainly using objective metrics. But
in a cloud gaming application, another parameter as to be taken into considera-
tion: the player himself. It is therefore interesting to analyze the behavior of the
player during a game session. Once important data for the player are known,
it becomes easier to adjust the compression, acting for example on some non
important areas. This type of approach acting on the perceived quality (com-
monly known as QoE) is related to subjective judgement. Assessment of QoE is
a hot topic in the media world [13], and it has been recently demonstrated that
the perceived quality is a very personal data [14].
As an exemple, in [15] an eye-tracker is used to highlight a correlation between
the ease of use of a website and the data collected by the eye tracker. In the cloud
gaming field, only a few papers are available, these include the work presented
in [16]. Authors introduced a conceptuel model for attention evaluation, based
both on salicency models and game logic. However the used model for visual
attention estimation on each picture is not design for games, it as been design
for static picture or video content. Moreover, although the authors suggest that
their system operates in real time, no study on the impact of the proposed model
as been presented.
Nevertheless, there is a common thread between all these solutions: they
always use a set of pixel to encode and adapt the video coding. All presented
solutions are using a standard version of H.264, We will use an optimized one
using each I and P pictures (B pictures are not used in cloud gaming video
compression systems: they introduce latency). Adaptive technologies are largely
used, but they are all the time confined to the pixel world. We will use an other
type of encoding system, which is based on vectorization. In order to effectively
366 P.-O. Rocher et al.

manage the various encoders, we choose to study the user behavior while palying
different games, and more particularly where his visual attention was focused.
We did intensive gaming tests by appealing several players acting on several
games in order to generate attention maps, which will subsequently be used by
the algorithm to finely manage encoders.

3 Transmodality
We introduce a new way of encoding a video stream using so called modalities.
A modality is a set of areas to be compressed regarding a specific encoder.
Each modality refers to (a) specific(s) part(s) in a frame, regarding an encoder
especially well suited for compressed corresponding areas in a picture. According
to our experiments we conclude that using more than one encoder will achieve a
better compression rate than using only one, while preserving the same quality in
terms of Peak Signal to Noise Ratio (PSNR). Conventional encoding approaches
consider a video as a set of pixels and use a single encoding scheme for the whole
picture. Our approach is based on the assumption that a dynamic partitioning
of the video frames and the approximation of some regions with a parametric
representation will reduce the video bitrate. Splitting a video stream into objects
and encoding them separatly is the basis of MPEG-4 part 2, but the way of doing
this work on a video stream is not covered by this standard.

3.1 Transmodality Definition

We define the Tansmodality as the fact of using several specific encoders to com-
press the same video file. The aim is to use a well-suited encoder to each region
and using different modalities for video encoding. In other words, a modality
is defined as being the compression of an area set by an encoder. We define
a class as a set of areas, each class being encoded using one specific encoder.
Using one or more modalities is not known by advance but it belongs to the
processed video itself. The decision to use 1, 2 or n modalities is taken in real
time during multimodal encoding, which means that shapes and area size are
potentially different for each frame. For instance, Fig. 1 is made of five areas z1
to z5 which are grouped into three classes m1 to m3 . An area is a contiguous set
of pixels, contiguity of two pixels is defined using the Kronecker operator (V4
neighborhood).
In the following section we focus on the transmoder whose role is to process
the video compression with multi-modalities.

3.2 Transmoder vs Transcoder

A transcoder is a software or hardware element which aims to modify the way
how a video file or stream is compressed. It first decodes a video file or stream,
and encodes it again using new parameters. This enables the application to mod-
ify the way this video will be transformed including three majors trends: the
 Video Stream Transmodality 367

Fig. 1. Areas and modalities in a sample picture.

output video size (spatial resolution), the frame rate (temporal resolution) and
the quality (resolution in PSNR). In this work, we introduce a new adjustable
parameter called modality. Doing so, we define a new application called trans-
moder, which is a transcoder with an ability of using modalities, as described
in previous sections. Our Transmoder as compared to transcoder uses an addi-
tional parameter that increase the output range of possible bitrates and thus
make it possible to distribute the video to a larger number of people. A trans-
moder is also able to conduct simple transcoding operations, so with regard to
our definition, the transcoder is a particular case of the transmoder.
In this work, we focus on implementing a transmoder that can handle two
modalities, the pixel and the vector ones. What follows therefore presents our
bimodal transmoder approach. In order to simplifiy, following parts of this article
will rely on a transmoder working on two modalities. The multimodal process is
then limited to a bimodal approach.

4 Transmoder Process

Our transmoder is built on three separated parts: a decoder, an encoder and

a sunder. The whole architecture is presented in Fig. 2. The decoder decodes
frames from a video file or stream. The sunder part is in charge of the partitioning
of each frame into n modalities, while taking into account attention maps. As
depicted in Fig. 2, the sunder output is a combination of n modalities denoted
M1 to Mn. Finally, the encoder uses the processed results of sunder as its input
and encodes the whole video in a transmoded stream. The remaining systems
blocks are explained in the following sections.

4.1 Attention Maps Generation

The state-of-the-art has shown that cloud gaming applications start to consider
attention (from a general point of view) to perform a compression adaptation.
368 P.-O. Rocher et al.

Fig. 2. Synoptical transmoder scheme.

Given this trend, we offer a new approach to refine attention models with the
aim of making them better.
The goal is firstly to determine which areas are important, depending both
on the game and on the player. In order to answer this question, we customized
our cloud gaming research plateform. It is now able to manage an eye-tracker so
when each picture is displayed, corresponding data (eyes position) is recorded
in a specified file. For our first expermiments, two games are available: Doom 3
and 0 A.D. First one is a First-Person Shooter (FPS) while the second belongs
to Real-Time Strategy (RTS) video game genre. The first phase of testing is
performed using a conventional video compression (no adaptation, constant Qp)
to make sure that the player’s attention is not disturbed. This stage will allow
the construction of visual attention maps for each game and each player. The
protocol to follow is quite simple: after calibration of the eye-tracking device,
users are ask to play one hour each game. To maintain consistency, all tests
are performed in the same room, with the same environment. Our tester panel
consists of three experienced players, who have completed three time the same
test with available games.
Before going further, consistency between the three tests made by a user with
a game is checked. This is done by computing the Pearson correlation coefficient
for each test, two by two. Our first conclusions show that players behave similarly
for the same game, so visual attention is game related. In order to reflect this
conclusion, per game final attention maps are built using all available player
data. These maps are considered as baselines for future quality evaluation.
 Video Stream Transmodality 369

4.2 The Sunder

The aim of Sunder is to take the decision of splitting (or not) the frame in several
modalities. All the required steps for such functionality as depicted in Fig. 2 are
explained later. But at this point we consider that the input of this whole process
is an uncompressed picture, typically in an Red Green Blue (RGB) raw format.
Some image processing treatments is first applied, then the picture is splitted
into modalities (see section “Modality Splitting”). Once both modalities are
splitted, respective encoding is conducted and some optimizations are applied
on the pixel encoding part (see section “Analysis”). The ouput of this process
is a so-called bimodal picture. While all necessary informations are available,
the video stream can be recorded. This is done by using packets, which are
usually containing one frame. Packetisation process is described in details later
(see section “Packetisation” of Sect. 4.3).

Modality Splitting. Modality splitting involves with splitting operation by

exploiting well known image processing filters. Previously, we explained our
choice to split a picture into two modalities. To operate efficiently, an image
processing filter appears to be the best choice. Our approach splits an image,
based on the level of details which can be performed using an edge detector
filter. There are various filters available for such requirements including Canny,
Laplace an Sobel. For scalibility reasons, we require an algorithm that exhibits
a good trade-off between performance an computation time. Under these condi-
tions, [17] and more recenlty [18] concludes that the Canny filter is the filter of
choice. Furthermore, this choice is affirmed by [3] in which the authors processed
videos. It should be noted that this process (and later described sub processes)
is not applied on the whole picture but only on non important regions, according

Fig. 3. Sample input picture.

370 P.-O. Rocher et al.

(a) Canny filtered input picture representation. (b) Macroblocked representation of vector areas (in
white).

Fig. 4. Canny filtered picture and its macroblocked version, based on 3 input picture.

to the game’s attention map. As an exemple, lets take the picture presented in
Fig. 3 as an input picture. After the edge detector process, we obtain a binary
picture as displayed in Fig. 4(a). As the reader can see, details in the picture
are spotted in white, and more uniform areas are in black. Some morphologi-
cal operations are now necessary to clean this mask (remove single points, close
holes. . . ).
At this point, we need to make a choice regarding modality encoding. We
choose to use one encoder from the pixel world and another one from the vector
one. Chosen pixel encoder is the well-known MPEG-4 AVC, the state-of-the-art
reference. As literature does not provide any efficient library for a raster to vector
conversion, we thus employ our own vectorization module. This one is designed
to suit our specific needs and is further described in section “Vectorization” of
Sect. 4.3. As we previously said, in the binary mask black parts represents a high
amount of details while white parts depicts more uniform areas. Based on the char-
acteristics of selected encoders, Black parts may be pixel coded whereas white
parts may be vector coded. At this point, white areas are candidates to be vector
coded, but further analysis (as described in see section “Analysis”) may down-
grade them in the pixel world to ensure a good compacity in the output video
stream. Because we choose to compress a modality using a MPEG-4 AVC encoder,
the binary mask needs to be adapted since this kind of encoder is based on a mac-
roblock definition (a 16 × 16 pixels size). This is simply done in checking each pic-
ture macroblocks: if all pixels are black, then the macroblock is black otherwise it
is white. The corresponding new macroblock mask is depicted in Fig. 4(b).

Analysis. This concludes as the most important step of transmoding. All deci-
sions made here are based on parameters set by the user at startup time. Cur-
rently a macroblock mask is present, the aim is now to process this mask and take
appropriate decisions. Important steps of this analysis are presented in Fig. 5.
 Video Stream Transmodality 371

Fig. 5. Logic analysis process in the sunder.

Labeling. The first operation involves a V4 neighbourhood labelling of white

areas. This will ensure that each vector area is separated from the others, and
provides an internal data structure easy to manipulate.

First Filter. Filter 1 as described in Fig. 5 is responsible of the deletion of some

area, based on surface size criterion as according to the threshold set by the user
(or by the default value). Consequently, some little areas which were planned to
be vector encoded is in place pixel encoded, because in this case a pixel encoder
offers better results. As an exemple, Fig. 5 shows a possible output result.

Largest Rectangle Decomposition. According to Fig. 5, we can now consider two

vector areas. These areas can have any shape, including holes. To make the
decoding stage possible, we will need to know two things: covered areas and
their respective definitions.
We need now to think on how to store efficiently the binary mask. We can
consider a lot of different solutions but in all cases, we need to keep a very good
compacity. In order to fulfill this need, we choose to decompose any vector area
into a set of largest rectangles. This kind of operation is known as a rectangle
decomposition of binary matrix, henceforth DBMR problem. More information
on this subject can be found in [19]. Most of the time, taking into account the
first two or three biggest rectangles is enough to keep a good approximation of
the intial shape.

Second Filter. Our second filter aims at deleting little areas (user set thresold)
that are moved in the pixel modality. At the end of this analysis, only important
372 P.-O. Rocher et al.

large rectangle areas are candidate for a parametrized representation. This

process is detailed in section “Vectorisation” of Sect. 4.3.
This is the last step of sunder process. A binary mask specifies if each area
has to be compressed using the first or the second modality. Thus encoding
operations, related optimizations, and stream generation can be conducted.

4.3 Stream Encoding

The modality splitting step is now finished. All pixel and vector areas are
known and respective encoding operations can be conducted. The first para-
graph describes our vectorization process while the second one is devoted to
optimizations. Bimodal video stream is then packetized and dumped in an out-
put video file or stream.

Vectorization. As described earlier, vectorization is not a trivial task, espe-

cially when the user wants to process a natural picture and expect good quality
results. The main problem of such encoding scheme includes complexity of such
a project, necessary time to process pictures and output obtained sizes. All
these parameters are intimately related to the picture content. The desirable
objectives at this point is the implementation of specific software that meets
our need along with a fast processing and a good quality (in term of PSNR).
Because as the previous operations provides only a specific texture type and
in order to achieve a uniform one, we can utilize a simple vectorization app-
roach. As a first implementation we choose to simply use a polynomial based
approach, inspired by what has been done in [3]. In order to vectorize an area,
we used sequence of retangles as graphic primitives, and a polynomial based
colour profil. Each polynomial expression is based on a static template of the
form: Z = a + bx + cy + dx2 + exy + f y 2 , computed using a least square based
regression. Each vector area have three unique polynomials, one for each picture
color channel. To control the quality of color profiles, respective correlation coef-
ficients are computed. If their values are too low (thus implies a poor modeling
quality), the corresponding area will fall back in the pixel mode.

Encoding Optimizations. Encoding a single video using a per area optimized

compression is not such a tedious task. The most difficult part is to use them
correctly while avoiding redundancy in the data stream and trying to limit the
impact on the encoding time. However, the only fact of adding some new lines
of code will inevitably increase the necessary computing power.
Starting from this fact, we tried to minimize the impact of the sunder in
optimizing the way encoder works. Both vector and pixel parts have been opti-
mized. After all areas are vectorized, a process checks if some of them can be
modeled using a same polynomial definition. Special treatment is reserved for
writing vector data: for example, the number of digits is limited and all the data
is compressed. The sunder process gives us some useful informations, which can
be directly sent to the vidoeo encoder. We can denote two optimizations: the
first one (O1) directly indicates to the encoder macroblocks to skip, and the
 Video Stream Transmodality 373

second (O2) manipulates the Qp parameter. Each of these two optimization is

macroblock based. The aim of O1 is to directly provide a list of macroblocks to
encoder that have to be skipped. Thus the analysis cost is reduced (minimize the
encoding time) and the stream is more compact, regardless the chosen encoding
preset. This optimization is only applied on P frames. O2 optimization is related
to I and P frames. Per macroblock Qp is modified to force the encoder to skip
some macroblocks (thus which are not used by the pixel modality). These two
optimizations can be used separately but also at the same time. They do not
impact the quality of the output stream.

Packetisation. For being able to add our vectorization data part, we modify
the way packetisation is done. Each vector data is just positionned after the
picture data, hence a frame is a combination of pixel and vector data. As the
splitting is done live in an automatic manner, some packets may have only a pixel
part, some others only a vector part. This stream is written in such a way that
any compliant MPEG-4 AVC decoder will be able to read it, however only the
usual pixel modality will be correctly displayed on the screen. In order to get the
full picture correctly constructed, a suitable decoder is necessary: final decoded
picture is a combination of the rendered vector part and of the traditional pixel
representation.

5 Performance Evaluation
We evaluate our solution using the Doom33 game. This video game is a First-
Person Shooter (FPS), so it comports a lot of action and rapid scenes. For testing
purpose, video is dumped in a raw file (RGB format) while the user is playing.
The obtained videos raw files are the result of different gaming sessions, there are
all different. We use our transmoder through our remote rendering chain. As a lot
of parameters can be tuned and will inevitably lead to different results, we choose
to use static parameters for testing. Both modalities coding parameters are fixed
to make comparisons possible. Furthermore the pixel video encoding strategy
uses a constant quantization parameter, which is set by default at Qp = 20.
All testing movies have been a 720p resolution, frame number varies from 300
to 500 pictures. Optimizations presented in section “Encoding Optimizations”
of Sect. 4.3 are used. Our transmoder implementation outputs XML files with
all necessary information, like frame sizes, made decisions. . . Results between a
classical approach (transcoder) and our approach are then analyzed. These first
tests results only use an objective metric (PSNR) but further assessment will
use subjective evaluation methods based on attention maps.

5.1 Implementation
Video encoding is a quite heavy task even on most recent computers. The
choice of a native programming language is then obvious. Because of code source
3
http://www.idsoftware.com/games/doom/doom3.
374 P.-O. Rocher et al.

portability but also external available libraries, the preferred choice for our imple-
mentation is C/C++. In order to accelerate the developping time and reduce the
source code size to maintain, some well known multimedia libraries were used,
such as FFmpeg and libx264 for decoding, encoding and manipulate the different
kind of codec and containers. Image processing related operations are done by
OpenCV, and our basic vectorization software uses the GNU Scientific library
(GSL). The whole program is available as a standalone executable capable of
transmoding any kind of videos, or as a library for direct splitting and encod-
ing. Obviously the transmoder is also capable of simple transcoding operations.
Although our model is designed for any number of modalities, our current imple-
mentation is limited to two: a pixel modality and a vector one.

5.2 Results
Table 1 presents obtained results in size between transcoded and transmoded
versions of a same input raw video. These experiments have been done using the
veryfast preset and the zerolatency tune. Column TC indicates the transcoded
file size, while TM(1) and TM(2) are transmoded output file size with differ-
ents settings (those are presented in section “Analysis” of Sect. 4.2). In TM(1)
experiment, both filters have been set to 1 and in the second test, TM(2) both
filters have been set to 2. Our transmoder outperforms transcoding operations in
all tested cases. Obviously, one can notice than chosen parameters influence the
final result. Gains in size ranging from 0.1% (Doom 3 D video) to 2.6% (Doom
3 A video). A sample snapshot is presented in Fig. 6.

Quality Assessement. Quality assessment is not a trivial job. Usually for

this task, some metrics as well-known PNSR or MSSIM are used to complete
surveys, but this is not sufficient. For now we need to take care of subjective
meaning and not only objective ones. First conducted tests used a PSNR metric
for quality assessment, with some particularities. Indeed, we choose to calculate
a per modality and a per framne PSNR value, aiming to preserve a maximum
amount of information for later analysis.
Table 2 outputs PSNR results for all tested videos, using a per picture com-
puting. We can notice that two times (Doom 3 A and Doom 3 D videos) out of
four, transmoding PSNR results are better than transcoding equivalent opera-
tion. When the transcoding operation is better in PSNR, the relative difference
is negligible.

Table 1. Obtained transcoded and transmoded file sizes (in bytes).

Video name TC TM(1) TM(2)

Doom 3 A 9907403 9824570 9644114
Doom 3 B 13178506 13133818 13077405
Doom 3 C 14912310 14876477 14831883
Doom 3 D 13151625 13145306 13134501
 Video Stream Transmodality 375

Table 2. Obtained transcoded and transmoded PSNR values (in dB).

Video name 0-TC O-TM(1)

Doom 3 A 31.8201 31.8211
Doom 3 B 32.3262 32.3253
Doom 3 C 35.4112 35.4086
Doom 3 D 31.0368 31.0379

Table 3. Per modality PSNR values (in dB) compared to original video.

Video name TC-P TM(1)-P TC-V TM(1)-V

Doom 3 A 34.142 34.1514 31.7812 31.7811
Doom 3 B 38.3652 37.8888 32.0904 32.0904
Doom 3 C 39.1548 39.1019 33.9563 33.9562
Doom 3 D 32.08 32.0538 31.1088 31.1088

For a better understanding of the process, and to make sure that PSNR
are similar in vector and pixel parts, we computed PSNR values per modality.
All PSNR computations have been done between the original raw video and
obtained transcoded and transmoded videos. Table 3 summarizes results for all
videos. One of the first observations is that vector modality PSNR is very simi-
lar, which means that our vectorization software is able to compete a pixel-based
encoder in such regions. In vector areas, differences between a transcoded and a
transmoded area is negligible. This last remark is also applicable to respective
pixel areas.

Fig. 6. A Doom 3 final rendered picture.

376 P.-O. Rocher et al.

5.3 Real Time Processing

Our first tests reveals that real time processing is possible with videos assuming
quite little spatial resolutions (320 × 240). In order to overcome this limitation,
we first optimize the whole architecture of the software, especially by adding
threading support in appropriate source code parts. Another way of increasing
performance, especially in the image processing world is to use the GPU. This
aspect has been brought to our program using OpenCV GPU related functions.

Table 4. Necessary time for transcoding and transmoding operations (in ms).

Experiment Doom 3 A Doom 3 B

TC 1 7062 6906
TM(1) 57078 57672
TM(2) 67766 63921
TM(1)-8T 16015 15937
TM(2)-8T 19703 20328

Table 4 outputs necessary time to conduct transcoding and transmoding

operations, without any software optimization. In this case, one can see that
transmoding operations take 7 to 8 times than simple transcoding. If we acti-
vate the multithreading support, using 8 threads significantly reduces necessary
time (8T suffix in the table). As doom 1 and doom 2 videos include 500 frames,
we can say that our system is able of live transmoding (≈ 40 ms per frame).
A 2 % improvement is quite little one, but this preliminary work comfort us
in exploiting other optimizations. Some optimizations have been done on both
encoders, but they can be further improved (accuracy, bitstream format. . . ). In
the mean time, the decoding process can also be improved. For exemple, at the
moment no image processing is done to reduce visual artifacts between pixel and
vector areas while decoding. Results have to be analyzed in the light of the used
cloud game - a FPS - which brings by definition a lot of action. Further work on
other kind of games will be conducted.

6 Conclusions and Future Works

In this paper, we presented a new type of video encoding system called trans-
moder. The video stream is splitted into regions that are encoded using several
modalities depending on the regions characteristics. We proposed an overall sys-
tem architecture for transmoding, that we implemented using two modalities,
pixel and vector encoding. Corresponding streaming architecture is presented.
We first split each frame into regions using an edge detector. We then deter-
mine the more relevant encoder for each region. We output a bimodal video
stream combining vector and pixel frames. We tested our approach in a cloud
 Video Stream Transmodality 377

gaming use case, using a FPS game. Performance analysis shows that our app-
roach outperforms state-of-the-art encoder for a large majority of testbed videos.
Although immature with respect to the transcoder baseline, the system could
compete with state-of-the-art solutions while other optimizations are still pos-
sible. This comfort us in continuing the exploration of transmoder algorithm
and architecture. Conducted tests present a reduction of the necessary bitrate
up to 2 %, based on objective metrics. Future research will allow us to optimize
even more our system, using for example GPU optimized algorithms (aiming
at reduce necessary processing time), but also per game attention maps profiles
to know precise locations of important and non important picture parts. In the
frame of our research on cloud gaming systems, we aim to integrate this mul-
timodal coding scheme in the realtime rendering chain. We therefore adapt our
approach for distributed processing in a cloud architecture.

References
1. Sun, J., Liang, L., Wen, F., Shum, H.Y.: Image vectorization using optimized
gradient meshes. In: ACM SIGGRAPH 2007 papers. SIGGRAPH 2007. ACM,
New York, NY, USA (2007)
2. Lai, Y.K., Hu, S.M., Martin, R.R.: Automatic and topology-preserving gradient
mesh generation for image vectorization. ACM Trans. Graph. 28, 85:1–85:8 (2009)
3. Zhang, S.H., Chen, T., Zhang, Y.F., Hu, S.M., Martin, R.: Vectorizing cartoon
animations. IEEE Trans. Vis. Comput. Graph. 15, 618–629 (2009)
4. Orzan, A., Bousseau, A., Winnemller, H., Barla, P., Thollot, J., Salesin, D.: Diffu-
sion curves: a vector representation for smooth-shaded images. ACM Trans. Graph.
27, 92:1–92:8 (2008)
5. Chaudhury, S., Mathur, M., Khandelia, A., Tripathi, S., Lall, B., Roy, S.D.,
Gorecha, S.: System and method for object based parametric video coding. U.S.
Classification: 375/240.16; 375/240.25; 375/E07.027; 375/E07.076 (2011)
6. Khandelia, A., Gorecha, S., Lall, B., Chaudhury, S., Mathur, M.: Parametric video
compression scheme using AR based texture synthesis. In: Sixth Indian Conference
on Computer Vision, Graphics Image Processing, 2008. ICVGIP 2008, pp. 219–225
(2008)
7. Szummer, M., Picard, R.W.: Temporal texture modeling. In: 1996 Proceedings
of the International Conference on Image Processing, vol. 3, pp. 823–826. IEEE
(1996)
8. Zhu, C., Sun, X., Wu, F., Li, H.: Video coding with spatio-temporal texture syn-
thesis. In: ICME, pp. 112–115 (2007)
9. Tripathi, S., Mathur, M., Dutta Roy, S., Chaudhury, S.: Region-of interest based
parametric video compression scheme. Communicated to the IEEE Transactions
on Circuits and Systems for Video Technology (2012)
10. Felzenszwalb, P.F., Huttenlocher, D.P.: Efficient graph-based image segmentation.
Int. J. Comput. Vis. 59, 167–181 (2004)
11. Tizon, N., Moreno, C., Cernea, M., Preda, M.: MPEG-4-based adaptive remote
rendering for video games. In: Proceedings of the 16th International Conference on
3D Web Technology. Web3D 2011, pp. 45–50. ACM, New York, NY, USA (2011)
12. Tizon, N., Moreno, C., Preda, M.: ROI based video streaming for 3D remote render-
ing. In: 2011 IEEE 13th International Workshop on Multimedia Signal Processing
(MMSP), pp. 1–6 (2011)
378 P.-O. Rocher et al.

13. Vankeirsbilck, B., Verslype, D., Staelens, N., Simoens, P., Develder, C., Demeester,
P., De Turck, F., Dhoedt, B.: Platform for real-time subjective assessment of inter-
active multimedia applications. Multimedia Tools Appl. 72, 749–775 (2013)
14. Lavignotte, A., Gravier, C., Subercaze, J., Fayolle, J.: Quality of experience, a
very personal experience!. In: 2013 24th International Workshop on Database and
Expert Systems Applications (DEXA), pp. 231–235. IEEE (2013)
15. Ehmke, C., Wilson, S.: Identifying web usability problems from eye-tracking data.
In: Proceedings of the 21st British HCI Group Annual Conference on People and
Computers: HCI... but not as we know it, vol. 1, pp. 119–128. British Computer
Society (2007)
16. Ahmadi, H., Khoshnood, S., Hashemi, M.R., Shirmohammadi, S.: Efficient bitrate
reduction using a game attention model in cloud gaming. In: 2013 IEEE Interna-
tional Symposium on Haptic Audio Visual Environments and Games (HAVE), pp.
103–108. IEEE (2013)
17. Heath, M., Sarkar, S., Sanocki, T., Bowyer, K.: Comparison of edge detectors: a
methodology and initial study. In: 1996 Proceedings of the IEEE Computer Society
Conference on Computer Vision and Pattern Recognition, CVPR 1996, pp. 143–
148 (1996)
18. Maini Raman, D.H.A.: Study and comparison of various image edge detection
techniques. Int. J. Image Process. 1–11 (2009)
19. Ferrari, L., Sankar, P., Sklansky, J.: Minimal rectangular partitions of digitized
blobs. Comput. Vis. Graph. Image Process. 28, 58–71 (1984)
STAR: Speech Therapy with Augmented Reality
for Children with Autism Spectrum Disorders

Camilla Almeida da Silva1 , António Ramires Fernandes2(B) ,

and Ana Paula Grohmann3
1
Faculdade de Computação e Informática, Universidade Presbiteriana Mackenzie,
São Paulo, Brazil
[email protected]
2
Centro Algoritmi, Universidade do Minho, Braga, Portugal
[email protected]
3
Associação para a Inclusão e Apoio ao Autista, Braga, Portugal
[email protected]

Abstract. Graphics based systems of Augmented and Alternative Com-

munication are widely used to promote communication in people with
Autism Spectrum Disorders. This study discusses an integration of Aug-
mented Reality in communication interventions, by relating elements
of Augmented and Alternative Communication and Applied Behaviour
Analysis strategies. An architecture for an Augmented Reality based
interactive system to assist interventions is proposed. STAR provides an
Augmented Reality tool to assist interventions performed by therapists
and support for parents to join in and participate in the child’s interven-
tion. Finally we report on the usage of the Augmented Reality tool in
interventions with children with Autism Spectrum Disorders.

Keywords: Autism spectrum disorder · Augmented reality · Aug-

mented and alternative communication · Human-computer interaction

1 Introduction
Children with Autism Spectrum Disorders (ASD) are affected with various
impairments in communication, social interaction and imagination, three major
components of self-development. Regarding communication issues, these range
from the total absence of language, 20 % to 30 % [1], to the lack of effective-
ness in the communication process, reinforcing the need to promote initiatives
to improve communication skills.
Applied Behaviour Analysis (ABA) interventions for autism in early child-
hood are an effective practice to improve socially relevant behaviours and
teach new skills, through several established teaching tools and positive rein-
forcement strategies. Augmented and Alternative Communication (AAC) based
approaches, in particular those based on graphics, are the most used in inter-
ventions for children with ASD. These interventions are considered to be highly
c Springer International Publishing Switzerland 2015
J. Cordeiro et al. (Eds.): ICEIS 2014, LNBIP 227, pp. 379–396, 2015.
DOI: 10.1007/978-3-319-22348-3 21
380 C.A. da Silva et al.

relevant in communication promotion, reduction of behavioural problems, and

environmental awareness [2].
The graphics based system usage is supported by the strong visual process-
ing ability seen in many children with ASD. Nevertheless, imagination impair-
ments caused by rigidity and lack of imagination can prevent the usage of some
symbols [3]. In [3] it is argued that some children with ASD, when confronted
with a symbol, only capture a set of lines, shapes and colours, and the usage
of communication symbols would lead to memorization and context association.
The authors show that changes in the background or width of the lines are
sufficient to prevent recognition on a previously learned symbol. Hence, they
conclude that these symbols are meaningless to those children. In [2] it is stated
that the lack of symbolic capacity is one of the main handicaps in communi-
cation, reflecting the difficulty in learning conventional or shared meanings of
symbols. To achieve communication, symbols must be meaningful to the child [4].
Computer animations, 3D graphics, and sounds, may provide deeper engage-
ment for children in their activities, gains in learning, and reduction of behav-
ioural problems, as suggested in several studies on the benefits of Information
Technologies in interventions with children with ASD [5–7]. The use of tangible
interfaces has also been suggested by [8] to promote alternative ways of collabo-
ration and communication, engaging the children for longer periods of time than
with regular interventions.
Augmented Reality (AR), merging virtual objects into real environments, can
be explored in some intervention areas for ASD children, such as self-awareness,
augmented communication, emotion awareness and identification, social com-
mitment and concept development [9]. These visually oriented approaches are
highly suitable due to the remarkable ability of most autistic people to excel at
visual spatial skills [10].
Considering the relevance of AAC approaches, and the impact of symbols
to promote communication, we performed a study on the usage of AR to assist
interventions and expand language with children with ASD. The interactive AR
system developed in this study uses both AAC and ABA principles to support
interventions focused on facilitating graphic symbol comprehension. The sys-
tem reinforces the communication symbols superimposing virtual objects and
animations over the real environment screen view.
The system’s human-computer interaction was previously discussed with
speech-language therapists to ensure its suitability for both therapists and chil-
dren. The system’s features include the creation of activities with AR graphics
and sound to enrich the symbol’s meaning, with both visual and audio rein-
forcements. The system is adaptable to each individual child, allowing to have
multiple activity configurations, and supports a number of different activity tem-
plates to increase the options for the therapist.
In order to evaluate the system, the application was tested in speech-language
sessions with four children with ASD. The tests fulfilled the framework proposed
by Moore [5], where computer assisted learning systems for ASD children should
focus on at least one on the major impairments; the projects should be based on
 STAR: Speech Therapy with Augmented Reality 381

current established practices; and evaluation should be performed in cooperation

with actual therapists.
In Sect. 2 we provide some background information on ASD. Related work
is described in Sect. 3. The architecture of our proposed system is described in
Sect. 4, and the working prototype is presented in Sect. 5. The evaluation of the
prototype is discussed in Sect. 6. Conclusion and possible avenues for future
work are presented in Sect. 7.

2 Autism Spectrum Disorder

ASD is a group of developmental brain disorders affecting individuals from all

races, cultures and socio-economic groups. It presents a wide range of possi-
ble symptoms and varying degrees of severity, the common feature being an
early developmental disruption in socialization processes [1]. Impairments in
communication, social interaction, and imagination - the triad of impairments
[11] - causes difficulties in understanding and using verbal and non-verbal com-
munication, to relate and empathize with other people, and rigidity and inflexi-
bility in thinking, language and behaviour.
The Centers for Disease Control and Prevention [12] estimate that about
1 % of individuals have ASD in Europe, Asia, and North America. ASD is more
common among boys than among girls, with a ratio of 5 to 1.

2.1 The Triad of Impairments and Behavioural Issues

The difficulty of socialization is the most remarkable feature, being more evident
in the first years of life. Even in the first few months of life, many infants with
ASD do not interact and avoid eye contact, in contrast to neurotypical infants
[13]. Social behaviour is not absent, but is developed with absolute lack of social
and emotional reciprocity response [14]. One of the consequences of this difficulty
is the lack or decline to imitate, one of the crucial prerequisites for learning and
a pre-linguistic skill to language development.
The imagination impairment is manifested by obsessive and ritualistic behav-
iours, literal understanding of language, resistance to changes in routine or envi-
ronment, inability to simulate, imagine and participate in pretend games. The
consequences of this inability affect several areas of thought, language and behav-
iour, preventing, for example, the development of skills of social reciprocity, lan-
guage construction, and acquisition of the word.
The difficulty in coordinating attention between people and objects - shared
attention - and the difficulty of learning the conventional meaning of symbols -
symbolic capacity - are the core deficits of communication.
The behavioural issues are the most influential regarding the social integra-
tion of a children with ASD. These include hyperactivity, lack of focus, mood
variability, unexpected movements, and uncommon reaction to sensory stimuli.
382 C.A. da Silva et al.

2.2 Interventions
The main goal of an intervention is to promote functional and spontaneous commu-
nication, social interaction and symbolic play, in order to increase functional inde-
pendence and quality of life. These interventions address communication, social
skills, unsuitable behaviours, among other common issues caused by ASD [15].
Intervention programs require a multidisciplinary base, involving behavioural ther-
apy, educational programs, speech-language and communication therapy [16].
Applied Behavioural Analysis methods, such as intense behavioural inter-
ventions, directed towards long term behaviour modification, have as goals to
increase and maintain adequate behaviours, reduce maladaptive behaviours,
teach new skills, and enhance the generalization ability to new contexts [15].
These methods have a short duration and are effective in both adults and
children [2]. Discrete Trial Training (DTT) is also an ABA method to enhance
skills such as attention, compliance, imitation, and discrimination learning,
among others [17]. It consists of a structured learning method whereby a complex
learning sequence is divided in small steps, with reinforcements and assistance
by the therapist where required.
Regarding ASD, the main goal of AAC is to enhance symbol acquisition, since
the development of mental models provides the ability to create abstractions
from concrete situations and use them when communicating [18].
The American Speech-Language-Hearing Association defines AAC as an area
of clinical practice focused on impairment and disability patterns of individuals
with severe expressive communication disorders [19]. AAC methods and tools
are also highly adaptable to the autistic child needs and personality.
Graphics based systems take advantage of the strong visual processing skills
found in many individuals with ASD, and have been proven effective, through
functional communication training, in increasing communication reception in
small children, and to replace disruptive behaviour such as aggression, self-
aggression and crying [2]. The environment, space, and time concepts are other
examples where AAC can be used.
The Picture Exchange Communication System (PECS), developed by Bondy
and Frost in 1985 [20], is an AAC method for functional communication based on
picture exchange. PECS follows ABA principles being structured in a sequence
of six steps, each with specific goals, instructions, and training procedures. Ini-
tially the child is taught to initiate requests through figures. Later the child will
start building sentences, and answering questions. Later stages require the child
to make comments. Several studies [21–23] have shown the benefits of PECS,
resulting in vocabulary acquisition and social behaviour improvements.

3 Related Work
In [6] the impact of an educational software in vocabulary acquisition is dis-
cussed. The software was designed to parallel the existing behavioural program,
adding audio reinforcements and object motion. The authors conclude that the
usage of the software enhanced the children’s focus, motivation and vocabulary
acquisition.
 STAR: Speech Therapy with Augmented Reality 383

3D Virtual Environments and AR techniques have also been used extensively

to assist interventions for children with ASD.
Project INMER-II [24] was designed based on a serious game in a virtual
environment related to shopping activities. The results show increased levels of
functional, symbolic comprehension, and imagination. Furthermore, one child
was able to successfully generalize the experience to a real environment.
Project AS Interactive was developed to support people with ASD to improve
social skills through 3D virtual environments simulating a cafeteria and a bus
[25]. Environments and learning objectives were designed in cooperation with the
participants’ teachers. User interaction is performed with a joystick and mouse.
The participants were able to successfully navigate in the virtual environment
and interacted with the system effortlessly. The study performed with teenagers
with ASD led to the conclusion that these simulated environments could be
successfully used to help people with ASD to engage in social interactions.
Virtual avatars with emotional facial expressions were used in project Vir-
tual Messenger [26]. The goal was to help users to relate emotions to facial
expressions. In the first stage, the user would be asked to select the emotion
corresponding to the avatar presented on screen, or to select a facial expression
that corresponds to an emotion. The second level requires the user to select an
emotion that is appropriate for a scenario presented on screen. When in the third
level, the user is presented with an avatar’s facial expression, and is required to
select a scenario that could cause the avatar’s expression. The results show that
people with ASD were able to increase their level of understanding of emotions
based on facial expressions.
Joint attention is one key issue in ASD. According to [27], joint attention is
a social interaction in which two people use gestures and gaze to share atten-
tion with respect to a third object or element of interest. Joint attention seems
to be related to communication as a requirement for its development. Interven-
tions that rehabilitate joint attention have shown correlated improvements in the
communication skills of the children, such as social initiations, positive affect,
imitation, play, and spontaneous speech [28]. In [29] a system that combines
VR simulation with an EEG data acquisition to evaluate the cognitive detec-
tion of joint attention stimuli was proposed. Two tasks to explore joint atten-
tion were developed: to identify joint attention clue; and follow joint attention
clue. Although the work is being developed considering people with ASD as the
target, a preliminary evaluation was performed with four people with no disor-
ders for proof-of-concept with encouraging results presenting 90 % classification
accuracy.
Mobis is a mobile AR system that aims to redirect the attention of children to
the objects used during therapy [30]. The system superimposes digital contents,
including text and visual shapes, on top of physical objects. The evaluation
was taken with seven teachers and twelve low-functioning children with autism,
between the ages of 3 and 8. Children rapidly learn how to manipulate the mobile
devices and used Mobis to explore their environment. The results indicate that
384 C.A. da Silva et al.

low-function children with autism can use AR technologies and a mobile device
as a visor, and that Mobis helped children to discriminate between and identify
new objects. Results also lead to the conclusion that Mobis has a positive effect
on the time the child remains on tasks, including when distractions are present,
and increased student engagement with people and objects.
In [9] a system based on AR with infrared markers was described. The sys-
tem superimposes the card symbols on the user’s image in real time aiming
at enhancing symbol understanding. However, this prototype was never tested
on children with ASD. Later, in [3] the Pictogram Room, a AR system using
Microsoft Kinect, was proposed having as main goals teaching self-consciousness,
body scheme and postures, communication and imitation, using serious games
to promote interaction with the children.
An AR marker-base system that substitutes real objects by virtual/imaginary
objects was proposed in [31] to promote pretend play based on common practices.
The virtual objects simulate appropriate behaviours when the real objects are
moved, like rotating wheels. Furthermore, a set of theme-related augmentations
are added into scene to reinforce pretence and encourage situational appropriate
responses. The same set of real objects can be replaced by different pretend
virtual objects to promote flexibility in pretend play. This work is still at an
early stage and therefore no testing with children with ASD has been performed.
Tentori and Hayes proposed the Mobile Social Compass framework for the
development of mobile AR systems. The Social Compass Curriculum is a behav-
ioural and educational curriculum for social skills training, based on stories and
visual paper clues to guide the child in both active and passive social interac-
tions. The framework focuses in the concept of Interaction Immediacy, providing
a set of visual clues to assist the child in anticipating situations [32].
MOSOCO is a mobile application implemented using the above described
framework, using AR and visual guidance, to assist children with ASD in social
interactions in real life. The system guides children in the basic social skills
defined in The Social Compass curriculum, encouraging them through interac-
tive resources. The systems evaluation was performed with three children and
the results show an increased social interactions, both in quantity and quality,
mitigating behavioural and social disruptions. The system also promoted social
integration of these children in neurotypical groups.
Another mobile AR application, Blue’s Clues, was proposed in [33], to help
children with ASD moving from one place to another. Blue’s Clues concept was
developed for a school environment, aiming at providing the necessary mobility
instructions, visual and audio clues, to guide students with ASD.
In [34] the authors have created an AR application where students handle plant
entities such as fruits leaves, flowers and seeds. Students are required to match
a template page where several fruits are presented with all the above items. The
system uses AR to match 3D and 2D plant entities, and provides visual cues, sur-
rounding the entities with red or blue circles as feedback to show if the markers
are correctly placed. Auditory and olfactory cues are also provided. The study was
performed with a large group of children including 11 cognitive disabled children,
 STAR: Speech Therapy with Augmented Reality 385

some of them with ASD. According to the authors, all the children, including dis-
abled, easily handled the markers and had no difficulty with technical elements.
Also, the perception and understanding of the 2D and 3D virtual entities, was effec-
tive for all children including the disabled ones. During evaluation the authors have
observed that disabled children were very enthusiastic when using the application
and showed a high motivation compared to most other pupils. Moreover, autistic
and trisomic children were able to express some positive emotions when confronted
to the application.
Artifact-AR [35] is an AR cognitive artifact for retraining and cognitive skill
development, that explores identification, memorization, comparison, and asso-
ciation of pictures and sounds. The user interacts with markers to perform the
tasks. To take full advantage of this system some sort of VR glasses are required.
Otherwise, the user must interact with the physical device, while the system
responds on a different location (the computer screen). This implies that the user
must be concentrated on the physical device to point at the right areas, while
at the same time checking the computer screen to see the system’s response.
The customization of the application can be done adding or removing sound
and images in a specific directory and edit some of the of the configuration files,
hence, the user doesn’t need programming skills to expand or personalize the
application. The system was tested by ten therapists to analysing the usability
aspects with encouraging results regarding its usability.

4 STAR’s Architecture

In this section we propose an AR based system that allows the creation of interac-
tive activities based on behaviour principles to assist interventions with children
with ASD.
STAR (Speech Therapy with Augmented Reality) design took the needs and
characteristics of children in ASD into account. We held meetings and applied
questionnaires to parents and experts in order to understand the activities under-
taken with children with ASD and identify the potential applications of AR
technology in interventions.
As result of this process, and literature review, we defined the requirements
of the system, namely providing:

– A broad enough range of AR activities templates to avoid repeating the same

task over and over again;
– A GUI activity manager where the therapist can define activities and system
behaviours according to needs and characteristics of each child. The therapist
does not have to be a programmer nor edit text or configuration files;
– Data generated from the sessions should be gathered and stored in a database
containing the history for each child;
– Video recording and annotation;
– Parent participation.
386 C.A. da Silva et al.

The AR subsystem and the GUI activity manager will be described in more
detail in Sect. 5.
Data gathered during the AR sessions can be of assistance to the therapist
to plan new sessions, namely by recalling the session history, and specific details
of each session, which activities were performed, for how long, where the child
had more difficulties, among other items.
While most of these items can be found abundantly in the literature, having
participating parents is not so common. The system would allow the parents,
in their own homes, to see the videos of the interventions by the therapist,
thereby being more aware of the work being developed with their children. The
video annotation becomes extremely useful as the therapist can use this resource
to provide parents with additional information, or direct their attention to a
particular detail on a session.
Session viewing, together with therapist coaching, could ultimately enable
parents to perform sessions with their children, acting as co-therapists. Partici-
pating parents, acting as co-therapists, has been reported in [36–38]. In [38], the
authors claim that parent training has been shown to be a very effective method
for promoting generalization and maintenance of skills in children with autism.
These sessions, between a parent and the child, could be streamed live to the
therapist who could guide the parents as the session progresses, or viewed later
and annotated by the therapist as feedback for the parents. This dual feedback
system, where parents access videos recorded in the therapist own sessions, and
the possibility of real time guidance, or at least feedback from the therapist
based on the parents sessions could provide a richer experience to everybody,
strengthening bonds between therapist, parents, and most importantly, the child.
The GUI activity manager would allow the therapist to elaborate and manage
activities. The therapist could then select a subset of these activities and make
them available to the parent’s software client, so they can act as co-therapists.
Data gathered during the parent-child session would then be sent back to the
central system, registered under the child’s history, and made available to the
therapist.
This proposal aims to fulfil the main characteristics of successful behaviour
intervention in ASD [39]:

– Training in various settings (house, school, clinics);

– Multiples facilitators (parents, teachers, therapists);
– Prompting, reinforcement and fading;
– Transparency and predicable structure of training/environment.

Note, however, that the system should only be seen as a facilitator, not as an
enforcer. The therapist should always be in charge and it is up to the therapist
to decide, on a child by child basis, whether, and which of these features should
be made available to the parents.
 STAR: Speech Therapy with Augmented Reality 387

5 The Prototype
In this section we propose an AR based system that allows the creation of inter-
active activities supported on AAC and ABA principles to assist interventions
with children with ASD, aided by a therapist. This system has been implemented
and tested with children with ASD, as detailed in Sect. 6.
The goal was not to impose a new methodology to the therapist, but to offer
a supporting tool that could enrich the current daily activities in the speech-
language sessions. Hence, all the development was initially discussed with speech-
language therapists, to ensure that the prototype would fit painlessly in the
current activities, and that its interface was clear and intuitive for both the
therapist and the children.
The prototype allows the design of activities for each individual child attend-
ing to the difference in degree of the ASD, the individual personality, and the
previous activities developed by the therapist. Within each activity, we aimed at
providing as much flexibility as possible. Several templates were provided for this
initial evaluation to guide the therapist in designing the individual activities.
The usage of AR is directed towards superimposing 3D models and anima-
tions on cards with communication symbols that the child holds or places on the
table. This creates a richer experience, and takes advantage of the visual skills
found in many children with ASD.
The selection of AR as the underlying technology for this project aims at:

– Promote a natural interaction with the computer through the use of tangible
interfaces;
– Facilitate the generalization of the acquired knowledge;
– Allow the enrichment of traditional interventions based on communication
symbols without imposing a new methodology on the therapy;
– Eliminate the distractions caused by traditional interfaces such as mouse and
keyboard.

In [40] is suggested that for the use of tangible interfaces to be intuitive and
natural, it is recommended that the selection of physical objects and interaction
metaphors should be familiar to users, allowing them to interact with the system
relying on previous skills and experiences. In our project we use communication
symbol cards which are already familiar to the children. The child interfaces with
the system by showing cards with communication symbols, which is a natural
behaviour for the child under intervention. Furthermore, the acquired knowledge
is implicitly associated with the communication symbols, hence, the child should
be able to recall its experiences afterwards.
The prototype was built with open software tools and APIS such as irrKlang
for audio, ARToolkit for the AR component and MySQL. These options aim at
providing a low cost solution, and multi-platform future support.
388 C.A. da Silva et al.

5.1 GUI Activity Manager

The activity construction process takes place in a simple support application

where the therapist can select a template and design interactive activities to
explore with the children.
Initially the therapist associates words with symbol cards, 3D graphics,
sounds and animations, and builds phrases with word sequences. The words
can be further defined as belonging to user-defined categories such as actions,
fruits, or animals.
The data set of words, phrases, and their associated media, is then used to
construct the activities based on the available templates. Each template provides
full parameterization including reinforcement definition, and error management.
Activity parameterization includes:

– Interaction procedure: choose between display or occlude markers to interact

with application;
– Interaction events behaviour: choose from (a) superimpose a virtual object,
(b) reproduce an audio (c) both or (d) none of them;
– Audio and visual reinforcements: In “beginning of activity”, “success on task”
and other events, the therapist can select to reproduce an audio file and/or a
selectable colour square to surround the mark;
– Sequence management: define when to proceed to next task in activity. Can
choose from (a) proceed only on attempt success; (b) proceed even on attempt
fails; or (c) proceed on success or after a configurable number of attempts.

5.2 Activity Templates

The available templates are: Free, Category, Discriminate, and Phrase. These
are based on stages I (Free), III (Category and Discriminate), and IV (Phrase)
of PECS. The possibility to use different templates also seeks to avoid having
the autistic child engaging in repetitive tasks for long periods of time.
The basic procedure consists in showing a card in the field of view of the web
cam. The recognition of the pattern in the card will trigger an action.
The Free template, the simplest one, allows both, child and therapist, to
interact freely with the prototype. When designing an activity with this template
the therapist can associate the action to be triggered when the card is shown or
occluded. This action, which can be a graphic, an animation, or a sound, can
act as a positive reinforcement depending on the intervention context, or just
be used to engage the child into exploring the system. Additional actions can
be specified when the card is occluded. As an example assume an activity based
on cards with animals. When the child shows a card the system may display on
screen a 3D model of the animal on top of the card. If the child covers the card
with the hand then the animal sound can be played.
In the Discriminate template the child is required to identify an object based
on a clue provided on screen, and present the card that matches the clue. Pos-
itive and negative reinforcements can be associated with each individual clue.
 STAR: Speech Therapy with Augmented Reality 389

The system supports the configuration for special cards that the therapist can
include in the activity. A ‘hint’ card will trigger a user-defined hint when shown
to the system. A ‘next’ card tells the system to move on to the next object.
Finally the ‘end’ card will end the activity. These cards, if provided by the ther-
apist, allow for a freer interaction from the child, since the child can control
the activity when possessing all three special cards, or some subset. The special
cards can also be used by the therapist to control the activity.
The Category template can be configured in two different ways. Initially the
system will present a hint on screen, namely a graphic or sound. The child is then
requested to select and show a card either with the respective category or with
an item in the same category. Positive reinforcements, applied when the answer
is correct, are configurable, as well as the actions to be triggered for showing the
wrong card.
In the Phrase template the child is requested to show a sequence of cards,
which represent the subject, verb and noun in a phrase, see Fig. 1 for an example
using symbols from the ARASAAC collection (available at http://www.catedu.
es/arasaac/descargas.php). When each card is placed on the table a reinforce-
ment can be triggered by the system. This reinforcement will hint the child if
the right card is being used, and also if the card is in the correct position within
the phrase. The system will detect when a phrase is complete and trigger the
user-defined positive reinforcement.

Fig. 1. An example of a phrase

6 Testing the Protopype

The test was performed in an association that supports children with ASD, the
Associação para a Inclusão e Apoio ao Autista, located in Braga, Portugal
(http://www.aia.org.pt/).
The prototype was tested by a speech-language therapist, the third author,
that performs interventions regularly with children, and four children diagnosed
390 C.A. da Silva et al.

with ASD. The goal was to test if the software was clear to the therapist, and
to evaluate its usage as a supporting tool in the interventions performed.
The therapist had 13 sessions with four boys diagnosed with ASD. The chil-
dren were between the ages of 6 and 10, had some level of orality, and attended
school.
Some children were previously assessed as participative in the speech-
language sessions, while others have an extremely passive behaviour and barely
engaged in spontaneous interaction and communication with the therapist. One
child had several verbal and motor stereotypies and expressed himself mainly
through echolalia.
A qualitative study was performed with data gathered through observation
by the team, including the therapist, in order to evaluate the suitability of the
software usage by the children, namely what behaviours were triggered by the
usage of AR, and if the children had benefited from the interaction with the
software. To support the study we recorded the sessions in video, and filled
observation grids with the interaction stages as well as observations that emerged
in the sessions with the children (Fig. 2).

Fig. 2. Children interacting with the prototype

6.1 Strategies

Based on the previous history with the children, as well as the existing interven-
tion plan for each child, the therapist set the following scripts for these sessions.
For three children the goal was set to: (1) identify animals, nutrients, and
daily objects; (2) Identify and designate gender; (3) Identify and designate
actions; (4) Discriminate and identify sounds; (5) Answer yes/no questions,
 STAR: Speech Therapy with Augmented Reality 391

together with the respective head nod or shake; and (6) Build sentences with
subject, verb and noun.
The other child had as goals to: (1) Enhance declarative skills; (2) Improve
language skills; and (3) Create and imagine using language.
The therapist created six activities based on the four available templates to
achieve these goals.

6.2 Evaluation
In general the children were highly motivated by the usage of AR. Superimposed
3D models and computer animations on top of cards displaying familiar symbols
caused a higher degree of engagement in the activities, and had the children
requesting the therapist to ‘play’ with the software when arriving.
The child with stereotypies, when in traditional interventions, used to keep
asking to go to the window to watch for cars passing by, and kept repeating
out of context phrases. After the AR based session had ended, when they were
ready to start a different activity (not related to this project), the child asked
the therapist for the dog marker and autonomously presented it to the system.
A joyfully expression emerged when the dog’s bark was heard.
A child defined by the therapist as passive, regarding communication and
behaviour, revealed to be highly motivated, repeating everything the computer
‘said’.
Another child, with high functioning ASD, offered an initial resistance to
any activity that didn’t involve creating stories. Since he had an interest in
animals, the therapist started to use the animal markers, and asked the child
to show one to the computer. When the system superimposed the 3D model of
the animal on the card the child was holding, the child became very interested
in the application. In particular, the child showed more enthusiasm for markers
associated with 3D models than those associated with sounds only. The therapist
took advantage of the sudden child’s interest and suggested that he could create
a story with the markers. The child selected only markers which were associated
with 3D models. During the session the child created two stories, one with insects,
and another with sea animals using the 3D models as characters. When the
session was over, the child asked for a photo of him playing with the animals to
show to his father.
The activities were easily grasped by the children, possibly due to the use
of tangible interfaces based on the cards with symbols, something which was
highly familiar to these children. The prototype did not require any new form
of communication, showing cards was something that was already present in the
sessions prior to this test. Hence, the interaction with the computer presented
itself as familiar to the child.
After 13 sessions the therapist reported her impressions. On a positive note,
the therapist stated that the children kept the interest on the activities for a
longer period of time, and inclusively that children had the initiative to ask
and use the software on their own. The children acquired new vocabulary and
consistently managed to perform correctly the proposed activities.
392 C.A. da Silva et al.

However, the therapist also noted that a particular child would sometimes
disregard everything else concentrating only on his image on the screen. The joint
attention was not as high and eye contact occurred less often, suggesting that
a balance must be achieved between the several different types of interventions
and activities.
The therapist concluded that an AR based approach has the potential to
present benefits in speech-language therapy although the social area may be
neglected with some children.

6.3 Detected Issues

Some issues came up during the test which are worth mentioning in this report.
First we had some situations where a card would be incorrectly recognized,
triggering an inappropriate action. Other times an action was triggered without
having any card on display. These issues came up due to the variations in lighting
during the day.
One child actually enjoyed the fact that the system made ‘mistakes’ and was
eager to correct these mistakes. The therapist also took these system misclas-
sifications and explored them within the intervention leading to an even more
engaging session. These unintended recognition failures must be fixed, as children
with more rigid cognitive processes may be unable to deal with these situations.
Nevertheless, this obviously leads to the question: should the system fail occa-
sionally on purpose? Should this feature be introduced in the configuration?
On another session when the system triggered a sound without a card being
shown, the child immediately sought for the associated card and showed it to
the system. Again, while these unattended failures must be avoided, they may
become part of an activity with some children, introducing an element of surprise
and breaking the pattern of the activity.
The team also noted that the Phrase template needed to be redesigned to
include variations and more reinforcements. The system should be able to provide
hints regarding the type of the missing elements and their relative position.
Furthermore, the system should allow reinforcements when, although all the
correct cards are in the table, the order is not correct. Another option to explore
is to start with incomplete phrases and ask the child to identify the type of the
missing element, as well as the specific missing card.
The last issue is related to the childs interaction when motor stereotypies are
present. In our study a child needed assistance from the therapist to be able to
show the cards. This can be solved in the context of the application if the cards
are laid on the table. Nevertheless, this issue also suggests a need for a different
form of interaction.

7 Conclusions

The present work has proposed an approach, relating elements from AAC and
ABA, to apply AR in communication and language interventions with children
 STAR: Speech Therapy with Augmented Reality 393

with ASD to mitigate the issues caused by the impairments in communication

skills and imagination. We exploited visual and auditory sensory stimulation,
combining immediate reinforcement and enrichment of the environment with
virtual objects and audio, in order to attract and motivate children.
The resulting prototype has been qualitatively evaluated in an association
that supports children with ASD and has therapists conducting interventions in
many areas. The sessions were conducted by a therapist and the behaviours of
the children and their interaction with the system was recorded on video and
using observation grids.
A common issue in qualitatively evaluations relates to the generalization of
results to the remaining population. We have no intention of suggesting that the
data presented in here can be generalized. This is particularly relevant in the
case of ASD due to the wide range of disorders and their varying degrees. This
study presents only a contribution to the development of technological systems
to support interventions.
Our study shows that the usage of information technologies, in particular AR
based systems, in speech-language interventions, complementing and supporting
the traditional approaches is an option that must be further explored. The results
suggest that the system usage could represent an added value since we could
observe greater motivation, word acquisitions, and commitment in performing
correctly, as observed by the therapist.
Due to the use of AR and tangible interfaces, familiar physical objects, and an
interaction procedure similar to the used in previous AAC sessions, where cards
are exchanged between child and therapist, the system was easily grasped by
the children. As expected, the children were able to use their past experience to
interact effortlessly with the system and hopefully they will be able to generalize
the acquired knowledge.
We believe that the positive results obtained in this study are, to some extent,
due to the engaging environment provided by the application, and its tangible
interface. These factors stimulate cognitive processing and lead to visual learning.
These results illustrate the potential of AR systems to be used in linguistic
skill development due to the significant increase in interaction and communica-
tion initiated by the children.
While technologies such as the ones in our prototype can provide very pos-
itive results, we must pay particular attention to avoid having the child com-
pletely immersed on some application details and neglecting social interaction.
The therapist plays a fundamental role in these ‘augmented’ interventions.

7.1 Future Work

We believe that the results presented show the need for deeper and further
studies, adding more interaction templates, extending their configuration, and
testing the system with more children.
Game oriented activity templates could provide an extra level of engage-
ment for the children. Making use of the markers as characters, one child was
394 C.A. da Silva et al.

encouraged to telling stories. Templates with more customizations options could

provide a suitable environment to training symbolic play.
Adding templates should be explored to increase the range of possibilities for
the therapist to conduct the interventions. The phrase template, as mentioned
before, would also benefit from a subdivision or at least more configuration
options.
The card recognition subsystem also needs to be fine-tuned to increase
robustness, hence suppressing unintentional card misclassification. These system
errors, due to the way children reacted to them, raise an interesting question that
we believe is worth pursuing: Would a system that intentional provides a wrong
answer be a way of further engaging at least some children? What is a ‘good’
wrong answer?
Children with motor stereotypies also require a different interaction proce-
dure. This condition makes it harder for the system to correctly classify the cards
when the child is holding the card, requiring the card to be placed on the table.
This is an important issue since a significant number of children with ASD suffer
from motor stereotypies.

In this study we focused on speech-language interventions, nevertheless, the

designed approach, combining behavioural methods and AR features, could make
this system suitable to support other interventions, such as sensorial and motor
based intervention or even music therapy.
Finally, the limited number of participating children does not reflect the
broad range of the spectrum, nor do they allow for any generalization in a par-
ticular subset of these disorders. While the generalization for the whole spectrum
of disorders is an impossible task, more participating children would allow for a
more relevant analysis, regarding the significance of the contribution of a system
such as the one presented in here when used in interventions.

Acknowledgements. Work partially funded by National Funds through FCT -

Fundação para a Ciência e a Tecnologia (Portuguese Foundation for Science and Tech-
nology) within project UID/CEC/00319/2013 and by Conselho Nacional de Desen-
volvimento Cientı́fico e Tecnológico (CNPq-Brazilian National Council for Scientific
and Technological Development) within program Ciência sem Fronteiras (Science with-
out Frontiers), grant 225248/2012-3.

References
1. Klin, A.: Autism and asperger syndrome: an overview. Rev. Bras. Psiquiatr. 28,
S3–11 (2006)
2. Committee on Educational Interventions for Children with Autism, N.R.C.: Edu-
cating Children with Autism. The National Academies Press (2001)
3. Herrera, G., Casas, X., Sevilla, J., Luis, R., Pardo, C., Plaza, J.: Pictogram room:
natural interaction technologies to aid in the development of children with autism.
Annuary Clin. Health Psychol. 8, 39–44 (2012)
 STAR: Speech Therapy with Augmented Reality 395

4. Avila, B.G.: Comunicao Aumentativa e Alternativa para o Desenvolvimento da

Oralidade de Pessoas com Autismo. Master’s thesis, Universidade Federal do Rio
Grande do Sul, Porto Alegre, Brasil (2011)
5. Moore, D., McGrath, P., Thorpe, J.: Computer-aided learning for people with
autism; a framework for research and development. Innovations Educ. Training
Int. 37, 218–228 (2000)
6. Moore, M., Calvert, S.: Brief report: vocabulary acquisition for children with
autism: teacher or computer instruction. J. Autism Dev. Disord. 30, 359–362 (2000)
7. Goldsmith, T.R., LeBlanc, L.A.: Use of technology in interventions for children
with autism. J. Early Intensive Behav. Interv. 1, 166–178 (2004)
8. Raffle, H., Farr, W., Yuill, N.: Collaborative benefits of a tangible interface for
autistic children. ACM SIGCHI (2009)
9. Herrera, G., Jordan, R., Gimeno, J.: Exploring the advantages of augmented reality
for intervention in asd. In: Autism Safari 2006–2nd Word Autism Congress (2006)
10. Grandin, T.: Thinking in Pictures. Bloomsbury Publishing, London (2009)
11. Wing, L.: The Autistic Spectrum: A Guide for Parents and Professionals. Constable
and Robinson, London (2002)
12. Centers for Disease Control and Prevention: Autism spectrum disorders (2013).
Accessed 15 October 2013. http://www.cdc.gov/ncbddd/autism/data.html
13. Strock, M.: Autism spectrum disorders (pervasive developmental disorders).
National Institute of Mental Health (NIMH) (2007)
14. Wing, L.: The continuum of autistic characteristics. In: Schopler, E., Mesibov, G.B.
(eds.) Diagnosis and Assessment in Autism. Current Issues in Autism, pp. 91–110.
Springer, Heidelberg (1988)
15. Myers, S.M., Johnson, C.P.: The council on children with disabilities: management
of children with autism spectrum disorders. Pediatr. 120, 1162–1182 (2007)
16. Gadia, C.A., Tuchman, R., Rotta, N.T.: Autism and pervasive developmental dis-
orders. J. Pediatr. 80, 83–94 (2004)
17. Smith, T.: Discrete trial training in the treatment of autism. Focus Autism other
Dev. Disabil. 16, 86–92 (2001)
18. Hobson, R.P.: Autism and the Development of Mind (Essays in Developmental
Psychology). Psychology Press, UK (1995)
19. Sevcik, R., Romski, M.: A.A.C.: More Than Three Decades of Growth and Develop-
ment (2000). http://www.asha.org/public/speech/disorders/AACThreeDecades/.
Accessed: 15 October 2013
20. Bondy, A.S., Frost, L.A.: The picture exchange communication system. Focus
Autism Other Dev. Disabil. 9, 1–19 (1994)
21. Charlop-Christy, M.H., Carpenter, M., Le, L., LeBlanc, L.A., Kellet, K.: Using
the picture exchange communication system (pecs) with children with autism:
assessment of pecs acquisition, speech, social-communicative behavior, and prob-
lem behavior. J. Appl. Behav. Anal. 35, 213–231 (2002)
22. Kravits, T.R., Kamps, D.M., Kemmerer, K., Potucek, J.: Brief report: increasing
communication skills for an elementary-aged student with autism using the picture
exchange communication system. J. Autism Dev. Disord. 32, 225–230 (2002)
23. Ganz, J.B., Simpson, R.L.: Effects on communicative requesting and speech devel-
opment of the picture exchange communication system in children with character-
istics of autism. J. Autism Dev. Disord. 34, 395–409 (2004)
24. Belen, S.O.: INMER-II: Sistema de Inmersin en Realidad Virtual para Personas
con Autismo. Tecnoneet 2004: Retos y realidades de la Inclusin Digital (2004)
396 C.A. da Silva et al.

25. Parsons, S., Leonard, A., Mitchell, P.: Virtual environments for social skills train-
ing: comments from two adolescents with autistic spectrum disorder. Comput.
Educ. 47, 186–206 (2006)
26. Fabri, M., Moore, D.: The use of emotionally expressive avatars in collaborative
virtual environments. In: Virtual Social Agents, p. 88 (2005)
27. Charman, T.: Why is joint attention a pivotal skill in autism? philosophical trans-
actions of the royal society of London. Ser. B: Biol. Sci. 358, 315–324 (2003)
28. Kasari, C., Freeman, S., Paparella, T.: Joint attention and symbolic play in young
children with autism: a randomized controlled intervention study. J. Child Psychol.
Psychiatry 47, 611–620 (2006)
29. Simões, M., Carvalho, P., Castelo-Branco, M.: Virtual reality and brain-computer
interface for joint-attention training in autism. In: Proceedings 9th International
Conference Disability, Virtual Reality and Associated Technologies, pp. 507–510
(2012)
30. Escobedo, L., Tentori, M., Quintana, E., Favela, J., Garcia-Rosas, D.: Using aug-
mented reality to help children with autism stay focused. IEEE Pervasive Comput.
13, 38–46 (2014)
31. Bai, Z., Blackwell, A.F., Coulouris, G.: Making pretense visible and graspable: An
augmented reality approach to promote pretend play. In: 11th IEEE International
Symposium on Mixed and Augmented Reality, ISMAR 2012, Atlanta, GA, USA,
5–8 November 2012. IEEE Computer Society (2012)
32. Tentori, M., Hayes, G.R.: Designing for interaction immediacy to enhance social
skills of children with autism. In: Proceedings of the 12th ACM International Con-
ference on Ubiquitous Computing, pp. 51–60. ACM (2010)
33. Escobedo, L., Tentori, M.: Blues clues: An augmented reality positioning system.
In: Child Computer Interaction Workshop (CHI) (2011)
34. Richard, E., Billaudeau, V., Richard, P., Gaudin, G.: Augmented reality for reha-
bilitation of cognitive disabled children: a preliminary study. In: Virtual Rehabili-
tation, pp. 102–108. IEEE (2007)
35. Kirner, C., Kirner, T. G.: Development of an interactive artifact for cognitive
rehabilitation based on augmented reality. In: 2011 International Conference on
Virtual Rehabilitation (ICVR), pp. 1–7. IEEE (2011)
36. Schopler, E., Reichler, R.J.: Parents as cotherapists in the treatment of psychotic
children. J. Autism Child. Schizophr. 1, 87–102 (1971)
37. Krantz, P.J., MacDuff, M.T., McClannahan, L.E.: Programming participation in
family activities for children with autism: parents’use of photographic activity
schedules. J. Appl. Behav. Anal. 26, 137–138 (1993)
38. Ingersoll, B., Dvortcsak, A.: Including parent training in the early childhood spe-
cial education curriculum for children with autism spectrum disorders. J. Positive
Behav. Interv. 8, 79–87 (2006)
39. Bölte, S.: Computer-based intervention in autism spectrum disorders. Focus
Autism Res. 9, 247–260 (2005)
40. Billinghurst, M., Grasset, R., Looser, J.: Designing augmented reality interfaces.
SIGGRAPH Comput. Graph. 39, 17–22 (2005)
 Interpretation of Construction Patterns for Biodiversity
Spreadsheets

Ivelize Rocha Bernardo1 ✉ , Michela Borges2, Maria Cecília Calani Baranauskas1,

( )

and André Santanchè1

1
Institute of Computing - Unicamp, Avenida Albert Einstein, 1251, Cidade Universitária,
Campinas, Brazil
{ivelize,cecilia,santanche}@ic.unicamp.br
2
Institute of Biology – Unicamp, Zoology Museum, Rua Charles Darwin,
s/n, Cidade Universitária, Campinas, Brazil
[email protected]

Abstract. Spreadsheets are widely adopted as “popular databases”, where

authors shape their solutions interactively. Although spreadsheets are easily
adaptable by the author, their informal schemas cannot be automatically inter‐
preted by machines to integrate data across independent spreadsheets. In biology,
we observed a significant amount of biodiversity data in spreadsheets treated as
isolated entities with different tabular organizations, but with high potential for
data articulation. In order to automatically interpret these spreadsheets we exploit
construction patterns followed by users in the biodiversity domain. This paper
details evidences of such patterns and how they can lead to characterize the nature
of a spreadsheet, as well as, its fields in a domain. It combines an automatic
analysis of thousands of spreadsheets, collected on the Web, with results from a
survey conducted with biologists. We propose a representation model to be used
in automatic interpretation systems that captures these patterns.

Keywords: Pattern recognition · Spreadsheet interpretation · Semantic mapping ·

Biodiversity data integration

1 Introduction

When producing spreadsheets, end-users have autonomy and freedom to create their
own systematization structures, with few formal requirements. However, the product is
driven to human reading, causing a side effect: programs provide poor assistance in
performing tasks, since they are unable to recognize the spreadsheet structure and to
discern its implicit schema – hidden in the tabular organization – from the instances and
consequently the semantics of this schema. Therefore, it is difficult to combine and
coordinate data among spreadsheets using conventional methods, because each new
different schema cannot be interpreted.
But, how much different they are in fact? We present in this paper evidences that
similarities in spreadsheets can indicate patters followed by groups. We consider that it
is possible to map these patterns to a respective semantic description, through the

© Springer International Publishing Switzerland 2015

J. Cordeiro et al. (Eds.): ICEIS 2014, LNBIP 227, pp. 397–414, 2015.
DOI: 10.1007/978-3-319-22348-3_22
398 I.R. Bernardo et al.

recognition of structural reasons which leads a user to interpret a spreadsheet in one way
and not another.
Thus, our strategy focuses on the detection of patterns to recognize similar spread‐
sheets. We argue that the specific way authors build their spreadsheets – i.e. the criterion
to define elements, the approach to spatially organize them and the relationship between
these elements – is directly related their daily experience in the community they belong.
The challenge of this research is to consider a computer system as a consumer of
spreadsheets besides the user. Our approach involves achieving a richer semantic inter‐
operability for data from spreadsheets through pattern recognition.
Most of the related work disregard these patterns to implement strategies for seeking
interoperability of tabular data. This paper argues that the structure, i.e. the organization
of spreadsheet elements, must be considered, since it leads to the identification of
construction patterns, which is related to the user intention/action. This technique allows
us to go towards the pragmatic interoperability layer [1].
This article is an extension of a previous work [2], in which we showed evidences
of spreadsheet construction patterns adopted by biologists based on an application
implemented by us, able to automatically recognize these patterns in the implicit spread‐
sheet schemas. To support our thesis we collected and analyzed approximately 11,000
spreadsheets belonging to the biodiversity domain.
In this paper we confront previous results with an exploratory observation of these
construction patterns based on a survey answered by 44 biologists.
The next sections are organized as follows: Sect. 2 gives an overview of some basic
concepts and our research, Sect. 3 details the process of collecting and analyzing spread‐
sheets employed by biologists, as well as research hypotheses and their evaluation;
Sect. 4 highlights evidences of construction patterns followed by biologists; Sect. 5
introduces our model to represent construction patterns; Sect. 6 presents Related Work
and Sect. 7 our concluding remarks and the next steps of this research.

2 Research Scenario

According to Syed et al. [3], a large amount of the information available in the world is
represented in spreadsheets. Despite their flexibility, spreadsheets were designed for
independent and isolated use, and are not easily articulated with data from other spread‐
sheets /files.
For this reason, there is a growing concern to make spreadsheet data more apt to be
shared and integrated. The main strategies convert them into open standards to allow
software to interpret, combine and link spreadsheet data [4–10].
Related work address this problem mainly by manual mapping to Semantic Web
open standards or by automatic recognition, relating spreadsheet elements to concepts
available on Web knowledge bases such as DBpedia (http://dbpedia.org).
Systematic approaches for data storage, such as databases, predefine explicit
schemas to record data. These schemas can be considered as semantic metadata for the
stored data. Spreadsheets, on the other hand, have implicit schemas, i.e. metadata and
data merged in the same tabular space.
 Interpretation of Construction Patterns for Biodiversity Spreadsheets 399

The central thesis behind our approach is that we can detect and interpret the spread‐
sheet’s schema by looking for construction patterns shared by research groups. We
propose in this paper a representation model able to capture such patterns, as well as to
be processed by machines. Results of our analysis in thousands of spreadsheets and in
a survey indicate the existence of such recurrent patterns and that they can be exploited
to recognize implicit schemas in spreadsheets.
There are several aspects that hinder the spreadsheet recognition and its implicit
schema, such as differences between columns order, the label used to identify fields and
their respective semantics etc.
Although related work explore a subset of the common practices in tabular data –
sometimes taking into account their context [11–13] – they do not define a mechanism
or model to independently represent these patterns. Since the knowledge about how to
recognize patterns is mixed with the programs, they cannot be decoupled from their
code. We claim here that a representation to materialize the knowledge about these
patterns as artefacts, independently of specific programs and platforms, enables to share,
reuse, refine and expand such patterns among users and applications.
This research is driven by a larger project that involves cooperation with biologists
to build biodiversity bases. We observed that biologists maintain a significant portion
of their data in spreadsheets and, for this reason, this research adopted the context of
biology as its specific focus.
We propose a model to represent construction patterns, departing from observations
conducted through incremental steps, including spreadsheets collecting/catalog, formu‐
lating hypotheses/models, exploratory observation of biology survey and evaluation.
These steps will be detailed in next sections.

3 Methodology

As previously mentioned, our approach to represent construction patterns was based on

a study of related work and field research in the biology domain.
Based on an initial analysis of how biologists of the Institute of Biology (IB) at
Unicamp created their spreadsheets, we defined a set of hypothesis and designed experi‐
ments and a survey to validate them. These data were also the basis to produce our first
model adopted in a process to automate the recognition of construction patterns, whose
design involved:
(i) preliminary collection and analysis of spreadsheet data: in a first moment, we
discussed with biologists of the Institute of Biology (IB) at Unicamp about the
practices applied when they work with spreadsheets, starting from their creation
until their reuse and manipulation;
(ii) formulation of hypotheses about spreadsheet construction patterns: departing from
the spreadsheets, we performed a visual analysis looking for commons elements
able to be identified by a system;
(iii) design and implementation of an automatic recognizer for these spreadsheets: we
implemented a systems – the SciSpread [2] – which was validated in three progres‐
sive groups of spreadsheets: 9 spreadsheets belonging to IB at Unicamp; 40
400 I.R. Bernardo et al.

spreadsheets from the Web selected manually; 11,000 spreadsheets from the Web
automatically selected by a crawler script;
(iv) exploratory observation of biology survey: we deployed an online form (available
in goo.gl/b1iEvl) to biology students, professors, researchers and professionals, in
order to verify according our hypothesis – formulated in (ii) – how biologists clas‐
sify and arrange their data. This step plus step (iii) were designed to evaluate and
refine the hypothesis. The results of both steps are confronted in the next section.

3.1 Initial Data Collection and Analysis

Our analysis started with 9 spreadsheets belonging to the IB, in which we identified
two main construction patterns, related to the nature of the spreadsheet: catalogs of
objects – e.g., specimens in a museum – and event related spreadsheets, e.g., a log of
samples collected in the field. We further will refer to these spreadsheet natures as
catalog and event.
In order to address the significant differences among spreadsheet types we classified
each field in six exploratory questions (who, what, where, when, why, how) [14]. It
enabled us to represent and recognize patterns in a higher level of abstraction, e.g., a
catalog spreadsheet has as initial fields the taxonomic identification – classified as what
question – on the other hand, a collection spreadsheet has as initial fields: date and
locality – classified as when and where questions, as illustrated in Fig. 1.

Fig. 1. Fields characterization.

The next step involved collecting more 33 spreadsheets on the Web to compose our
sample. To search spreadsheets belonging to the biology domain, we applied domain
related keywords as criterion.

3.2 Hypothesis

According to the observation of these spreadsheets, we proposed the following pattern-

related hypotheses:
 Interpretation of Construction Patterns for Biodiversity Spreadsheets 401

H1: most of the spreadsheets organization follows the pattern of columns as fields and
rows as records;
H2: most of the spreadsheets organization follows the pattern of columns as fields and
rows as records;
H3: the first fields of a spreadsheet often define its nature, e.g., catalog or event, as well
as its construction pattern.

We developed a system – SciSpread – to automatically recognize schemas based on

these hypotheses (see details at [2]). We found evidences, based on our hypothesis, that
patterns can drive the recognition of the spreadsheet nature in a context, to make its
schema explicit and to support its semantic annotation.

4 Evidences of Construction Patterns

The evidences of Construction Patterns are based in two steps. The first was based in an
automatic interpretation did by SciSpread system [2], and the second is an exploratory
analysis of a survey conducted with biologists.
Thus, the next subsections explain (i) the considerations about the scenario that we
implemented the system and observed the evidences, (ii) the survey conducted with
biologists in order to verify and refine the hypothesis, and (iii) the comparative results
about (i) and (ii).

4.1 Evidences Based on Automatic Recognition of Web Spreadsheets

Based on our preliminary assumptions presented in Sect. 3, we looked for patterns

concerning: schema layout (e.g., column labels), order and grouping of spreadsheet
fields etc. A set of hypotheses – presented in the previous section – was defined and we
developed an initial version of the automatic recognition system to validate these
hypotheses [2].
The system was tuned to recognize all spreadsheets of this initial sample, whose
nature fit in our context. We further randomly collected more 1,914 spreadsheets on the
Web, finding them through the Google search engine, based on keywords extracted from
previous spreadsheets: kingdom, phylum, order, biodiversity, species, identification key
etc. The system recognized 137 spreadsheets (7 %) of all 1,914 spreadsheets collected.
The manual analysis of these spreadsheets showed that the system correctly recognized
116 spreadsheets and incorrectly recognized (false positives) 21 spreadsheets. Even
though the latter spreadsheets have the expected construction pattern, they do not address
the focus of our study, which are spreadsheets used for data management.
Increasing our sample size to 5,633 spreadsheets, the system recognized 7 %; subse‐
quently, increasing to about 11,000 spreadsheets, the system recognized 10.4 %, which
corresponds to 1,151 spreadsheets, in which 806 were classified in the catalog and 345
in the event.
We selected a random subset of 1,203 spreadsheets to evaluate the precision /recall
of our system. The percentage of automatic recognition of the spreadsheets in the subset
was approximately the same as the larger group. Our system achieved a precision of
402 I.R. Bernardo et al.

0.84, i.e. 84 % of retrieved spreadsheets were relevant; are call of 0.76, i.e. the system
recognized 76 % of all relevant spreadsheets; and an F-measure of 0.8. The accuracy
was 93 % and the specificity 95 %, i.e. among all spreadsheets that the system does
classified as not relevant, 95 % were in fact not relevant.
The recognition rate of approximately 10.4 % of the spreadsheets must consider that
they were collected through a Web search tool. According Venetis et al. [12], these
search tools treat tabular structures like any piece of text, without considering the
implicit semantics of their organization and thus causing imprecision in the search
results. We further show an analysis of the data extracted from spreadsheets.

4.2 Exploratory Analysis Based on a Survey Conducted with Biologists

The survey was performed with 44 biologists, where 36 are from Brazil, 4 are from
France, 3 are from EUA and 1 is from Colombia, Denmark and England.
The distribution about their main activity is 11 % are professors, 23 % are researchers,
5 % work in companies, 20 % are post-doctoral students, 25 % are PhD students and
17 % are master students, where 79,5 % work with spreadsheets always or often, 14 %
work sometimes and 2 % never use spreadsheets.
We developed an online form with 10 questions to analyze: how biologists organize
their specimens’ catalogs or samples collected in the field; if the arrange of the fields in
a spreadsheet influences its nature; which fields work like identifiers according to the
spreadsheet nature; and, if it is possible to capture and map the patters biologist behavior
when producing an spreadsheet through a survey with exploratory questions (this survey
is available in goo.gl/b1iEvl).
The questions were divided in 2 blocks, in the first block we would like to know the
vocabulary which describes catalog and collection spreadsheets. Thus, the biologists
chose 8 elements from a list of 17 (author, altitude, collector, common name, date, field
number, group, latitude, life stage, locality, longitude, museum, note, species, source,
taxonomy identification, time, other) ranked by relevance. In the second block, we
presented 3 spreadsheet images and asked to biologists to classify them according to:
catalog, field sampling, both types, neither and other. If the option was “other”, an extra
text field must be filled with the description of the type more adequate to the image.
Compared to spreadsheets retrieved from the Web, the characteristics analyzed in
the survey indicated a more specialized set of spreadsheets, i.e. some spreadsheets types
retrieved from the Web and classified by the system as catalog or event (field sampling),
cannot be related to a specific context as a museum catalog or a research field sampling.
Therefore, the results of the survey refined our hypothesis about construction patterns
and in a future work we will increase the sampling with more biologists.
In the following subsections, we will confront results obtained from the automatic
recognition of Web spreadsheets by our SciSpread system introduced in Sect. 4.1 –
which we will refer in abbreviated form as SciSpread abbreviated form – and those
obtained from the survey introduced in this section – which we will refer in abbreviated
form as Survey.
 Interpretation of Construction Patterns for Biodiversity Spreadsheets 403

4.3 Comparative Results: Scispread X Survey

Pattern for Schema Location. Our SciSpread system identifies the schema of a
spreadsheet guided by the adopted vocabulary, which varies according to the type of the
spreadsheet: catalog or event. The scatter chart in Fig. 2 shows the percentage of terms
extracted from a given spreadsheet recognized in a vocabulary against the row in the
spreadsheet they were recognized. We observe that the spreadsheets schemas were
concentrated on the initial lines, since the percentage of matching terms per line
decreases exponentially as we move away from the initial lines and the most of the terms
are located in the initial lines. Therefore, there is a tendency of positioning schemas at
the top followed by their respective instances.
Percentage of Matching

spreadsheet

Schema Row

Fig. 2. Terms by schema of initial lines.

Even though we did not design a specific experiment to validate this hypothesis in the
survey, we assumed this pattern in its examples, when we asked to biologists to recognize
spreadsheet images. No biologist reported having difficulty to recognize the schema and
the answers indicate they successfully recognized it.

Predominance of Terms and Spatial Distribution. In this stage of the analysis, we

verified how much the predominant terms and their disposition in the schema can indi‐
cate the spreadsheet nature: catalog or event (see explanation of these natures in
Sect. 3.1). In order to perform a comparative analysis among proportions and positions
of the fields in spreadsheets of the catalog type, we present a radar chart in Fig. 3
(SciSpread) and Fig. 4 (Survey).
While fields and their positions were automatically recognized in SciSpread, in the
survey we requested to the interviewee to select and sort fields for a catalog spreadsheet
schema. The schema fields were grouped in one of the six exploratory questions and they
were weighted according to their position in the schema – the weight has the value one in
the leftmost field and decreases to half of its value for each field position towards the right.
In catalog spreadsheets, we observed that, charts have smaller differences and both
have a strong tendency to “what” questions, validating that spreadsheets recognized as
catalog tend to have many fields that answer the “what” question appearing in the initial
positions of the schema. The quantities of the other five questions were no significant
404 I.R. Bernardo et al.

Fig. 3. SciSpread - Proportions among fields of catalog spreadsheets.

Fig. 4. Survey - Proportions among fields of catalog spreadsheets

in the SciSpread spreadsheets and were a bit more significant in the survey spreadsheets,
with emphasis to “who” and “where” fields. It delineates a pattern for catalog spread‐
sheets, which tend to have more fields to identify and detail objects – specimens in this
case – in the beginning.
We interpret the increase in the provenance questions (who, where and when) due
to the specialization of spreadsheets in the survey, i.e. while we do not have control of
the context of the SciSpread spreadsheets and it is not possible to attest the purpose of
its catalog spreadsheets, the survey spreadsheets were clearly directed to museums,
requiring a more strict control of provenance.
Following the same approach, in Fig. 5 (SciSpread) and Fig. 6 (survey) we show
the proportions of fields in event spreadsheets. We can note that there are differences
among the proportions of the fields. As in the previous case, our interpretation is that
the differences are due to the specialization in the survey spreadsheets. While in the
SciSpread spreadsheets the provenance fields – mostly the “when” fields but also the
“who” fields – are highly predominant, appearing in the beginning, in the survey
spreadsheets the difference compared to “what” fields is less remarkable, but still exists.
 Interpretation of Construction Patterns for Biodiversity Spreadsheets 405

Fig. 5. SciSpread - Proportions among fields of event spreadsheets

Fig. 6. Survey - Proportions among fields of event spreadsheets

The predominance of provenance fields, comparing event to catalog spreadsheets, is

enough in both cases (SciSpread and survey) to distinguish their nature. In the survey,
we had the opportunity of discussing about the relationship among fields related to
“when” and “where” questions. For biologists these fields usually appear together and
there is an extra field – usually named field number – which works as an index for a
“when” and “where” specification. The relationship among these fields/questions can
be considered as a sub-pattern in an automatic analysis.
The previous charts show that it is possible to distinguish between catalog and event
spreadsheets, but we need identify which words are the most used by the biologists. In
the previous paper [2] we presented detailed observations about the distribution of fields
identified by the SciSpread, in this paper we will present values based on the Survey –
first block questions.
Thus, the next bar chart (Fig. 7) show which fields are more used by biologists in
each spreadsheet nature. The vertical axis describe the quantity of biologists that
answered the question and the horizontal axis describe the fields that they chose,
according catalog or event spreadsheets.
406 I.R. Bernardo et al.

Fig. 7. Comparative terms quantities between spreadsheets nature.

Using the result of the previous bar chart, we show in Fig. 8 the distribution of these
fields in the columns, i.e. if the biologists must create a new catalog or event spreadsheet,
how they will arrange these fields in columns? This distribution indicates that, consid‐
ering the western behavior, the tendency is to organize most import things from the left
to the right. We infer that species, taxonomy identification and museum can be identifiers
of catalog spreadsheets, and the fields date, species, locality and field number can be the
identifiers of event spreadsheets, because in both types the respective fields appear in
greater amount in the first columns.

Fig. 8. Comparative terms location between spreadsheets nature.

 Interpretation of Construction Patterns for Biodiversity Spreadsheets 407

After we consider the vocabulary and the structural organization of the fields – first
block questions, it is important analyze how the biologist classifies the spreadsheets –
second block questions. According to the spreadsheet shown in Fig. 9, 48 % of the
biologists considered this spreadsheet as catalog. The explanation to the decision is due
to the detailed information and complete taxonomy identification.
In the second image, we copied the same spreadsheet and we changed the order of
the fields, aiming at identifying if the arrangement of the fields is important to charac‐
terize the nature of a spreadsheet. The modified spreadsheet is shown in Fig. 10.
The charts in Fig. 11 show in the vertical axis the percentage of biologists that
answered the questions and in the horizontal axis the possible choices to classify the
spreadsheet images. According to these charts, 40 % of the biologists continue classi‐
fying as catalog spreadsheet, but analyzing these charts, we observed that, even though
most of the biologists do not consider the second spreadsheet as event, there is a signif‐
icant number of biologists who changed their opinion to event, and this was due to the
simple fact that we changed the fields disposition.

Fig. 9. Spreadsheet 1 - used in the Survey

Fig. 10. Spreadsheet 2 - used in the Survey.

Fig. 11. Comparative results about spreadsheets classification.

408 I.R. Bernardo et al.

Therefore, we consider that the order of the fields is important to decide the nature
of the spreadsheet, and we add the observation that there is a common understanding
among biologists that if a spreadsheet has many fields and complete information, this
spreadsheet is more related to a catalog than about events. On the other hand, if the
information is synthetic and objective, this spreadsheet is about events. This differen‐
tiation is due to the context. Biologists do not have time to detail information and classify
specimens during the collection in the field; therefore the conditions limit the answers.
The third spreadsheet, shown in Fig. 12, was classified by biologists classified as
event. According to our hypothesis, this spreadsheet must be classified as catalog,
considering only the fields. We concluded that the biologists classified as collection
despite the lack of date and location fields, due to the quantity of fields, which is limited,
characterizing a sample collection in the field.
This information is important since we were not considering in the automatic inter‐
pretation of this data type of spreadsheet the relevance of the detail level of the fields.
Another point to consider in an automatic interpretation of catalog spreadsheets is that
they never start with a “common name” field.
Through this comparative results, we verified that: the schema usually appear on the
top of spreadsheets; the fields arrangement influenced in a spreadsheet characterization;
the identifiers of catalog spreadsheets can be a taxonomy identifier or a museum number
and of event spreadsheets can be a date plus locality information or field number.

5 Representation of Construction Patterns

This section details the model, proposed in this paper, to capture and represent construc‐
tion patterns in spreadsheets, which can be interpreted and used by machines. The char‐
acteristics of this model were based on field observations reported in the previous
section. Therefore, even though we intend to conceive a generic model to represent
patterns in spreadsheets for data management in general, in the present stage our analysis
is focused in biology spreadsheets.
As detailed before, the schema recognition step involves analyzing patterns used by
users to organize their data, which we argue to be strongly influenced by the spreadsheet
nature inside a domain. Departing from our spreadsheet analysis, we produced a system‐
atic categorization of construction patterns observed in biology spreadsheets, which
supported the design of a process to recognize these patterns. Our process to recognize
construction patterns and consequently the spreadsheet nature is focused on the schema
recognition
Our representation approach considers that there is a latent conceptual model hidden
in each spreadsheet, which authors express through patterns. How authors conceive
models and transform them into spreadsheets is highly influenced by shared practices

Fig. 12. Spreadsheet 3 - used in the survey.

 Interpretation of Construction Patterns for Biodiversity Spreadsheets 409

of the context in which the author is inserted, e.g., a biologist author cataloging speci‐
mens from of a museum. Her reference to build the catalog will be the specimens them‐
selves, but also the usual strategy adopted by biologists of her community to tabulate
data from specimens. Thus, the construction patterns and the respective hidden concep‐
tual models to be represented here reflect community or domain patterns and models.
Our analysis shows that a catalog spreadsheet contains taxonomic information of a
specimen (“what” question) concentrated in the initial positions, defining their role as
identifiers. On the other hand, an event spreadsheet contains temporal and location fields
in the initial positions.
A visual analysis of the spreadsheet structure gives us directions of how the pattern
is organized, e.g., schema up /instances down; identifier on the left, as a series of
progressively specialized taxonomic references. To express these characteristics of the
pattern in a computer interpretable representation, we represent them as qualifiers [2]
identified by the prefix “q”, and they are categorized as follows:
Positional qualifier – characterizes an element in a pattern according to its absolute
position within a higher level element. There are four positional qualifiers: left (q ←),
right (q →), top (q↑) and bottom (q↓).
Order qualifier (q#) – characterizes an element in a pattern according to its relative
order regarding its neighboring elements.
Label qualifier (q@) – indicates that the label characterizes the element. In the
example, the label species identifies that this column refers to species.
Data type qualifier (q$) – characterizes the predominance of one data type in the
instances of a given property.
Range qualifier – specify if neighbor elements have generalization /specialization
relations. The qualifier (q>) indicates that the left one is more general than the right
one and (q<) the opposite.
Classified qualifier – characterizes instances of a given property that are arranged in
ascending order (q +) or descending order (q-).
Redundancy qualifier (q =) – characterizes redundancy of information in instances of
a property. Such redundancy is typical, for example, in non-normalized relations
among properties and composite properties, in which the values of a sub-property are
broader or more generic of a related sub-property – usually the value of one sub-
property embraces the value of the other.
Besides the qualifiers, we associate the relation of elements with one of the six explor‐
atory questions (who, what, where, when, why, how). This association will subsidize
the characterization of construction patterns in a more abstract level. For example,
looking at other kinds of catalog spreadsheets, outside the biology domain, we
observed they define “what” fields as identifiers and they appear in the leftmost posi‐
tion (q ←).

5.1 Formalizing the Model to Represent Patterns

In this subsection, we will present a more formal representation, to be stored in digital

format and to be read and interpreted by machines. This representation takes as a starting
point the conceptual model implicitly expressed through the pattern. Figure 13 shows
410 I.R. Bernardo et al.

Legend:
Specimen
property
domain domain domain

[what] q# [when]
identifier collected obs
q$
range
range

Specimen
Date
Identifier
Hour
domain domain domain domain
domain domain
q@#> q@#> q@#> q@# q@#^ q@#
kingdom phylum gender species date hour
q$= q$= q$= q$ q$= q$

Fig. 13. Conceptual model for catalog spreadsheets annotated with qualifiers.

the representation of the construction pattern. The model is based on the OWL Semantic
Web standard. The ovals represent classes (owl:class) and rectangles represent proper‐
ties (owl:ObjectPropertyorowl:DatatypePropert).
The root class –Specimen in the figure –is related to the spreadsheet nature; in this
case, instances of the Specimen class represent the instances of the spreadsheet, which
catalogs specimens. A class will have a set of applicable properties, represented by a
domain edge (rdfs:domain). Properties of this model are related to fields extracted from
the spreadsheet. Range edges (rdfs:range)indicate that values of a given property are
instances of the indicated class. For simplicity, the diagram omits details of the OWL
representation.
Properties in our model are annotated by qualifiers presented in the previous section.
Properties can be annotated in OWL through the owl:AnnotationProperty. In this case,
annotations are objects that specify the qualifier and the pattern they are related.
Qualifiers as annotations are depicted above and/or below the properties they qualify.
A qualifier above a property indicates that it is applied to the relationship between the
property and the class to which it is applied by the domain relation. For example, the
qualifier q ← (left positional qualifier) is represented above the identifier property, indi‐
cating that when this property is used as a field in a spreadsheet describing a Specimen,
we expect that it will appear in the left position.
A qualifier below the property means that it applies to property values – instances
in the spreadsheet. For example, the qualifier q$ = below the kingdom property indicates
that a specific type (string) and redundancy are observed in the values of this property
in the instances.
Properties are also annotated as answering one of the six exploratory questions.
These annotations are depicted in Fig. 13 inside brackets. There are additional concerns
in the OWL model that are necessary to bridge it to the implicit spreadsheet schema,
which are also represented as annotations: the order of properties and their relation with
labels. This OWL representation allows us to digitally materialize building patterns of
spreadsheets, to be shared by users and applications.
 Interpretation of Construction Patterns for Biodiversity Spreadsheets 411

6 Related Work

As discussed in Sect. 5, a fundamental characteristic of spreadsheets used for data

management is the separation between schema and instances. The schema is presented
above (q↑) or left (q ←) and instances is below (q↓) or right (q →).
This observation appears in all the papers of related work, whose purpose is to
recognize the implicit schema of spreadsheets. Syed et al. [3] point out that this challenge
leads to a more general problem of extracting implicit schemas of data sources –
including databases, spreadsheets etc. One approach to make the semantics of spread‐
sheets interoperable, promoting the integration of data, is the manual association of
spreadsheet fields to concepts in ontologies represented by open standards of the
Semantic Web.
Han et al. [6] adopt the simplest approach to devise a schema and its respective
instances, called entity-per-row [4]. In this approach, besides the schema, each row of
the table should describe a different entity and each column an attribute for that entity,
for example, each column corresponds to an attribute – e.g., Date, Genus, Species
etc. – and each row to an event – a collection of a specimen. Han et al. [6] and several
related work assume the entity-per-row organization to support the process of manually
mapping attributes, to make them semantically interoperable. Initially, the user must
indicate a cell whose column contains a field which plays the role of identifier–equivalent
to the primary key of a database.
Langegger and Wob [15] propose a similar, but more flexible, solution to map
spreadsheets in an entity-per-row organization. They are able to treat hierarchies among
fields, when a field is divided into sub-fields, for example, fields Date, Latitude and
Longitude refer to when and where the species was collected. It is usual that authors
create a label spanning the entire range above these columns – e.g., labelled as “Field
Number” – to indicate that all these fields are subdivisions of the larger field. This hier‐
archical perspective can be expressed in our model, since a property can be typed
(rdfs:range) by a class, which in turn has properties related to it.
RDF has been widely adopted by related work as an output format to integrate data
from multiple spreadsheets, since it is an open standard that supports syntactic and
semantic interoperability. Langegger and WOB [15] propose to access these data
through SPARQL [16] – a query language for RDF. Oconnor et al. [4] propose a similar
solution using OWL.
Abraham and Erwig [10] observed spreadsheets are widely reused, but due to their
flexibility and level of abstraction, the reuse of a spreadsheet by people outside its
domain increases errors of interpretation and therefore inconsistency. Thus they propose
a spreadsheet life cycle defined in two phases: development and use, in order to separate
the schema of its respective instances. The schema is developed in the first cycle, to be
used in the second cycle. Instances are inserted and manipulated in the second cycle
guided by the schema, which cannot be changed in this cycle.
Another approach to address this problem is automating the semantic mapping
using Linked Data. Syed et al. [3] argue that a manual process to map spreadsheets is
not feasible, so they propose to automate the semantic mapping by linking existing
data in the spreadsheets to concepts available in knowledge bases, such as DBpedia
412 I.R. Bernardo et al.

(http://dbpedia.org) and Yago (http://www.mpi-inf.mpg.de/yago-naga/yago/). Yago is

a large knowledge base, whose data are extracted, among others, from Wikipedia and
WordNet (http://wordnet.princeton.edu). The latter is a digital lexicon of the English
language, which semantically relates words.
Among the advantages of the last approach, there is the fact that such bases are
constantly maintained and updated by people from various parts of the world. On the
other hand, the search for labels without considering their contexts can generate ambig‐
uous connections, producing inconsistencies. Thus, there are studies that stress the
importance of delimiting a scope before attempting to find links.
Venetis et al. [12] exploit the existing semantics in the tables to drive the consistent
manipulation operations applicable to them. The proposal describes a system that
analyzes pairs of terms heading columns and their relationship, in order to improve the
semantic interpretation of them. Authors state that a main problem in the interpretation
of tabular data is the analysis of terms independently. This paper tries to identify the
scope by recognizing a construction pattern, which is related to a spreadsheet nature
inside a context.
Jannach et al. [11] state that the compact and precise way to present the data are
primarily directed to human reading and not for machine interpretation and manipula‐
tion. They propose a system to extract information from web tables, associating them
to ontologies. They organize the ontologies in three groups: 1. core: concepts related to
the model disassociated from a specific domain; 2. core + domain: domain concepts of
a schema related to the information to be retrieved; 3. instance of ontology: domain
concepts of instances. These ontologies aim at gradually linking the information to a
semantic representation and directed by the user’s goal.
Among these solutions, we note that some of them address individual pieces of
information inside spreadsheets – devoid of context – and others consider the importance
of identifying and characterizing the context. Even though all approaches rely on
construction patterns of spreadsheets, none of them proposes a model to represent,
exchange, reuse and refine these patterns, which is one of the main contributions of this
work.

7 Conclusions and Future Work

This paper presented our thesis that it is possible, from a spreadsheet structure, to recog‐
nize, map and represent how users establish construction patterns, which are reflected
in the schema and data organization. One of our main contributions here is an exploratory
observation of these patterns through a survey with biologists, as well as the refinement
of hypothesis for construction patterns.
Our process also involves the comparative analysis between the SciSpread system
and survey results. None of the related work departs from the characterization of the
underlying conceptual models and their association with construction patterns, to cate‐
gorize spreadsheets according to the nature of information they represent, and to recog‐
nize them.
 Interpretation of Construction Patterns for Biodiversity Spreadsheets 413

Our studies presented here have focused in the area of biodiversity. We intend to
investigate its generalization to other domains of knowledge, extending this strategy to
a semiotic representation.

Acknowledgements. Work partially financed by FAPESP (2012/16159-6), the Microsoft

Research FAPESP Virtual Institute (NavScales project), the Center for Computational
Engineering and Sciences - Fapesp/Cepid 2013/08293-7, CNPq (grant 143483/2011-0, MuZOO
Project and PRONEX-FAPESP), INCT in Web Science (CNPq 557.128/2009-9), CAPES, as well
as individual grants from CNPq.

References

1. Tolk, A.: What comes after the Semantic Web - PADS Implications for the Dynamic Web,
pp. 55–62 (2006)
2. Bernardo, I.R., Santanchè, A., Baranauskas, M.C.C.: Automatic interpretation spreadsheets
based on construction patterns recognition. In: International Conference on Enterprise
Information Systems (ICEIS), pp. 1–12 (2014)
3. Syed, Z., Finin, T., Mulwad, V., Joshi, A.: Exploiting a Web of Semantic Data for Interpreting
Tables, pp. 26–27 (2010)
4. O’Connor, M.J., Halaschek-Wiener, C., Musen, M.A.: Mapping master: a flexible approach
for mapping spreadsheets to OWL. In: Patel-Schneider, P.F., Pan, Y., Hitzler, P., Mika, P.,
Zhang, L., Pan, J.Z., Horrocks, I., Glimm, B. (eds.) ISWC 2010, Part II. LNCS, vol. 6497,
pp. 194–208. Springer, Heidelberg (2010)
5. Zhao, C., Zhao, L., Wang, H.: A spreadsheet system based on data semantic object. In: 2010
2nd IEEE International Conference on Information Management and Engineering, pp. 407–
411 (2010)
6. Han, L., Finin, T.W., Parr, C.S., Sachs, J., Joshi, A.: RDF123: from spreadsheets to RDF. In:
Sheth, A.P., Staab, S., Dean, M., Paolucci, M., Maynard, D., Finin, T., Thirunarayan, K. (eds.)
ISWC 2008. LNCS, vol. 5318, pp. 451–466. Springer, Heidelberg (2008)
7. Yang, S., Bhowmick, S.S., Madria, S.: Bio2X: a rule-based approach for semi-automatic
transformation of semi-structured biological data to XML. Data Knowl. Eng. 52(2), 249–271
(2005)
8. Ponder, W.F., Carter, G.A., Flemons, P., Chapman, R.R.: Evaluation of Museum Collection
Data for Use in Biodiversity Assessment. 15(3), 648–657 (2010)
9. Doush, I.A., Pontelli, E.: Detecting and recognizing tables in spreadsheets. In: Proceedings
8th IAPR International Workshop Document Analysis System - DAS 2010, pp. 471–478
(2010)
10. Abraham, R., Erwig, M.: Inferring templates from spreadsheets. In: Proceeding 28th
International Conference on Software Engineering - ICSE 2006, vol. 15, p. 182 (2006)
11. Jannach, D., Shchekotykhin, K., Friedrich, G.: Automated ontology instantiation from tabular
web sources—The AllRight system☆, Web Semant. Sci. Serv. Agents World Wide Web
7(3), 136–153 (2009)
12. Venetis, P., Halevy, A., Pas, M., Shen, W.: Recovering semantics of tables on the web. Proc.
VLDB Endow. 4, 528–538 (2011)
13. Mulwad, V., Finin, T., Syed, Z., Joshi, A.: Using linked data to interpret tables. In:
Proceedings of the International Workshop on Consuming Linked Data, pp. 1–12 (2010)
414 I.R. Bernardo et al.

14. Jang, W., Seiie, Ko, Eun-Jung and Woo: Unified user-centric context: who, where, when,
what, how and why. In: Proceedings of the International Workshop on Personalized Context
Modeling and Management for UbiComp Applications, pp. 26–34 (2005)
15. Langegger, A., Wöß, W.: XLWrap – querying and integrating arbitrary spreadsheets with
SPARQL. In: Bernstein, A., Karger, D.R., Heath, T., Feigenbaum, L., Maynard, D., Motta,
E., Thirunarayan, K. (eds.) ISWC 2009. LNCS, vol. 5823, pp. 359–374. Springer, Heidelberg
(2009)
16. Pérez, J., Arenas, M., Gutierrez, C.: Semantics and complexity of SPARQL. ACM Trans.
Database Syst. 34(3), 1–45 (2009)
Enterprise Architecture
 Integrating Business Information Streams in a Core
Banking Architecture: A Practical Experience

Beatriz San Miguel ✉ , Jose M. del Alamo, and Juan C. Yelmo

( )

Center for Open Middleware (COM), Universidad Politécnica de Madrid,

Campus de Montegancedo, E-28223 Pozuelo de Alarcón, Madrid, Spain
{beatriz.sanmiguel,jose.delalamo,
juancarlos.yelmo}@centeropenmiddleware.com

Abstract. Traditional commercial and retail banks are under great pressure from
new competitors. They must rise to the challenges of understanding their customer
actions and behaviors, and be ready to meet their expectations even before they
explicitly express them. But the ability to know customers’ demands in nearly
real-time requires the evolution of existing architectures to support the detection,
notification, and processing of business events to manage business information
streams. This paper describes a practical experience in evolving a core banking
enterprise architecture by leveraging business event exploitation, and includes
the definition of business events; the design of a reference architecture and its
integration points with the legacy architecture, as well as the description of an
initial governance approach. Furthermore, as the core banking architecture is a
critical infrastructure we have evaluated the performance of the evolved archi‐
tecture so as to understand whether or not it can meet the banks’ quality levels.

Keywords: Enterprise Architecture (EA) · Event-Driven Architecture (EDA) ·

Event · Core banking · Business information · Middleware

1 Introduction

Business information has become a critical asset for companies and it of even more value
when obtained and exploited in nearly real time. The correct distribution of this infor‐
mation to all interested parties and its accurate, subsequent use and processing by
different applications allow companies and organizations to react quickly to changes in
their environment and obtain multiple profits.
In order to fulfill the aforementioned requirements, Event Driven Architecture
(EDA) was devised. EDA allows systems and applications to deliver and respond to real
time information (events), helping to support business needs from an IT management
point of view [1]. Thus, it has associated both technological benefits and business
advantages. As regards the former, EDA provides a loose coupling between its compo‐
nents, which reduces dependencies and allows modifications without giving rise to side
effects. A many-to-many communication pattern is also applied, facilitating the reusa‐
bility of information and the freedom to act independently with the received information.
All of the above creates an adaptive and flexible architecture that results in business

© Springer International Publishing Switzerland 2015

J. Cordeiro et al. (Eds.): ICEIS 2014, LNBIP 227, pp. 417–433, 2015.
DOI: 10.1007/978-3-319-22348-3_23
418 B. San Miguel et al.

advantages. EDA enables faster, more agile and more responsive business processes,
enhancing the informed decision-making model and the automation, processing and
motoring of operational activities, among other business advantages.
Companies can follow different strategies to introduce the EDA approach and
thus exploit the streams of business information. However, these strategies obvi‐
ously have to coexist and comply with the existing software architecture solutions
and legacy systems. In this paper we describe our experience in evolving a core
banking Enterprise Architecture (EA), leveraging real time business information
processing and exploitation.
The work presented in this paper has been carried out in the context of the Center
for Open Middleware (COM), a joint technology center created in 2011 by Santander
Bank (together with its technological and operational divisions ISBAN and
PRODUBAN) and Universidad Politécnica de Madrid. COM is the incubator of an open
software ecosystem aiming at developing middleware solutions and experimenting with
new software architecture approaches.
There are different technologies hosted under the COM umbrella and one of the key
ones is an EA called BankSphere (BKS). Created by the Santander Group, BKS is a set
of integrated design tools, and a deployment and runtime environment that speeds up
the development of new bank software such as applications for customers, call center
staff or bank branch workers. It provides great flexibility and control, thus saving time,
resources and money. Since its introduction, BKS has constantly evolved to fulfil
Santander requirements, and it is now required to enhance the generation and exploita‐
tion of real time business information.
In this paper we present our findings on how to introduce an event-oriented approach,
architectures, tools and technologies and their potential application and integration into
a core banking EA. Specifically, we describe how to incorporate and use real time busi‐
ness information in the BKS context, identifying the key elements necessary and inte‐
grating them into BKS, while minimizing interference with the existing architecture and
procedures.
The remainder of this paper is structured as follows. Section 2 covers related work
and puts our work in perspective. Section 3 gives an overview of the background of
EDA’s main concepts. In Sect. 4 we present a high level description of the BKS char‐
acteristics and architecture. Then, we address and detail the proposed solutions to intro‐
duce EDA in BKS in Sect. 5. It includes the definition of business events; the design of
a reference architecture, which identifies the integration points with the specific EA, and
the description of the initial governance approach to manage the new elements.
Section 6 describe an operational validation though a case study and a non-functional
validation focussing on performance. Specifically, we describe the evaluation of two
main metrics: throughput (messages per second supported) and latency (delay experi‐
enced by the messages). Finally, Sect. 7 concludes the paper and introduces areas of
future research.
 Integrating Business Information Streams in a Core Banking Architecture 419

2 Related Work

Diverse studies have tackled the introduction of business events in existing architectures
of different domains. Most of them describe general approaches for EDA and SOA
integration such as [1, 2] or [3], while others address only the specific issues of EDA
integration such as modelling, simulation, methodologies, performance, etc. For
example, [4] proposes an EDA modelling notation and its associated simulation
language; [5] describes a unified event ontology and a methodology for event-driven
business processes; and [6] explains a business-oriented development methodology for
event processing.
The papers reviewed provide the theoretical basis to evolve EAs towards the EDA
paradigm. Most of them include a validation exercise through a case study in the field
of application. However, these case studies are usually academic or simplified examples,
not practical experiences for real-world EAs, since most companies, and banks in partic‐
ular, do not usually publish the evolution experiences of their core EA.
Our contributions focus on this last point, a real-world EA evolution and its associ‐
ated solutions, which we think are of the utmost interest for engineers and practitioners.

3 Event Driven Architecture

EDA is a software architecture style based on multiple entities communicating asyn‐

chronously via announcements or notifications, known as events [7]. Instead of the
traditional synchronous, request-response interaction model, where a requestor asks for
services or messages and waits for an answer from a replier; in EDA, events are trans‐
mitted in a fire-and-forget mode. In other words, events are communicated without a
previous request and without being concerned about what happens with them afterwards.
Basically, an event is a change in a state within a particular system or domain that
merits attention from other systems [2]. The term has been given other meanings,
depending on the context. It can refer to the actual occurrences (the things that have
happened), which are also known as instances of a particular type of event. On the other
hand, we can use ‘event’ or ‘notification’ to specify the particular communication of an
event instance. Generally, the word ‘event’ is used in both cases without distinction. We
will use ‘event instance’ or ‘event notification’ where its distinction is relevant.
We can think about different types of event taking place in a company, such as events
related to low-level technical information, software activity, user actions or business
data. Furthermore, we may also consider events happening outside the company (e.g.
stock markets, social networks or any other data sources). By way of example, low-level
technical events can be information from sensors, ATM status, network data or activity
in many other devices. Software events can indicate calls to methods, execution of serv‐
ices or exceptions in the execution of a program or a process. We may understand user
events as actions or information generated by both customers and workers of a company.
Finally, this paper focuses on business events.
420 B. San Miguel et al.

Fig. 1. Generic EDA layers.

These events are generated by the core company activities and represent relevant
information that has an impact on its economic development and management. For
instance, in a financial institution, business events can derive from the registration of
new customers, canceling of services, money withdrawals, or the contracting of products
such as credits, mortgages, etc.
A generic EDA is made up of three core layers: producers, channel and consumers
(Fig. 1). The process begins at the producer layer, detecting, creating and sending events
through a channel, and ends when consumers receive these event notifications and carry
out a specific task (automatically or with human intervention).
Producers can contain a software subcomponent called preprocessor to add intelli‐
gence to the event publication. It can carry out different tasks such as, filter, prioritize
or homogenize the produced events. Thus, only the most relevant events will be sent or
all event notifications will have the same format.
The channel is responsible for transporting event notifications between producers
and all associated consumers. It usually takes the form of a Message Oriented Middle‐
ware (MOM) [8], which is a software infrastructure that can send and receive messages
between distributed systems, regardless of platforms, technologies and resources used.
A MOM can use different messaging models such as point-to-point or publish/
subscribe. In the first model, only one consumer receives a particular event notification,
while in the second, more than one consumer may express their interest in a set or subset
of event types, in order to be notified when a producer generates their registered interest
[9]. Technically, it is usually achieved thanks to an intermediary entity known as broker
that receives all event notifications from producers and routes them to the subscribed
consumers, using queues that store event notifications if necessary.
There are three characteristic examples of standards for MOM that implement the
publish/subscribe model: the Data Distribution Service for Real-Time Systems (DDS)
[10]; the Java Message Service (JMS) [11]; and the Advanced Message Queuing
Protocol (AMQP) [12].
DDS was created by the Object Management Group (OMG) and is the first middle‐
ware standard that defines publish/subscribe messaging model for real time and
embedded systems. It supports a lot of Quality of Service (QoS) policies to control many
aspects of data delivery and quality and provides a robust architecture. However, DDS
has few open source implementations. For its part, JMS is a widely used Java MOM
API from Sun Microsystems but it is not a messaging system in itself. JMS only defines
the interfaces and classes needed to communicate two Java applications and does not
 Integrating Business Information Streams in a Core Banking Architecture 421

include the specification of brokers or queuing systems. Finally, AMQP is an open

standard protocol for MOM, originated in the financial services industry in 2006. It
enables client applications to communicate (sent or received messages) with messaging
brokers, using messaging queues.
On the other hand, consumers can be any entity such as software components, appli‐
cations or systems that react to the received notifications. For example, consumers can
create new event notifications, invoke a service, initialize a business process, increase
a value or notify humans to carry out manual tasks. There is a special kind of consumer
that is known as an event processing engine, which encompasses the set of computational
procedures to carry out operations with events such as reading, creation, transformation,
deletion or correlation [13]. Because of its importance, it is frequently considered as an
independent layer in EDA.
There are three styles of event processing which may be used together [14]: simple,
stream and complex. The former is the most basic process: an event is received and it
produces an action. On the other hand, Event Stream Processing (ESP) [15] continually
receives all kinds of events (ordinary and notable) and, through established rules or
queries on the flow of data, it decides whether or not to forward events (or information
on them) to other consumers. Finally, Complex Event Processing (CEP) [16] relates
different event types from various sources to produce new events or extract relevant
information.

4 BKS Banking Service Platform

BKS is a set of development tools created by the Santander Group that allows program‐
mers to create new banking applications quickly. It includes a design framework inte‐
grated with the Eclipse IDE, and a deployment and runtime environment based on Java
Enterprise solutions and web technologies.
BKS has been designed to allow the reuse of software components and simple, fast
programming. The former is achieved through its Service Oriented Architecture (SOA)-
like approach, where pieces of software are developed and exposed to be reused by other
components. The latter is carried out by using a visual programming environment which
allows programmers to design applications through usable graphical user interfaces
(GUI). It hides the code details and lets programmers to use graphic symbols that repre‐
sent software components.
Simply put, BKS programmers can create presentation and business flows by reusing
previously implemented software components. The business flows are exposed by a
facade and can be used to create banking applications. An application is usually consti‐
tuted by a main presentation flow that calls different business flows, which in turn call
backend operations or services.
A BKS application is typically turned into a Java Enterprise Edition application,
exposing the application through a web module (WAR) and implementing the business
logic through various Enterprise Java Beans (EJB). It is then executed in a runtime
environment provided by BKS.
422 B. San Miguel et al.

The execution of BKS applications at runtime is as follows (Fig. 2). First, a request
for an application is detected and redirected to the operation container. It invokes
different initial operations and the states defined by the presentation flow. These states
can call business flows through a common facade. At this point, the execution entails
invoking backend operations and external components defined by the business logic.
Finally, the request ends when all presentation states have been executed and the control
is returned to the user.

Fig. 2. Basis of BKS applications at runtime.

5 Proposed Solutions

In the previous sections we have reviewed the foundations of EDA and the main features
of BKS. Converging on a solution that brings the best of these architectures requires
extending current BKS capabilities and identifying the key integration points that inter‐
fere minimally in the existing architecture and the associated procedures.
We include the proposed solutions to evolve an EA towards the EDA paradigm in
the following subsections. The specific results are: the definition of the new banking
business events; the design of a reference architecture to integrate EDA that allows
business event generation and exploitation and identifies the specific integration points
with BKS; and finally, a description of the initial governance approach to manage the
new EDA elements.

5.1 Business Event Definition

The event definition entails deciding what semantic and data each event instance must
contain, and which data-exchange format is assigned to event notifications. Given that
 Integrating Business Information Streams in a Core Banking Architecture 423

there is no standard or a generally adopted event format, and there is a huge variety of
business event types with different meanings and aims, the event definition is one of the
most problematic issues in EDA integration.
We have carried out a study with the possible alternatives that are used or can be
used to communicate and later, process business events, concluding with the following:
– There is a wide range of event definitions in different formats that addresses specific
issues in a company and changes depending on the purpose of the event, the domain,
or the business layer [17].
– Event notifications can be implemented by any data-exchange format such as XML,
JSON, Google Protocol Buffers, CSV, ASN.1 or Hessian [18] and [19].
– Event processing engines can use Java Objects, expressions or JSON-based or tag-
delimited languages to represent events [13].
Moreover, we have examined several initiatives proposed in the field of web services,
such as the WS-EventDescription [20] or the WS-Notification [21]. They are not specific
event definitions and cover the description of communication protocols between web
services. Events in other domains, for example the specification Activity Streams [22]
for social web application, have also been reviewed.
Clearly, there is no agreement on how to represent event-related information and
some solutions can be used either jointly or separately. However, there is a trend towards
formats that allow the inclusion of two differentiated parts: header and body [13, 14].
The header includes metadata information: generic event information such as the name,
identifier, occurrence time or producer identification, or can describe the event type. On
the other hand, the body or payload contains the specific data on the event instance.
We have decided to follow the aforementioned structure using XML as the repre‐
sentation format. We have defined a general XML Schema Definition (XSD) that
contains the basic structure for all the event types we are managing (Fig. 3).
In our structure, the header has three basic elements common to all events:
– eventType: indicates the type of event according to the hierarchy of the Santander
Group’s business event catalogue. It has an attribute denominated as a category that
specifies the nature or domain of the event and the value is a text string that contains
two parts separated by a dot, indicating the business area and the specific type.
– createdTime: contains the timestamp of an event occurrence.
– createdBy: identifies the event producer that generates the event.
The body is limited by an element called logRecordAppData that contains a reference
to other separated XSD. There is an XSD for each event type and each of them describes
the specific data of an event type. It is important to highlight that we have divided the
header and body definitions into different XSDs to allow specialized actors to handle a
specific part.

5.2 Reference Architecture

An EDA process starts with the detection and creation of events. Although BKS does
not include any EDA layers, it uses relevant business information that can be mapped
424 B. San Miguel et al.

to banking business events. As a result, event producers have been detected as the only
integration point between EDA and BKS architectures (Fig. 4).

Fig. 3. XML event definition.

Fig. 4. EDA integration with BKS.

We have detected two main ways of incorporating event producers in BKS appli‐
cations: explicit and implicit.
– Explicit. BKS programmers must include a call to an event producer component to
generate business events, configuring the exact values for the business event instance.
In other words, they have to decide where to include the creation of the event within
the business logic and moreover, obtain the context data that corresponds to event
instances.
This solution has several drawbacks. First, programmers have to acquire new
responsibilities and understand new concepts related to business areas that differ from
their daily technical work. Secondly, this incorporation into existing banking appli‐
cations, which are currently in production, implies their modification and can entail
risks in stable applications.
 Integrating Business Information Streams in a Core Banking Architecture 425

– Implicit. Here, the incorporation of event producers is almost transparent for

programmers. They are strictly limited to defining the business logic and the business
event generation is associated with calls to business flows.
The main disadvantage of this solution is the low quality of the generated events.
They correspond to calls to functional business methods but they do not necessarily
tie in with business event definitions. To solve this last point, a preprocessor can be
included. It can create real business events based on execution traces obtained from
calls to the facade. However, it requires an in-depth analysis of the context and each
functional component to be related to business events.
Both alternatives can coexist in BKS. We have specifically proposed to use explicit
business event generation for new BKS applications and implicit generation for existing
applications.
The remaining of the EDA layers (channel and consumers) will be new elements in
BKS. At present, BKS already incorporates a stable commercial messaging system for
logging the application. It allows execution traces to be stored in a database to be batch
processed. Therefore, a MOM that allows the publish/subscribe model to distribute
business events in real time has to be incorporated.

5.3 Governance Approach

The incorporation of EDA elements in BKS entails the creation of new operational and
organizational processes that allow the Santander Bank to govern business events.
Governance is a wide discipline that can be applied on multiple perspectives of a
company such as that related to EA, IT, data, business or SOA [23]. Basically, it seeks
to define a global structure for establishing and ensuring how the company resources
sustain and extend the organization strategies. To begin with, we have identified the
organizational processes involved in the creation, use and reuse of business events in
the BKS context. This new process has been called event lifecycle and has been defined
according to the existing procedures in the bank.
The event lifecycle (Fig. 5) describes the different activities or tasks that must be
carried out in design time to define, incorporate and use business events generated by
applications. In previous sections, we noticed that BKS programmers know the func‐
tional specification and logical model of their applications. However, they ignore the
business value of their components. Consequently, other stakeholders must participate.
We have identified four main actors in the event lifecycle: the project leader of an
event-oriented application that wants to use a specific event, the project leader respon‐
sible for the BKS application that generates the event and the corresponding program‐
mers of each project. Members of the quality department participate in approving the
different stages.
The lifecycle starts with an identification stage: the event-oriented project leader
detects the need to consume a specific business event to take advantage of it and exam‐
ines whether the specific banking business event exists. If it is already incorporated into
any BKS application, a subscription to this event is made. Otherwise the new event is
defined.
426 B. San Miguel et al.

Fig. 5. Event lifecycle.

The definition process begins with a formal request for the incorporation of a new
business event (application stage). The request goes through a validation stage and if
approved, the event is defined and incorporated into a BKS application. The quality
department validates this last step again. The definition process ends with the event
subscription and use.

6 Validation

In order to validate the previous solutions, we have developed a proof-of-concept test

bed in an open source environment. It incorporates several event producers, a MOM, a
CEP, and different event consumers that give shape to a banking case study. The case
study supports the operational validation of our contributions as well as the evaluation
of some non-functional aspects, such as performance.

6.1 Case Study

The case study consists of a wire transfer scenario whose aim is to demonstrate the
feasibility of the proposed solutions and the value added to the business by EDA. The
scenario takes into account Santander customers who are sending and receiving money
from the same bank or others. It must incorporate the technologies and mechanisms that
allow the detection, distribution and use of the associated business events. Moreover, it
must show any of the multiple possibilities for exploiting these events in real time.
We have identified two main business event types: sent wire transfer and received
wire transfer. The former represents the events of orders that Santander customers carry
out to transfer a certain amount of money to other bank institutions. Conversely, received
wire transfers are orders from customers of other banks to Santander customers. Each
 Integrating Business Information Streams in a Core Banking Architecture 427

event type contains the specified header and the following information in the body:
session identification, IP address, source account, target account and amount of the
transfer.
Our scenario includes the following logical entities (Fig. 6):
– Two event producers. Each of them generates a different business event for wire
transfers and publishes it through a MOM.
– One MOM distributes the received events to all the associated consumers.
– A CEP acts as a consumer and receives all the previous events. It extracts relevant
information and displays it in a visualization dashboard. Moreover, if applicable, it
generates a new event type that indicates that an individual (not a corporate entity)
has received a wire transfer above a threshold. This new event type is called user
alert.
– An application displaying a wire transfer dashboard that shows relevant information
on the business events of sent and received wire transfers.
– Two consumers that react to the user alert event. One of them is a simulation of a
Customer Relationship Management (CRM) that displays records of the wire trans‐
fers received by Santander users. Moreover, it can manage, assign and create alerts
to call-center software with the aim of carrying out commercial actions. The other
consumer is a user notification system that sends mails and/or Short Message Service
(SMS) to Santander users that have activated the real time notification service to be
informed about their transactions.

Fig. 6. Architecture of the case study.

We have developed the previous components using open source tools and applica‐
tions, and the Java programming language. Event producers have been entirely imple‐
mented as Java preprocessors that use information from different BKS applications.
They generate event notifications based on the proposed XML event definition for sent
and received wire transfers and publish them through a channel.
428 B. San Miguel et al.

The channel is implemented by an open source MOM called RabbitMQ [24] that
supports the AMQP in its 0.9.1 version. AMQP covers different messaging models such
as publish/subscribe. Its main advantage is to be interoperable, thus allowing consumers
and producers to use any programming language or data format.
As regards the CEP consumer, we have selected an open source solution available
for Java such as Esper, and for.NET such as NEsper [25]. It allows large volumes of
events to be processed by applying ESP and CEP. Basically, Esper allows applications
to store queries and run the event streams through to obtain information or new event
streams. Queries are written using Event Processing Language (EPL) [13] and [26] that
is similar to the Structure Query Language (SQL) of databases. The differences are: the
queries are carried out through an event stream, the data available for querying within
the stream is defined by views and the basic units of data are events and their specific
information.
We have obtained information on the received events with EPL queries such as the
total number of sent or received transfers, the accumulated amount or the amounts per
second for each event type. We also have drawn up a ranking of the top five banks for
the most sent accumulated amounts. We extract extra information on the received events
by carrying out an enrichment process. Thus, we have located the source of the transfers.
In particular, the associated Santander bank branches that are the source or destination
of transfers.
All the previous information has been displayed in a web application called Wire
Transfer Dashboard. We have implemented it using the Google Charts that allow a lot
of chart types such as maps, tables, or line or column diagrams to be incorporated. We
used WebSocket technology to communicate Esper with the web application.
Finally, we have used Esper to obtain a new event stream with all the received wire
transfers that are more than €3,000 and whose target account belongs to an individual,
not a corporate entity. We have developed two consumers that react to it: CRM and the
user notification system, implemented by Java web applications.

6.2 Performance Evaluation

We have carried out a series of performance measures to evaluate whether the selected
open source tools fulfil the basic requirements for our case study. In particular, we have
begun by verifying two main metrics, namely throughput and latency, using different
configurations in two environments.
The throughput, also known as the message rate, measures how many messages per
second can be supported. It is similar to the bandwidth (amount of raw data in MegaBytes
that may be transmitted per second), but refers to the number of discrete messages (in
our case events) that can be produced or consumed. On the other hand, latency indicates
the delay experienced by messages from producers to consumers. Generally, latency is
explained from the channel point of view, but producers and consumers can also inter‐
fere, depending on how the environment is configured. In our performance evaluation,
we have considered production and consumption rates of throughput and latency.
The measures were repeated using different persistency configurations. Specifically,
we dealt with persistent and non-persistent messages. The former guarantees that a
 Integrating Business Information Streams in a Core Banking Architecture 429

message will not be lost, and so it could survive a crash in the EDA layers. To ensure
it, producers have to flag the messages as persistent before publishing, and the broker
components have to be configured as durable [27]. Durability is a similar concept to
persistence but applied to RabbitMQ components. By default, these components are
transient or non-durable and therefore, they do not survive reboots and could be lost on
failovers.
We have compared the performance of the two types of message in different
scenarios. First, we considered a real time scenario, in which one producer publishes
events and one consumer reacts to them as soon as they arrive. We have called it scenario
1:1. Moreover, we have also defined a scenario 1:n in which more than one consumer
has participated. In this case, the variable n indicates the number of consumers in the
different tests.
We have worked with two different environments to compare the results. First, we
used two virtual machines, one hosts the messaging broker and another hosts the
consumers and producers. In this environment each machine has an Intel Xeon E5520
@2.27 GHz × 4 core processor, running 2 GB RAM, 30 GB disk capacity and Ubuntu
12.04 LTS, 64 bit server. The virtual machines are hosted in a shared network of the
university. The second environment includes two physical machines with greater
features, an Intel Core i7-3720Q @2.60 GHz × 8 core processor, 16 GB RAM and 80 GB
disk capacity, running in an isolated network.
In the following subsections we will describe the different tests carried out in each
scenario and the results obtained. It is important to point out that each test was run three
times and that the measures obtained were averaged.

Scenario 1:1. It is the more general case in which there is only one producer and one
consumer, publishing and receiving messages, respectively.
We have carried out different tests to check how the production and consumption
throughput varies with the message size in the described environments. First, we
performed an in-depth test in which a producer sent bursts of ten thousand messages
through the broker to a producer. We repeated the test increasing the message payload

Fig. 7. Throughput for producers in scenario 1:1.

430 B. San Miguel et al.

from 181 bytes to 12 kilobytes with a 120 byte interval. The tests were repeated for both
the persistent and non-persistent configurations, and in both the virtual and physical
environments.
Figure 7 summarizes the results for the throughput of producers, which are very
similar but a little lower for consumers. The observed throughput decreases when the
message size increases. There is also a notable difference between using virtual or phys‐
ical machines.
Based on the analysis of the business events managed by Santander, we estimated
that the payload of our messages (event notifications) would be between 2 and 5 kilo‐
bytes. For these figures, we observe that the throughput is greater than 4,000 messages
per second in both cases (virtual and physical environments), doubling the amount in
the physical machines.

Fig. 8. Latency in the different environments, scenario 1:1.

In addition to the previous tests, we continued increasing the message size until the
saturation throughput was reached. In the virtual-machine environment, we observed
steady values of around 100 messages per second, for message sizes of 100 kilobytes
and over. For the case of physical machines, we identified the message size of 600
kilobytes as the saturation point. Here, the throughput falls to 200 messages per second
and it is almost stable for bigger messages. The previous values are production
throughput but, as in the previous test, consumption throughput figures are very similar
but with a smaller difference.
As regards the latency, it practically maintains a value of slightly above 100 ms in
the range from 181 bytes to 12 kilobytes. Specifically, we obtained an average delay of
119–125 ms with physical machines and 148–158 ms with virtual machines.
Finally, as expected, we perceived that the latency increases with the size of the
messages. In the virtual machine environment the latency fluctuates sharply (Fig. 8, left)
while it is more continuous in the physical machine environment (Fig. 8, right). We
think that this difference in the behavior is brought about by the network configuration,
since it is a shared network in the virtual environment but it is an isolated network in
the physical environment.

Scenario 1:n. This scenario includes one consumer publishing messages, which are
delivered to n consumers through the channel. In this scenario we used just the physical
 Integrating Business Information Streams in a Core Banking Architecture 431

machine environment, since the goal was to understand how many consumers can be
hosted in the same machine.
As in the previous scenario, we varied the size of the messages and calculated the
throughput and latency in each consumer. However, we have reduced the number of
sizes and increased the number of sent messages for each size. Specifically, the producer
published 40,000 messages for each of the four message sizes tested, namely 1 byte, 1
kilobyte, 10 kilobytes, and 1 megabyte.
We started by executing a test with one producer and one consumer, and we increased
the number of consumers iteratively at the end of each test. We detected the saturation
of the scenario with 6 consumers, when the machines crashed.
The throughput drops when a consumer is added to the scenario. It is more
pronounced in the case of using persistent messages (Fig. 9, right) than non-persistent
messages (Fig. 9, right).

Fig. 9. Throughput for consumers, scenario 1:5.

The results obtained are within the target range of message payload (2-5 kilobytes)
and therefore, they suggest that RabbitMQ can be applied in BKS. Moreover, although
using physical machines would cover larger amounts of events, virtual machines do also
meet our requirements, providing greater flexibility and adaptation to variable condi‐
tions.

7 Conclusions

We have analyzed how to introduce business event streams in a legacy, core banking
architecture, and have described our practical experience on this journey. Supporting
business event processing at an architectural level, would allow programmers to create
software bank applications quickly, efficiently and proving high performance. In turn,
this architectural evolution would support traditional commercial and retail banks in
dealing better with new entrants in the banking domain by leveraging business event
exploitation.
We have analyzed the technical challenges that this integration involves, and have
proposed a set of solutions to meet them, including a business event definition based on
an XML structure with our own semantic; a reference architecture to integrate business
events generation, notification and processing that identifies the specific integration
432 B. San Miguel et al.

points with the core banking EA; the definition of an event lifecycle that allows the
incorporation and use of business events without interfering with the existing legacy
infrastructure; and the related governance processes. These solutions have been success‐
fully validated in a proof-of-concept test bed that uses open source tools. Furthermore,
we have carried a performance evaluation of the selected tools to understand whether
they would meet the bank quality requirements.
Since our first results have demonstrated the workability of our approach, the future
points towards a further analysis the governance approach. A core banking architecture
requires solutions that allow the cataloguing and managing of events to optimize their
production, reuse and consumption. The monitoring of events and their lifecycle will
also ensure the consistency of the solution.

Acknowledgements. The work presented in this paper has been carried out in the context of the
Center for Open Middleware (COM), a joint technology center created by Universidad Politécnica
de Madrid, Banco Santander and its technological divisions ISBAN and PRODUBAN.

References

1. Malekzadeh, B.: Event-Driven Architecture and SOA in collaboration - A study of how Event-
Driven Architecture (EDA) interacts and functions within Service-Oriented Architecture
(SOA). Master of Thesis in IT-Management, Department of Applied Information
Technology, University of GothenBurg (2010)
2. Taylor, H., Yochem, A., Phillips, L., Martinez, F.: Event-Driven Architecture: How SOA
Enables the Real Time Enterprise. Pearson Education, Boston (2009)
3. Marechaux, J.L.: Combining Service-Oriented Architecture and Event-Driven Architecture
using an Enterprise Service Bus. Technical report, IBM Developer Works (2006)
4. Clark, T., Barn, B.S.: A Common basis for modelling service-oriented and event-driven
architecture. In: 5th India Software Engineering Conference, pp. 23–32. ACM, New York
(2012)
5. Weigand, H.: The pragmatics of event-driven business processes. In: 7th International
Conference on Semantic Systems, pp. 211–218. ACM, New York (2011)
6. Vidackovic, K., Kellner, I., Donald, J.: Business-oriented development methodology for
complex event processing: demonstration of an integrated approach for process monitoring.
In: 4th ACM International Conference on Distributed Event-Based Systems, pp. 111–112.
ACM, New York (2010)
7. Levina, O. H., Stantchev, V.: Realizing event-driven SOA. In: 4th International Conference
on Internet and Web Applications and Services, pp. 37–42. IEEE, Washington (2009)
8. Snyder, B., Bosanac, D., Davies, R.: ActiveMQ in Action. Manning Publications, Stamford
(2011)
9. Eugster, P., Felver, P., Guerraoui, R., Kermarrec, A.: The many faces of publish/subscribe.
J. ACM Comput. Surv. 35(2), 114–131 (2003)
10. Data Distribution Service Portal. http://portals.omg.org/dds/
11. Oracle, Java Message Service Specification - version 1.1. http://www.oracle.com/
technetwork/java/docs-136352.html
12. AMQP, Advanced Message Queuing Protocol. http://www.amqp.org/
13. Etzion, O., Niblett, P.: Event Processing in Action. Manning Publications, Greenwich (2010)
 Integrating Business Information Streams in a Core Banking Architecture 433

14. Michelson, B.M.: Event-Driven Architecture Overview – Event-Driven SOA Is Just Part of
the EDA Story. Technical report, Patricia Seybold Group (2006)
15. Cugola, G., Margara, A.: Processing flows of information: From data stream to complex event
processing. J. ACM Comput. Surv. 44(3), 15 (2012)
16. Hurwitz, J., Nugent, A., Halper, F., Kaufman, M.: Big Data For Dummies. Wiley, For
Dummies (2013)
17. Becker, J., Matzner, M., Müller, O., Walter, M.: A review of event formats as enablers of
event-driven BPM. Bus. Inf. Process. 99, 433–445 (2012)
18. Aihkisalo, T.: A performance comparison of web service object marshalling and
unmarshalling solutions. In: 2011 IEEE World congress on Services, pp. 122–129. IEEE,
Washington (2011)
19. Maeda, K.: Performance evaluation of object serialization libraries in XML, JSON and binary
formats. In: 2st International Conference on Digital Information and Communication
Technology and it’s Applications, pp. 177–182. IEEE, Bangkok (2012)
20. Davis, D., Malhotra, A., Warr, K., Chou, W.: Web Services Event Descriptions (WS-
EventDescriptions). W3C Recommendation, World Wide Web Consortium (2011)
21. OASIS Web Services Notification (WSN) TC. https://www.oasis-open.org/committees/
tc_home.php?wg_abbrev=wsn
22. Activity Streams. http://activitystrea.ms/
23. Brown, W.A., Laird, R.G., Gee, C., Mitra, T.: SOA Governance. Pearson Education, Boston
(2009)
24. RabbitMQ, Messaging that just works. http://www.rabbitmq.com/
25. EsperTech, Event Series Intelligence: Esper & NEsper. http://esper.codehaus.org/
26. Davis, J.: Open Source SOA. Manning Publications, Greenwich (2009)
27. Videla, A., Williams, J.W.: RabbitMQ in Action: Distributed Messaging for Everyone.
Manning Publications, New York (2012)
 PRIMROSe: A Graph-Based Approach
for Enterprise Architecture Analysis

David Naranjo(B) , Mario Sánchez, and Jorge Villalobos

Universidad de los Andes, Bogotá, Colombia

{da-naran,da-naran,jvillalo}@uniandes.edu.co
http://ticsw.uniandes.edu.co/

Abstract. Enterprise Models are the central asset that supports Enter-
prise Architecture, as they embody enterprise and IT knowledge and
decisions. Static analysis over this kind of models is made by inspecting
certain properties and patterns, with the goal of gaining understanding
and support decision making through evidence. However, this is not a
straightforward process, as the model in its raw form is rarely suitable
for analysis due to its complexity and size. As a consequence, current
approaches focus on partial views and queries over this model, leading
to partial assessments of the architecture. In this paper, we propose a
different approach to EA analysis, which consists on the incremental
assessment of the architecture based on the interaction of the user with
visualizations of the whole model. We implemented our approach in a
visual analysis tool, PRIMROSe, where analysts can rapidly prototype
custom functions that operate on topological or semantic properties of
the model, combine partial insights for sounder assessments, associate
these findings to visual attributes, and interact with the model using
multiple visualization techniques.

Keywords: Enterprise architecture · Visual analysis · Enterprise

models · Model analysis

1 Introduction
Thirty years since its inception, Enterprise Architecture (EA) has evolved from
a method for reconciling business and IT to a relatively mature discipline. Enter-
prise Modelling is one of the subjects where EA can deliver real value on the orga-
nization, and consists on the development of Enterprise Models (EMs), which
are the embodiment of all the collected information about the enterprise under
several perspectives. Building EMs typically comes with a high price tag that is
payed for when they are analyzed, that is, when additional knowledge is created
by processing and reworking previously defined facts [1]. The knowledge gained
from analyzing EMs serves to support decision making processes in architectural
and stakeholders boards, and to lower risk.
A very important concern with analysis is that it is far from being a trivial
task. Most of the times, analyzing an EM involves the formulation and reformula-
tion of hypotheses, as well as the composition of different insights in order to get
c Springer International Publishing Switzerland 2015
J. Cordeiro et al. (Eds.): ICEIS 2014, LNBIP 227, pp. 434–452, 2015.
DOI: 10.1007/978-3-319-22348-3 24
 PRIMROSe: A Graph-Based Approach for Enterprise Architecture Analysis 435

to sound assessments. Furthermore, the scope of analysis processes is arbitrary

and not necessarily known a priori. It may range from a full-fledged impact
analysis over the entire model, to an in-depth analysis on a specific domain
where issues were detected during the early stages of the analysis process. For
instance, if an analyst needs to assess the business process architecture of an
enterprise, he would use a pertinent and proven method (e.g. Flow Analysis) for
this evaluation, which differs from say, a security (e.g. vulnerability) assessment.
Furthermore, it would be a good idea to use a combination of several methods,
in order to arrive to more powerful insights.
On top of that, the reasoning process behind an analysis is rarely expressed
and documented because it heavily depends on the experience of the analyst.
This lack of a traceability mechanism forces the analyst to guess the rationale
behind past decisions.
Also, we have to take into account that EA modelling tools offer different
features and thus restrict the kind of analysis that they support [2]. Some of
the characteristics that result in limitations include (a) the modelling approach,
(b) the metamodels supported, and (c) their analytical capabilities, which range
from model conformity checks to generation of pre-defined views and the possi-
bility to query the model.
Given known and important characteristics of EMs, such as being large,
complex, typed, and structural in nature [3], visualizations are becoming more
and more used to support analysis methods. However, most modelling tools only
provide the capacity to visualize (by means of diagrams or views) partial models
that are subsets of an EM. While there is the notion of an integration of these
views to form an unified model, it is rarely possible to apply analysis techniques
over the whole EM.
The problem with this, as evidence suggests, is that applying analysis with-
out an overview of the whole model can possibly lead to information loss and
reaching false conclusions [4]. Furthermore, interactive exploration of the large
volumes of data by visual means, appears to be “... useful when a person sim-
ply does not know what questions to ask about the data or when the person
wants to ask better, more meaningful questions” [5]. This precisely reflects what
precedes most ad-hoc analyses and explains why visualizations are progressively
considered less as a product (diagrams) than a medium.
Taking into account the issues discussed above, we consider that a platform
that enables structural analysis of EMs, and is supported by its visual exploration
and interaction, can facilitate the tasks of an analyst. Thus, we can formulate our
research question as follows: How can we provide an useful and flexible method
for inspecting facts on an Enterprise Model, and what is the architecture behind
an analysis tool that supports the visualization of the whole model, displaying
these facts incrementally?
The goal of this paper is to present a conceptual framework –and a tool
that implements it– that addresses these issues. This framework supports the
formulation of analytical functions that enrich the model, and allows their visual-
ization through an extensible set of visualization techniques. This work is based
on the usage of Overview Visualizations that display the underlying topology of
an Enterprise Model and help the analyst to incrementally find new structural
436 D. Naranjo et al.

properties and patterns. This conceptual framework was implemented in PRIM-

ROSe, an advanced platform for the analysis of metamodel-independent EMs,
which provides feedback continuously as new insights are generated during an
analysis process.
The structure of this paper is as follows: First, in Sect. 2, we make a short
literature review of similar approaches. Section 3 offers an overview of our app-
roach, followed by Sect. 4, which provides its conceptual framework. Section 5
describes the architecture and implementation of PRIMROSe, and Sect. 6, illus-
trates the use of the tool with an ArchiMate model. Finally, Sect. 7 will discuss
results and future steps.

2 Related Work
In general, EA Visual Analysis is grounded in the wide array of previous work in
Software Visualization. Of interest is the work of Panas et al. [6], where the authors
describe a framework and architecture for the configuration of model-to-view and
view-to-scene transformations using a graph-based approach. Based on the rela-
tion between visual attributes and views, the authors start from a model graph,
which is translated almost directly to a view graph by filtering unused proper-
ties from the model. The authors also introduce Visual Metaphors as a collection
of common Visual Representations, i.e. families of visual objects fitting together.
These metaphors are used to visualize properties of a model, using techniques such
as graphs, trees, or more complex 3D representations (e.g. city maps).
This framework, however, only deals with visualization: it requires a pre-
processing of the model that is left to the user. Moreover, its focus lies on the
visual analysis of vertices of the graph, giving edges the same importance. This
is a downside, taking into account that EA Analysis is pretty much based on
relationships [3]. Finally, the process is one-way: it is not clear how to travel the
way back from visualization to further analysis.
Chan et al. [7] describe a Visual Analysis tool for bottom-up enterprise analy-
sis, based on the incremental reconstruction of hierarchical data. Analysis is
made by the exploration of the model, starting with an initial view of an entity,
and adding elements to a graph visualization by selecting concepts and relations
in the metamodel. This is complemented with a set of filtering, searching, and
abstraction methods. This bottom-up exploration is useful to manage the com-
plexity of models, and can be a complement to top-down analysis. A downside
of this approach is that it assumes that the analyst knows where to start, which
is a problem in models that span thousands of elements. Furthermore, earlier
processing and analysis is again a prerequisite, and there is no integration with
graphical frameworks.
In the field of Model Driven Engineering, the Eclipse Foundation presents
Zest [8], a library that aims to amplifiy the visual capabilities of modeling edi-
tors. Built on top of the Graphical Editing Framework, it offers a family of graph
layouts, such as Spring, Tree, Radial and Grid algorithms. The library allows the
processing of a model graph, operating in terms of the attributes of the model
elements. Also, it can selectively highlight elements and relations by develop-
ing view operations that modify visual attributes. A disadvantage, however, is
 PRIMROSe: A Graph-Based Approach for Enterprise Architecture Analysis 437

that each view must be developed from scratch. While the authors focus on an
easily to program framework, there is no explicit way to compose and process
independent view operations depending on the outcome of an analysis.
Recent approaches on the Visual Analysis of EA focus on view-based mod-
elling and analysis. Buckl et al. and Schaub et al. [9,10] describe the conceptual
framework and requirements behind the generation of domain independent inter-
active visualizations that comply to predefined stakeholder Viewpoints, linking
an abstract View Model with the EA Information Model. With a focus on non-
technological stakeholders, the authors provide a tool [11] that allows the design
of ad hoc visualizations that filter the model taking into account aspects such
as access rights of a stakeholder to the information.
In [12], Roth et al. further enhance this framework with a pattern matching
algorithm that supports the mapping of information and view models, based on
the information demand and offering. The tool provides a set of configurable
visualization techniques, such as a Gantt Charts, Matrices, Bubble Charts, and
Cluster Maps. While this allows the analysis of the Enterprise Model by non-
technical business experts, it makes difficult to provide flexible and specialized
analysis to architects. As described in Sect. 1, the generation of these views deal
with the communication of the architecture.
In summary, there are some aspects that current research is not addressing,
leaving a gap in the field of Enterprise Model Analysis:

– We could not find approaches that allow the composition/combination of dif-

ferent analysis methods, i.e. incremental processing of the model by operating
on previous analysis routines.
– Approaches seldom provide a clear division between analysis and visualization,
the latter commonly being just a product of the analysis, e.g. diagrams, instead
of a medium for interactive analysis.
– Support for ad-hoc analysis is limited, and often implies the development of
tailored analysis tools from scratch.
– We could not find approaches that take full advantage of the several topo-
logical properties of Enterprise Models seen as networks/graphs, such as the
differentiation of relations between elements, discovery of paths, clusters, or
graph metrics.
– Current approaches are often tied to a concrete graphical library/framework,
offering a limited set of visualization techniques. In addition, the composition
of several techniques on the same representation is not possible.

3 Visual Analysis of EA Models

Visual Analysis takes advantage of the ability of people to discover patterns
easily, and revolves around giving shape - or Finding the Gestalt [13]- of infor-
mation, in order to uncover outliers, bad smells, and interesting or unusual
groups/clusters. In this aspect, the human visual system is one of the most
sophisticated in nature, and shape is one of the most important visual attributes
to characterize objects [14].
438 D. Naranjo et al.

On the other hand, the complexity of Enterprise Models demands new meth-
ods for inspecting their properties and finding interesting facts about them.
Thus, Visual Analysis appears as a valuable field with several ideas that we can
take advantage of. This section will describe the Visual Analysis process that
starts with an Enterprise Model and ends with the results of analysis that derive
on assessments about the architecture. This kind of analysis is incremental, and
it is guided by the interaction of the analyst with the model.

3.1 Visual Exploration and Interaction

In their study about the interactive nature of visualizations, Chi and Riedl [15]
provide a conceptual model and a classification of interactive tasks. They propose
the notion of operators that transform a data model under a series of stages in a
Visualization Pipeline (see Fig. 1). This results on a view of the data, mediated
by Analytical and Visualization models.
Wickam et al. [16] take this idea further, asserting that any visualization tech-
nique has the (often implicit) notion of a pipeline. However, they also mention
the fact that this pipeline metaphor breaks down when user interaction is consid-
ered: on each transformation stage of the process, user interaction (e.g., grouping,
collapsing, zooming) can take the visualization to another transformation step.
Thus, instead of being sequential, the Visual Analysis processes operate in a sense-
making loop, or dialog, between the user and the data in a visual form.
In the context of EA, Schaub et al. [10] describe a conceptual framework
with requirements for interactive visualizations of EA models. In particular,
the interaction type is selected depending on the type of analysis required. For
example, in order to perform ‘what-if’ analyses, the framework provides the
means to generate dynamic views conformant to a viewpoint metamodel, thus
aligned to the concerns of a stakeholder and his access to information.

3.2 A Process for EA Visual Analysis

At this point we want to make a parallel with the field of Visual Analytics, which
can be regarded as the transformation from data to insights by a concatenation
of several sub processes, such as visualizing data sets and generating hypothetical
models from them.

Fig. 1. Visualization pipeline from [15].

 PRIMROSe: A Graph-Based Approach for Enterprise Architecture Analysis 439

Fig. 2. The EA visual analysis process.

Visualization is a semi-automated analytical process, where humans and sys-

tems cooperate using their respective distinct capabilities for the most effective
results [17]. The user modifies visualization parameters repeatedly [18], allowing
the analyst to gain insights by directly interacting with the data, and coming up
with new hypotheses that can be validated, again with visual interaction [19].
This process is based on the economic model of visualization proposed by
van Wijk [5,20], where a visualization is a time-dependent image (instead of a
static one), and a gain in knowledge is based on the perception of the image and
knowledge acquired from previous interactions. Inspired by this model, we define
EA Visual Analysis as an iterative process between an Analyst and a Visual
Analysis System, where hypotheses are generated and refined by the means of
interaction with Visualizations.
This process, described in Fig. 2, begins with an Import of the Enterprise
Model, which is transformed into a graph structure (part of the Analytical
Abstraction - see Fig. 1). This model can be Analyzed, i.e. processed under a
series of Analytical Functions that operate in terms of its structure and seman-
tics, adding new knowledge that was not explicitly present in the model. For the
first iteration of the process, this stage will be a lightweight processing, as our
priority is to visualize and explore the model in its totality.
Posterior to this processing, the analyst is able to Visualize the model struc-
ture with several Visualization Techniques. We use these visual representations
as a memory aid to amplify cognition - that means, we transform data into
images to derive insights, using pattern recognition from the human visual sys-
tem to process visual information.
As the analyst starts to Interact with visualizations of the model, Hypothe-
ses (which start as expectations, i.e. weak formulations) get refined over time.
This interaction modifies the parameters of a visualization, both with view and
data operators. These last operators parametrize and activate the Analytical
Functions for further processing of the model.
440 D. Naranjo et al.

Within each iteration, these formulations are confirmed or denied, as the

analyst starts to associate visual patterns with EA patterns [21] that are present
from knowledge and experience. Finally, when the Analyst has acquired sound
insights on the model, he is able to Communicate results from the analysis in
terms of Assessments of the architecture.

3.3 Requirements for Visual EA Analysis

The complexity of depicting large models has been largely examined, and two key
concerns that surface in their visualization are: (a) the use of algorithms for the
automatic placement of elements of the model to minimize visual complexity [22],
and (b) the need of automated abstraction mechanisms that reduce information
overload [23].
With these issues in mind, Naranjo et al. [4] defined a collection of requirements
from the Visual and EA Analysis perspectives. By exploring the concept of ‘holis-
tic’ or ‘total’ overview visualizations, and in the context of Visual Analysis applied
to Enterprise Models, these requirements were used to evaluate the gap between
what is currently offered by popular EA modelling tools, and what is possible
with general purpose visualization toolkits. EA Analysis Requirements provide
the guidelines that complement and support the process described in Sect. 3.2.
Another important, but often overlooked issue, is to maximize the effective-
ness of these visualizations, that means, to provide an overview of the model that
is expressive enough to support the tasks of an analyst. In [3], the authors examine
the effectiveness of four overview visualization techniques: Force-directed graphs,
Radial graphs, Sunbursts and Treemaps, and further prescribe use cases (i.e. Ana-
lytical Scenarios) for EA Visual Analysis. These cases include the diagnosis of
Enterprise Models, that is, to discover anomalies in their structure, such as iso-
lated sub-graphs of the model. This preemptive aspect of analysis is largely unex-
plored, but we consider that it is where valuable insights are generated, in the
same manner as a physician can identify pathologies with a view to a MRI scan.

4 PRIMROSe: Conceptual Framework

The goal of this section is to present the conceptual framework at the base of
PRIMROSe, and the way it is structured to support the application of analysis
and visualization functions.
Figure 3 presents a trivial model that will be used throughout this section
to illustrate our conceptual framework. This figure represents a small excerpt
of an Enterprise Model that relates elements from domains such as Strategy,
Infrastructure, or Business Process Architecture. In this model, A and C are
Business Processes, while B is an Application Component and D is a Macro
Process that references processes A and C. For the purpose of illustrating an
analysis over this model, we will try to assess the consequences of removing
process A. This should have an impact on B, C and D, and also implies the
removal of relations a, b and e.
 PRIMROSe: A Graph-Based Approach for Enterprise Architecture Analysis 441

Fig. 3. Small fragment of an enterprise model.

Fig. 4. Transformation stages of the enterprise model, from the model graph to a visual
graph.

In order to support even simple analysis such as the one presented in the
previous section, it is necessary to have the capacity to identify or select, and
group, individual elements in the model. Taking into account that in Enterprise
Architecture analysis working with the relations is as important as working with
the elements, the underlying data structures for the analysis are not the raw
Enterprise Models. Instead of that, our approach is based on graphs that are an
homeomorphism on the EM, which means that they are topologically equivalent
to it [24], but make relations first-level elements. We now describe these graphs
and the way in which they are built, starting from what we call a Model Graph
(see Fig. 4(a)).

Model Graph. A directed graph GM = (V (GM ), E(GM )), where V (GM ) is a

set of vertices and E(GM ) is a set of edges. Each vertex in V (GM ) references one
element of the original Enterprise Model, along with its attributes. Each edge in
E(GM ) references a relationship in the model between the corresponding pair of
elements.

The second data structure, which can be built directly from the Model Graph,
is what we call an Expanded Graph (see Fig. 4(b)):

Expanded Graph. A directed

 and bipartite graph GE = (V (GE ), E(GE )),
where V (GE ) = V (GM ) E(GM ) is the set of vertices, and E(GE ) is the set
442 D. Naranjo et al.

of edges. Each of these edges connects a vertex from V (GM ) and an edge on
E(GM ), or the other way around.

The Expanded Graph contains exactly the same information as the original
Enterprise Model, that is, no new knowledge has been added. In order to do
so, and thus really start the analysis process, we need to define the third data
structure, which is precisely what we call the Analysis Graph:

Analysis Graph. A directed graph GA = (V (GA ), E(GA )), where V (GA ) =

V (GE ) S is the set of vertices, and E(GA ) is the set of edges, each one con-
necting a pair of vertices. S is the set of new vertices that are not present in
V (GE ), and they are called Selectors.

Where this graph differs from the previous one is on the introduction of a special
type of vertex called Selector, which serves to group vertices in GE , that is,
elements or relationships of the original EM:

Selector. A vertex in V (GA ) that is not present in V (GE ), but has edges that
point to vertices of V (GE ).

Ultimately, selectors are the elements in an Analysis Graph that reify the knowl-
edge acquired through an analysis process. Within the proposed framework,
selectors are added by means of the application of functions that operate over
Analysis Graphs:

Analysis Function. Is a graph rewriting function fA : GA × P → GA , with

P = {p0 , p1 , ..., pn } the set of parameters, which vary depending on the specific
function. These functions have an unique identifier, and can be either (1)Generic,
i.e. can be applied to any Enterprise Model, or (2)Metamodel specific. An Analy-
sis Function may add additional attributes to existing vertices.
Considering these functions and the available data structures, we can now illus-
trate the analysis process applied to the sample model. For this, we now define
5 atomic operations which incrementally process GA (see Fig. 5) and ultimately
result in a graph where it is trivial to answer the question “which elements will
be affected by the removal of process A (Product Sales)?”.
– f0 is a function that adds the domain of an element as an attribute. In the
example, elements A, B, and C are grouped in the same domain because they
represent ArchiMate [25] concepts, while element D is classified in another
domain.
– f1 is a function that adds a selector (S1 ) which groups edges that connect
elements from different domains.
– f2 is a function that adds a selector (S2 ) to vertices that satisfy an expression
entered as a parameter. In this case, the only element selected is A, which
refers to the Process where the attribute name equals “Product Sales”.
 PRIMROSe: A Graph-Based Approach for Enterprise Architecture Analysis 443

Fig. 5. Analysis pipeline for the example.

– f3 is a function that selects neighbors of a given vertex. In this case, it selects

the elements B, C, and D, that are pointed by A, and introduces the selector
S3 to reference them.
– f4 is a function that selects the relations between elements in groups of ele-
ments defined by selectors. In this case, it selects a, b, and c, which are the
relations between elements selected by S2 and S3 .

After all the sequence of Analysis Functions is executed, we have GA with

all the additional facts added, and affected vertices pointed by selectors. Now
we have to translate this data into visual information. To do this, we use Visual
Rules, that are pairs of (1) a Visual Attribute (e.g. color, shape, position) and
(2) a function that returns the value of such attribute for a given vertex:

Visual Rule. A rule Ri = {A, fR } is a pair of A a Visual Attribute, and

fR : V → X a function, where V ∈ V (GA ) and X is a value of A.

A Visual Rule returns the value (e.g. ‘red’) of a visual property (e.g. color) when
an attribute of a vertex holds certain condition. In order to apply multiple rules
to a set of vertices pointed by a selector, we use a Visual Decorator :

Visual Decorator. A function fV : GA × Si × R → GA , with Si a given selec-

tor, that removes Si and its associated edges, and applies a set of visual rules
R = {R0 , ..., Rn } to all target nodes of such edges.

This function effectively removes the selector, and enriches the selected nodes
with visual information. However, we need a final structure that embodies the
translation of an Analysis Graph to a Visual Graph:

Visual Graph. A directed graph GV = (V (GV ), E(GV )) that is a subgraph

of GA , where V (GV ) is a set of n vertices, such that n = |V (GV )| = |V (GE )|.
Each vertex contains only visual properties.

Having GV with all the information needed to visualize the analyzed model, we
can generate all the necessary graphic files of a given format (e.g. GraphML,
444 D. Naranjo et al.

JSON, DOT) required by a particular graphical toolkit. Its result is a Visual-

ization that the architect uses to propose and validate (or deny) hypotheses in
a sense-making loop. Finally, as the architect finds some evidence that supports
his suspicions, he would like to focus on the relevant elements and relations,
and communicate his assessments to decision makers. For this purpose we use
Filters:

Filter. A function fF : GV × P r → GV , that removes vertices (and their respec-

tive edges) of GV that satisfy a predicate P r.

These filters are applied sequentially, and the resulting graph is exported to a
model, which is a View of the original Enterprise Model.

5 PRIMROSe: Application Framework

Before presenting the architecture that realizes the conceptual framework, we

highlight some aspects to bear in mind:

Incremental Analysis. As described by the process in Fig. 2, EA Visual Analy-

sis is incremental, starting with lightweight processing in the initial stage of
analysis, and with the application of additional processing on demand, given by
the interaction of the user with the tool. In this order of ideas, Analysis Func-
tions should be applied in a composite manner, e.g. as a pipeline (see Sect. 6.2),
with functions given in terms of Selectors created on previously applied functions
(see f3 and f4 of Fig. 5).

Reusable and Extensible Functions. One of the pillars of PRIMROSe are

analysis functions described in terms of elements of the model and/or metamodel,
complemented with basic graph functions that are independent of the EM and
its metamodel.

Non-destructive Analysis. As it could be noted by the reader in Sect. 4,

Analysis Functions cannot remove nodes from the Analysis Graph. Filtering is
made explicit by the user in terms of the visualization, not the data, i.e. elements
are visually hidden, but present in the Analysis Graph.

Independence from the Visualization Toolkit. Currently, there is no

general-purpose graphical toolkit that satisfies all of the visual requirements
for the Visual Analysis of EMs [4]. Each one has its own strengths in various
aspects, so the user should select which one to use, depending on the visualiza-
tion technique and capabilities needed for a specific analysis, or design their own
graphical library for EA-specific visualizations.
 PRIMROSe: A Graph-Based Approach for Enterprise Architecture Analysis 445

Customizable Visualizations. Selectors of the Analysis Graph must be

mapped to visual attributes of a visualization. This mapping has to be translated
into toolkit-specific code and input data.

5.1 Architecture

The PRIMROSe architecture is divided into four main components, and four
stages (Analysis, Mapping, Visualization, and Filters) that conform the
global pipeline (see Fig. 6):

Fig. 6. PRIMROSe architecture.

Graph Manager. This component is responsible of translating the Enterprise

Model into an Expanded Graph GE through the transformations defined in
Sect. 4.

Project Manager. This component administers the different projects of an

user. A Project is composed of an Enterprise Model, as well as the aggregate
descriptors needed for the analysis. A Pipeline Descriptor is a sequence of Oper-
ations, each one with its own parameters. For instance, an operation of the
Analysis Pipeline contains an instance of an Analysis Function to be executed,
as well as the parameters needed for the function. The Visual Pipeline, on the
other hand, contains the Visual Decorators. Finally, the Filter Pipeline is a
sequence of Filters.

Pipeline Engine. Its purpose is to manage and apply the different stages of
the global pipeline. For each stage, that also is a pipeline, the engine composes
and applies the respective operations sequentially.

1. Analysis: Involves the transformation of the Analysis Graph using the

Analysis Functions referenced in the Analysis Pipeline.
2. Mapping: Processes the Visual Decorators in the Visual Pipeline and builds
the Visual Graph.
446 D. Naranjo et al.

3. Visualization: Taking into account that each toolkit ‘knows’ how to visualize
a graph (or a similar structure) in its own fashion, the Visualization stage
translates the processed graph GV into tool-specific artifacts, such as code
and/or input files, e.g. a GraphML or a JSON file that contains the data to
be visualized.
4. Filters: Transforms the Visual Graph by removing vertices and edges with
the Filters in the Filter Pipeline.

User Interface. This component serves both as an editor for the Pipeline
Descriptors, and also as a workbench for visual Analysis. The user can explore
the model through View Operators that modify the visualization without further
processing or visual mapping (e.g. zooming or panning), as well as control the
different stages of the global pipeline.

6 Illustrating the Approach

In order to illustrate the implementation and interaction with the tool, we will
make use of a small case study using ArchiMate as Enterprise Metamodel. As
described by its authors in [26], the ArchiSurance case study is an example of a
fictitious insurance company, which has been formed as the result of a merger of
three previously independent companies, and describes the baseline architecture
of the company and then a number of change scenarios.
We start by creating a Project containing the model, which is imported by
the Graph Manager (see Fig. 7). Until now, we have not done much.
We would want to analyze this model in order to acquire new knowledge. For
instance, and for the sake of a small example that fits in this paper, we asked
ourselves some questions: What is the relative importance of each element in
the model? Also, related to this, what is the impact of a change in the model?,

Fig. 7. ArchiSurance model graph

 PRIMROSe: A Graph-Based Approach for Enterprise Architecture Analysis 447

e.g. the removal of the Customer Business Role. In order to perform this kinds
of analyses, we need (1) A function that finds the connectivity degree of each
node, (2) A function that finds paths and neighbors of a given model element.

6.1 Analysis Functions

Each Analysis Function comes with an unique identifier, and receives as input
the Analysis graph, in addition to custom parameters defined by the creator of
the function. Its output is the modified graph, with additional selectors, and
even new attributes.
This selection of relevant vertices is the backbone of Analysis Functions. In
order to select subsets of vertices of the graph, the Pipeline Engine exposes some
methods to Analysis Functions that allow them to query the model.
In order to elucidate the definition of Analysis Functions, we created two
very simple functions that operate with the topology of the model:

Degree Calculator. Adds an attribute –degree– to every vertex that represents

an element of the model. Its value is the number of incoming and outgoing edges.
Finally, it adds a selector for all the modified vertices of the Analysis Graph.

Impact Analysis. Given the id of a vertex, it adds a selector to adjacent

vertices and their respective edges. It also adds a selector to the source element.
The following snippet of code shows the Java class that is used to implement
the Impact Analysis function:

public class ImpactAnalysis extends AnalysisFunction{

private PipelineEngine engine;

...
public static Collection<Selector> execute(Collection<Parameter>
parameters) {
ArrayList<Selector> selectors = new ArrayList<Selector>();
Long source = parameters.iterator().next().getValue().asLong();

List<Vertex> neighbors = engine.getNeighbors(source,

pipelineEngine.getAnalysisGraph());
List<Vertex> relations = engine.getIncomingEdges(source);

Vertex s0 = new Selector();

s0.addAttribute("name","sourceElement",String.class);
selectors.add(s0);

Vertex s1 = new Selector();

s1.addAttribute("name","impactedElements",String.class);
selectors.add(s1);
448 D. Naranjo et al.

AnalysisGraph graph = engine.getAnalysisGraph();

Long id = graph.addVertex(s1);

for(Vertex v : neighbors){
graph.addEdge(id, v.getId());
}

for(Vertex v : relations){
graph.addEdge(id, v.getId());
}
}
...
}

6.2 Analysis Pipeline

As described in Sect. 3, analysis is a dynamic process where the flow of control is
constantly changing between the user and the system, with incremental process-
ing oriented by the addition of operations to the Analysis Pipeline. At the same
time, analysis is also incremental, where insights are obtained progressively.
This is the Analysis Pipeline for our example, in JSON format:
[
{
"id": "o1",
"function": "org.primrose.functions.DegreeCalculator",
"parameters": []
},
{
"id": "o2",
"function": "org.primrose.functions.ImpactAnalysis",
"parameters": [{"source","9"}]
}
]

The pipeline starts by executing the DegreeCalculator function, which does

not require any parameter. After the execution finishes, we run the Impact
Analysis function, that receives the ID of the source vertex as a parameter.
We can create more complex pipelines, where we would like to combine dif-
ferent functions. However, at the same time, we would like to preserve the encap-
sulated nature of each function. For these reasons, the Pipeline Engine also has
a registry of the selectors created on each function, and a method that returns
them given a function ID.

6.3 Visual Pipeline

The Visual Pipeline is a series of Visual Decorators. Each one of them is in
control of a Visual Attribute, is assigned to a selector, and has its respective set
of Visual Rules:
 PRIMROSe: A Graph-Based Approach for Enterprise Architecture Analysis 449

[
{"id": "v1",
"function": "elementSize",
"parameters":
[
{"selector":"degreeElements"},
{"attribute":"size"},
{"rules":[{"expression":"allElements",
"value":"attribute(degree)"}]}
]},
{"id": "v2",
"function": "impactedElementsColor",
"parameters":
[
{"selector":"impactedElements"},
{"attribute":"color"},
{"rules":[
{"expression":"allElements","value":"#FF4D4D"},
{"expression":"allRelations","value":"#FF4D4D"}
]}
]
}
]

The Mapping component works over the Visual Graph, which is a graph
that only contains visual information, despite having the same structure as the
Expanded Graph. In this case, we are inserting visual attributes defined on each
Visual Decorator:
First, we obtain all the elements of the Analysis Graph pointed by the selector
degreeElements, which is the selector added by the DegreeCalculator function,
and then insert the value of the attribute degree into the size attribute of the
Visual Graph.
For the second Visual Decorator, we obtain all the elements of GA pointed by
the selector impactedElements, which is the selector added by the ImpactAnalysis
function, and then insert the value #FF4D4D, a shade of red, into the color
attribute of the Visual Graph.

6.4 Results

Figure 8(a) is the visualization when we execute the first function, DegreeCal-
culator, and Fig. 8(b) is the final visualization, where we have executed both
functions. Here we can identify clearly both questions that we formulated at the
beginning of this section (see Fig. 9).
450 D. Naranjo et al.

Fig. 8. Incremental appliance of analysis functions.

Fig. 9. Detail of Fig. 8(b)

7 Conclusion

This paper delineates the Visual Analysis of Enterprise Models, emphasizing on

the interactive nature of this activity, and taking into account that there is a
reasoning process – which goes in parallel– in the brain of the analyst.
Seeing this analysis more as a dialogue than the production of automated
and partial results, the contribution of this paper lies in the conceptual frame-
work and architecture that enables the incremental production and refinement
of hypotheses that end in assessments that support decision making.
We designed this PRIMROSe framework (and tool architecture) supported
by a set of requirements from various perspectives, also taking into account the
structural properties of Enterprise Models. Analysis over these models is made
with non-destructive functions that select and decorate an analytical abstraction.
This Analytical Model is then mapped to a Visual Graph, which is transformed
 PRIMROSe: A Graph-Based Approach for Enterprise Architecture Analysis 451

into the necessary artifacts that are needed to depict the results on a given
visualization toolkit. The user interacts with the visualization and returns the
flow of control to the system, allowing the user to deepen on more detailed
analyses.
We omitted the last stage of the process, Communicate (see Fig. 2), which
deals with the transformation of a visualization and its associated insights into
a view by using filters. We think this is the meeting point between PRIMROSe
and similar approaches (see Sect. 2) that complement and enhance analysis.
Extension points for the framework include the traceability of the whole
process, which seems a promising field for complementing and enhancing EA
documentation, as it would provide evidence of the rationale behind analysis.
Moreover, the Analysis Graph should be preserved throughout the lifecycle of
the Enterprise Model, as it allows the preservation of the additional facts that
are introduced.
On the other hand, we are currently evaluating and augmenting the tool with
more complex scenarios involving different EMs of large enterprises that span
several thousands of elements and relations. As with every Visual Analysis tool,
user feedback shapes the supported functionality, as well as design considerations
that involve its usability. This evaluation consists of a given Enterprise Model
and a set of Analytical Scenarios, which are complex questions that require some
method of analysis to answer. Users will be invited to use PRIMROSe and fill
a questionnaire addressing both the Analysis Component and the Visualization
Component, in functional (e.g. accuracy, efficiency) and usability (e.g. location
of elements, interactive operations) aspects. This will help us shaping the limi-
tations of the tool, measure its effectiveness, and assess the minimal set of basic
functions that are useful for the different kinds of EA Analysis.

References
1. Buckl, S., Gulden, J., Schweda, C.M.: Supporting ad hoc analyses on enterprise
models. In: EMISA. GI, vol. 172, pp. 69–83. LNI (2010)
2. Schekkerman, J.: How to Survive in the Jungle of Enterprise Architecture Frame-
works: Creating or Choosing an Enterprise Architecture Framework. Trafford,
Victoria (2006)
3. Naranjo, D., Sánchez, M., Villalobos, J.: Connecting the dots: examining visualiza-
tion techniques for enterprise architecture model analysis. In: Grabis, J., Kirikova,
M., Zdravkovic, J., Stirna, J. (eds.) PoEM of Short Paper Proceedings of the 6th
IFIP WG 8.1 Working Conference on the Practice of Enterprise Modeling, CEUR-
WS, vol. 1023, pp. 29–38 (2013)
4. Naranjo, D., Sánchez, M., Villalobos, J.: Visual analysis of enterprise models. In:
Workshops Proceedings of the 16th IEEE International Enterprise Distributed
Object Computing Conference, EDOCW 2012. IEEE Computer Society, September
2012
5. Fekete, J.-D., van Wijk, J.J., Stasko, J.T., North, C.: The value of information visu-
alization. In: Kerren, A., Stasko, J.T., Fekete, J.-D., North, C. (eds.) Information
Visualization. LNCS, vol. 4950, pp. 1–18. Springer, Heidelberg (2008)
452 D. Naranjo et al.

6. Panas, T., Lincke, R., Lwe, W.: Online-configuration of software visualizations with
vizz3d. In: Naps, T.L., Pauw, W.D. (eds.) SOFTVIS, pp. 173–182. ACM (2005)
7. Chan, Y.H., Keeton, K., Ma, K.L.: Interactive visual analysis of hierarchical
enterprise data. In: Proceedings of the 12th IEEE International Conference on
Commerce and Enterprise Computing. CEC 2010, pp. 180–187. IEEE Computer
Society, Washington (2010)
8. The Eclipse Foundation: Zest: The eclipse visualization toolkit (2013). http://
www.eclipse.org/gef/zest/. Accessed 15 October 2013
9. Buckl, S., Ernst, A.M., Lankes, J., Schweda, C.M., Wittenburg, A.: Generating
visualizations of enterprise architectures using model transformations. In: EMISA.
GI, vol. P-119, pp. 33–46. LNI (2007)
10. Schaub, M., Matthes, F., Roth, S.: Towards a conceptual framework for interactive
enterprise architecture management visualizations. In: Modellierung. GI, pp. 75–
90, vol. 201. LNI (2012)
11. Hauder, M., Roth, S., Pigat, S., Matthes, F.: A configurator for visual analysis of
enterprise architectures. In: ACM/IEEE 16th International Conference on Model
Driven Engineering Languages and Systems (MODELS 2013), Miami (2013)
12. Roth, S., Hauder, M., Zec, M., Utz, A., Matthes, F.: Empowering business users
to analyze enterprise architectures: structural model matching to configure visual-
izations. In: 7th Workshop on Trends in Enterprise Architecture Research (TEAR
2013), Vancouver (2013)
13. Buja, A., Cook, D., Swayne, D.F.: Interactive high-dimensional data visualization.
J. Comput. Graph. Stat. 5(1), 78–99 (1996)
14. Backes, A.R., Casanova, D., Bruno, O.M.: A complex network-based approach for
boundary shape analysis. Pattern Recognit. 42(1), 54–67 (2009)
15. Chi, E.H.-h., Riedl, J.: An operator interaction framework for visualization sys-
tems. In: Proceedings of the 1998 IEEE Symposium on Information Visualization.
INFOVIS 1998, pp. 63–70. IEEE Computer Society, Washington (1998)
16. Wickham, H., Lawrence, M., Cook, D., Buja, A., Hofmann, H., Swayne, D.: The
plumbing of interactive graphics. Comput. Stat. 24, 207–215 (2009)
17. Kohlhammer, J., May, T., Hoffmann, M.: Visual analytics for the strategic decision
making process. In: De Amicis, R., Stojanovic, R., Conti, G. (eds.) GeoSpatial
Visual Analytics. LNCS, pp. 299–310. Springer, Heidelberg (2009)
18. Jankun-Kelly, T., Ma, K.L., Gertz, M.: A model and framework for visualization
exploration. IEEE Trans. Vis. Comput. Graph. 13(2), 357–369 (2007)
19. Keim, D.A.: Information visualization and visual data mining. IEEE Trans. Vis.
Comput. Graph. 8(1), 1–8 (2002)
20. van Wijk, J.: The value of visualization. In: (2005) Visualization VIS 2005, pp.
79–86. IEEE, October 2005
21. Buckl, S., Ernst, A., Lankes, J., Matthes, F., Schweda, C.: Enterprise architecture
management patterns - exemplifying the approach. In: (2008) 12th International
IEEE Enterprise Distributed Object Computing Conference EDOC 2008, pp. 393–
402, September 2008
22. Fruchterman, T.M.J., Reingold, E.M.: Graph drawing by force-directed placement.
Softw. Pract. Exper. 21, 1129–1164 (1991)
23. Egyed, A.: Automated abstraction of class diagrams. ACM Trans. Softw. Eng.
Methodol. 11, 449–491 (2002)
24. Ray, S.: Graph Theory with Algorithms and its Applications. In Applied Science
and Technology. Springer, Heidelberg (2012)
25. The Open Group: ArchiMate 2. 0 Specification. Van Haren Publishing (2012)
26. Jonkers, H., Band, I., Quartel, D.: Archisurance case study. Accessed 4 April 2014
 A Review and Evaluation of Business Model
Ontologies: A Viability Perspective

A. D’Souza1(B) , N.R.T.P. van Beest3 , G.B. Huitema2 ,

J.C. Wortmann2 , and H. Velthuijsen1
1
School of Communication, Media and IT, Hanze University of Applied Sciences,
Zernikeplein 11, Groningen, The Netherlands
{a.d.souza,h.velthuijsen}@pl.hanze.nl
2
Department of Operations, University of Groningen, Nettelbosje 2,
Groningen, The Netherlands
{g.b.huitema,j.c.wortmann}@rug.nl
3
Software Systems Research Group, NICTA Queensland, 70-72 Bowen Street,
Spring Hill, 4000 Brisbane, Australia
[email protected]

Abstract. Organisations are increasingly becoming interdependent in

order to create and deliver superior value to their customers. The result-
ing business models of such organisations are becoming increasingly com-
plex and difficult to design, because they have to deal with multiple
stakeholders and their competing interests, and with dynamic and fast
paced markets. Hence, in order to ensure the long-term survival of such
firms, it is crucial that their business models are viable. Business model
ontologies (BMOs) are effective tools for designing and evaluating busi-
ness models. However, the viability perspective has largely been ignored,
and the current BMOs have not been evaluated on their capabilities to
facilitate the design and evaluation of viable business models. In order
to address this gap, current BMOs have been assessed from the viabil-
ity perspective. To evaluate the BMOs, a list of 26 criteria is derived
from the literature. This list of criteria is then applied to assess six well-
established BMOs. The analysis reveals that none of the BMOs satisfies
all the criteria. However, the e3-value satisfies most of the criteria, and it
is most appropriate for designing and evaluating viable business models.
Furthermore, the identified deficits clearly define the areas for enhancing
the BMOs from a viability perspective.

Keywords: Business model ontology · Business model assessment · IS

alignment · Viability

1 Introduction
Organisations operate in an increasingly dynamic and networked setting that
involves many stakeholders. The resulting business complexity requires new busi-
ness models. Consequently, this also requires a change in the corresponding enter-
prise architecture (EA).
c Springer International Publishing Switzerland 2015
J. Cordeiro et al. (Eds.): ICEIS 2014, LNBIP 227, pp. 453–471, 2015.
DOI: 10.1007/978-3-319-22348-3 25
454 A. D’Souza et al.

An important example of the increased complexity is the energy industry.

The influx of new technologies, changing customer needs, environmental con-
cerns, and government policies, is causing a shift towards a decentralised energy
industry. This also implies a shift from a centralised monopolistic or linear logic
of value creation to a decentralised non-linear logic of value creation. The decen-
tralised energy industry is characterised by cleaner decentralised energy gener-
ation assets (e.g. solar panels), smart grids, and new stakeholders, such as the
prosumers (both large and small scale). The traditional consumers are taking on
the role of prosumers, who not only consume energy, but also produce energy
with the help of technologies such as solar panels. Consequently, the traditional
simplistic and linear business models have to be changed to deal with such new
stakeholders and technologies. Furthermore, new services are needed to support
such new stakeholders. For example, the amount of energy delivered to the grid
by the prosumers will have to be metered and they have to be compensated for
it. This means that the appropriate metering, billing, and accounting services
will have to be designed and implemented. In addition, the energy industry is
a heterogeneous mix of stakeholders with conflicting interests. For example, the
system operator (responsible for maintaining the grid infrastructure) is interested
in maintaining a robust infrastructure at minimal cost; the political stakeholders
are interested in promoting green energy in order to reduce CO2 emissions, and
in creating new jobs; the prosumers are interested in a sustainable lifestyle, and
at the same time want lower energy bills. Hence, this shift towards a decentralised
energy system increases the technical and business complexity [1]. Consequently,
the enterprise architecture1 (EA) as well as the physical technologies (e.g., solar
panels, windmills) that support this new market setting are highly distributed.
Furthermore, its components are owned and managed by different stakeholders.
It is essential that such complex enterprise architectures (comprising the busi-
ness processes, enterprise information system (EIS) infrastructure, and the phys-
ical technologies) are derived from a viable business model. That is, all the
participating organisations and stakeholders are able to capture value, such that
they are committed to the overall business model [3]. In addition, the envisioned
business model should be technologically viable [4]. However, this new way of
doing business in a networked setting dramatically increases the complexity of
designing viable business models, due to the competing interests of stakeholders
and the new technologies that enable new ways of creating value [3]. Moreover,
the technology infrastructure not only has to align with the overall business
model [4], in some cases it is an explicit part of the business model.
The design and evaluation of business models is supported by business model
ontologies (BMOs). BMOs can also be used to conceptualise and communicate
business models [5]. The existing BMOs are conceived from different perspectives
and are used for different purposes. However, the viability perspective has been
largely neglected. As such, the capabilities of BMOs to support the design and
1
An enterprise architecture is defined as “a coherent whole of principles, methods, and
models that are used in the design and realisation of an enterprise’s organisational
structure, business processes, information systems, and infrastructure” [2].
 A Review and Evaluation of Business Model Ontologies 455

evaluation of viable business models remains unclear, particularly in complex

business settings.
Therefore, this paper presents a list of fundamental criteria to which a BMO
should comply in order to facilitate the design and evaluation of viable busi-
ness models and, therefore, enterprise architectures. These criteria are subse-
quently used to assess six well-established BMOs. Out of the six ontologies,
four were specifically conceived to represent business models (e3-value, VNA,
BMC, and EBMS), while the other two (VSM and REA) were conceived for
different purposes. The VSM was developed for reorganising production sys-
tems from the lean manufacturing perspective [6]. The REA was originally pro-
posed as a generalised accounting framework, which allows accountants and
non-accountants to maintain information about the same set of phenomena
in terms of resources exchanged, events (economic exchanges), and the agents
involved in the exchanges [7]. However, they could possibly be used as BMOs
(for more details we refer to Sect. 4.2). Therefore, we refer to these modelling
ontologies as BMOs.
The BMOs are assessed against the derived criteria, in order to select the
most appropriate BMO, and to identify the deficits and areas for improvement
from a viability perspective. This will allow future research to enhance BMOs
to fully support viability as an explicit design focus of business models.
Accordingly, the paper is structured as follows. Section 2 provides a discus-
sion on related work. Section 3 describes the criteria as provided in literature.
Subsequently, Sect. 4 applies the criteria to assess current BMOs. Finally, the
paper is concluded in Sect. 5, along with some directions for future research.

2 Related Work

The boundaries of traditional enterprises are shifting from a single organisation

to a network of organisations [2]. This has led to enterprise architectures being
developed, owned, and operated in a highly distributed manner, which in turn
has lead to misaligned and inflexible enterprise architectures [4,8]. Furthermore,
for an enterprise architecture to be effective, it has to enable the business strat-
egy. However, the distance between strategy and enterprise architecture is very
large, because it is hard to conceive and design enterprise architectures based
on general strategy statements [8,9]. Therefore, scholars have argued that the
business model concept helps to address these challenges by conceptualising and
translating the strategy into a blueprint that describes how business is carried
out [8,10–12]. The business model, as shown in Fig. 1, is the linking pin between
strategy and the enterprise architecture [10–12]. As such, it is critical that the
enterprise architectures (i.e., the business processes, technical infrastructure –
both ICT and the physical technological infrastructure –, and the organisational
structure) are derived from or aligned with a viable business model.
The business model concept is relatively young, and scholars are constantly
debating the meaning and scope of the concept [11,13]. They are debating the
scope of business models on several fronts such as strategy and operational detail
456 A. D’Souza et al.

Fig. 1. Business model in relation to strategy and enterprise architecture.

[13]. However, since we consider business models to be a linking pin between

strategy and enterprise architecture, we make the following distinction between
them. A business model translates the strategy into a blueprint that describes
how value is created, exchanged, and captured among the stakeholders. In addi-
tion, it also describes the organising logic of the key components of a busi-
ness model. Strategy is concerned with how to create and maintain competitive
advantage [11,14,15]. At a meta level, the enterprise architecture describes the
process of designing and implementing the organisational structure, business
processes, technology (ICT and physical technologies), and infrastructure. Fur-
thermore, scholars from different disciplines are studying the concept of business
models, which makes it difficult to agree on a common definition. There has been
considerable interest in integrating the multidisciplinary views and arriving at
a common definition [11,13]. It is important to distinguish between business
models and the individual components of the EA in particular, business process
models, because in the past, business process models have been used wrongly
to represent business models [16]. Looking at business models from the viabil-
ity perspective they should focus on value and viability. The business process
models focus on describing how the business model should be implemented. It
describes the activities, the sequence of activities, and the actors performing
them. Business models and business process models are closely related, but they
are not the same. According to [4] modelling business processes is the next logical
step in designing enterprise architectures, after a viable business model has been
designed. Therefore, the business process models should align with the business
model, especially in terms of viability.
Scholars are also interested in the design of viable business models. They
focus on identifying the factors, characteristics, and conditions that lead to
viable business models. Furthermore, they provide guidelines and evaluation
criteria for business models [17–19]. The viability of business models is mainly
studied from two perspectives. Some use a qualitative approach to conceptualise
and analyse viable business models [17,18], whereas others use BMOs to con-
ceptualise, design, analyse, and evaluate viable business models [4,20]. BMOs
are a reliable way of conceptualising, designing and evaluating business models
[11]. They provide a common language to conceptualise and represent business
models, and leave little room for misinterpretation of the business model.
BMOs and business models are closely related. A BMO is a language, which
can be used to conceptualise and communicate any number of business models.
For example, e3-value can be used to conceptualise and communicate business
models of companies [14].
 A Review and Evaluation of Business Model Ontologies 457

There has been some interest in the past to compare BMOs for different pur-
poses. In [5], a framework is proposed to compare BMOs to find similarities and
differences with the goal of integrating BMOs. In [21], a framework is proposed
to assess BMOs from a taxonomical perspective. However, no attempts have
been made to assess existing BMOs from a viability perspective.

3 Derivation of Criteria

This section describes the research design, and how it is applied to distil a set
of criteria, which are subsequently used to assess the BMOs.

3.1 Research Design

Figure 2 visualises the research design. The Conceptual Model Analysis Frame-
work (CMAF) [22] is proposed to help researchers to compare and analyse the
BMOs. The CMAF is a generic and flexible framework, which we will use to
derive the criteria from literature in a systematic manner. The literature review
will yield a list of criteria that we subsequently use to compare, assess, and select
the most appropriate BMO for designing and evaluating viable business models.
Furthermore, the revealed deficits can be used to enhance the BMOs in context
of designing and evaluating viable business models.
However, the CMAF framework assumes that the process of comparing and
analysing BMOs will automatically lead to the selection of an appropriate BMO.
This implies that the user is required to come up with a set of preferences (and

Conceptual model analysis framework (CMAF)

Develop Assessment Framework for BMOs

Development process of the assessment framework

Use CMAF to structure

Derive criteria for
literature review from Subject criteria to MCDA Final list of criteria
assessment
viability perspective

Evaluation of BMOs

Evaluation process of BMOs

Assess whether or not

Select BMOs the BMOs support each Result
of the derived criteria

Fig. 2. Research design.

458 A. D’Souza et al.

hence a subset of relevant criteria) based on which the user will select a BMO.
Consequently, there is a need for a set of criteria specifically tailored to assess a
BMO on its capabilities to facilitate the design and evaluation of viable business
models.
The CMAF framework consists of three dimensions, namely conceptual focus,
unit of analysis, and level of analysis. These dimensions will be used to distil
a set of criteria to assess the BMOs. In addition, we have also made use of
the existing literature in the strategic management domain and the business
model domain (also including BMO literature). Following [23], we analyse the
literature from the domains mentioned above, and distil a set of criteria that are
relevant, understandable, complete and concise, and judgementally independent
(i.e. the preference for one criterion should not be dependent on other criteria).
Furthermore, the criteria are distilled in such a way that they are operational,
which means that they are readily applicable to BMOs. Additionally, attention is
paid to the simplicity versus complexity condition. In order to ensure simplicity
(without sacrificing the complexity) of the criteria and to ensure the criteria are
operational, some of the criteria are further decomposed into a set of lower level
criteria. Finally, in Sect. 4.1 the criteria are checked to ensure that they comply
with the redundancy condition.

3.2 Conceptual Focus

The conceptual focus dimension helps to synthesise a perspective through which

we view business models. It defines the functionality of the BMOs, the compo-
nents that should be modelled and analysed, and the granularity at which they
are modelled. Hence, it strongly influences the unit of analysis, and the level of
analysis dimensions. Therefore, the goal of this subsection is to synthesise a per-
spective through which we view business models. The perspective through which
we view business models is viability. Consequently, the BMO should focus on
the design and evaluation of viable business models. In the following subsection,
we define the concepts of business model and viability.
Business Model. In the continuous debate on the scope of business models,
some common ground can be identified on the definition of a business model [11].
A business model describes how business is carried out; it includes a description
of the stakeholders (e.g. customers and partners), their roles, value proposition
for other stakeholders involved, and the underlying logic of value creation, value
exchange, and value capture at organisational level and at network level. Fur-
thermore, it defines the business architecture (organising logic of all the key
components, such as information, value creation activities, stakeholders, and
value exchange relationships) that enables the value creation, value exchange,
and value capture logic [11,12,24,25].
Viability. A business model should be viable both in terms of technology and
in terms of value [4]. A business model is viable in terms of technology when the
underlying physical technologies can support the business model and when the
 A Review and Evaluation of Business Model Ontologies 459

ICT infrastructure can support the information services required by the busi-
ness model to work. This can be achieved by considering which type of infor-
mation services (and, therefore, the ICT infrastructure) and physical technology
infrastructure is needed to support the business model. In addition, the capa-
bilities of new technologies should be considered to determine how they could
lead to better and new ways of doing business [8]. For example, if the value
proposition of a business model is self-sufficiency in terms of energy use, the
underlying technological infrastructure should support the creation and delivery
of this value proposition. If the underlying technologies are unable to supply
the energy demand then the business model is not technologically viable. Con-
sequently, the customer will not be self-sufficient and as a result the business
model will fail. A business model is viable in terms of value when all the partic-
ipating organisations/stakeholders are able to capture value, such that they are
committed to the business model [3].
Value. Value is the core component of a business model, and it plays an impor-
tant role in making business models viable [12]. Figure 3 shows that value is
composed of exchange value (e.g. euros, dollars etc.) and use value (e.g. ben-
efits of a product or service enjoyed by a customer, benefits derived by other
stakeholders) [26,27]. Use value concerns the desired benefits end users derive
out of a product or a service. The concept of use value was extended by [26]
to include benefits realised by stakeholders other than end users, such as gov-
ernmental organisations and society. This implies that the business model could
include a broader set of stakeholders other than end users and the company
selling the product or a service. This is especially evident in the energy industry.
For example, let us consider a community-owned small-scale solar farm in the
Netherlands. The business model of the solar farm involves several stakehold-
ers, such as the community members, solar farm operators, local municipality,
and the subsidising agency. These stakeholders are interested in different types
of value. For instance, the solar farm operator is purely interested in exchange
value (profits), while the community members are interested in exchange value
(return over investment) as well as use value (reduction of CO2 , jumpstarting

Exchange value
(e.g. revenue)

Value

Use value
(e.g. CO2 reduction,
conservation of the
environment)

Fig. 3. Composition of value.

460 A. D’Souza et al.

the local economy by hiring local suppliers and installers). Similarly, the local
municipality and the subsidising agency are interested in use value, i.e. reducing
CO2 emissions to meet the EU sustainability goals, and at the same time in
boosting the local economy.

3.3 Unit of Analysis

The unit of analysis describes the functionality of the BMOs and the components
to be modelled and analysed. BMOs can model and analyse business models at
organisational level as well as at business network level. The organisational level
analysis includes the components within the organisation, such as value cre-
ation activities. The business network analysis includes the components within
the business network, such as organisations and relationships among them. The
functionality and the components to be modelled are derived from literature,
and are directly influenced by the conceptual focus. In addition, the functional-
ity and the components that should be modelled and analysed are presented as
a list of criteria.
Based on the definition of the business model, we posit that the BMOs should
conceptualise, encompass, and model the following concepts, functionality, and
components. BMOs should conceptualise and model business models at two lev-
els: at organisational level and at network level [22,28]. Further, they should con-
ceptualise the stakeholders within the network, and how they create, exchange,
and capture value. Additionally, they should be able to represent the business
model architecture of viable business models and the business rules that govern
them.
Business models have been studied from different perspectives [11]. Hence,
we further explore these concepts below, to gain a better understanding of the
concepts it encompasses from the perspective of viability and what it means to
BMOs.

Value Creation. Value creation is a central concept in management literature.

It is the increased value (exchange value and use value) that two or more parties
enjoy when they engage in mutually beneficial transactions [27]. Furthermore,
value could also be created for other stakeholders participating in the business
model, even though they do not directly engage in transactional relationships,
for example, political stakeholders providing subsidies to green energy produc-
ers [26]. Business models should be able to generate enough value to keep the
stakeholders committed to the business models. If the stakeholders are unable to
capture enough value, they will not be committed to the business model, which
consequently renders it unviable [3]. Therefore, value creation is crucial from a
viability perspective. In context of business models, value creation cannot be
explained by a single theory, such as a resource based view [28]. Therefore, to
truly understand value creation in context of business models, a nexus of several
theories is necessary [3]. Hence, several theories have been reviewed that utilise
different units of analysis to explain value creation [28]. Further, it is argued
that the units these theories analyse are the sources of value creation [3,28].
 A Review and Evaluation of Business Model Ontologies 461

Following the lead of [3,28], we posit that the BMO should model and analyse
the different sources of value creation, because by definition of business models,
it is crucial to understand how value is created. Table 1 highlights the theories
and the sources of value creation they analyse.
We acknowledge Schumpeters theory of entrepreneurship, which posits that
an entrepreneur is the source of value creation [28]. However, we do not review
this theory in context of this research, because the conceptual focus used to
view viability of business models is at the organisational and the network level,
and not at the entrepreneur level. An overview of the criteria concerning value
creation is shown in Table 2.

Value Capture. Value capture is the amount of value retained by each stake-
holder within the business model [26]. The amount of retained value is concep-
tualised in terms of use value and exchange value. The value captured in terms
of exchange value is the total amount of revenue a stakeholder is able to retain

Table 1. Sources of value creation.

Theory Source of value creation

Value chain framework Value creation activities
Resource based view Resources
Business networks Business networks
Transaction cost economics Transactions (value exchange relationships)

Table 2. Assessment criteria concerning value creation.

No Criteria Sources
1. Value creation [27]
1.1. Model value creation by each stakeholder [25]
1.2. Model sources of value creation [28]
1.2.1. Value creation activities [28]
1.2.2. Resources [28]
1.2.3. Business network [28]
1.2.4. Transactions [28]

Table 3. Assessment criteria concerning value capture.

No Criteria Sources
2. Value capture [27]
2.1. Model value captured by each stakeholder [3]
2.2. Model captured value in terms of use value [27]
2.3. Model captured value in terms of exchange value [27]
462 A. D’Souza et al.

(i.e. profit) [27]. The value captured in terms of use value is the total benefits
realised by a stakeholder from a product and/or a service, or by participating in
a business model [26].
Successful business models ensure that the stakeholders participating in the
business models are able to capture value, such that they are committed to the
business model. If not, the business model will not be able to attract and retain
competitive stakeholders. Hence, this could lead to the business model being
rendered unviable. Consequently, the BMO should be able to model and analyse
the amount of value captured by each stakeholder. An overview of the criteria
concerning value capture is shown in Table 3.

Value Exchange. Value exchanges are relationships formed among stakehold-

ers to exchange value. At the organisational level, these relationships can be
analysed within the organisations, and at a dyadic level. Organisations form
dyadic relationships with partners, such as channel partners and customers [3].
Adopting the business network approach implies that the value exchange rela-
tionships are among the basic building blocks of the business network [20]. Conse-
quently, it calls for a systemic approach, where the value exchanges are analysed
not only from a focal organisations perspective, but also from a business networks
perspective. This involves the analysis of the entire business network formed to
produce, govern, and deliver the products and services to the end user [3]. Hence,
this implies that the BMO should be able to conceptualise and model the value
exchanges at organisational level and at network level. An overview of the criteria
concerning value exchange is shown in Table 4.

Table 4. Assessment criteria concerning value exchange.

No Criteria Sources
3. Value exchange [3]
3.1. Model value exchanged at organisational level [3]
3.1.1. Use value at organisational level [27]
3.1.2. Exchange value at organisational level [27]
3.2. Model value exchanged at business network level [20]
3.2.1. Use value at business network level [27]
3.2.2. Exchange value at business network level [27]

Business Model Architecture. Business model architecture is the organising

logic of how the key components that enable value creation, value capture, and
value exchange relationships are organised [29]. Visualising the business model
architecture helps in gaining a deeper insight into the business model. Further,
 A Review and Evaluation of Business Model Ontologies 463

Table 5. Assessment criteria concerning the business model architecture.

No Criteria Sources
4. Represent the business model architecture [29]

it is a useful and effective technique used to brainstorm and identify alternative

configurations of the business model. Hence, in context of designing viable busi-
ness models, visualising the business model architecture of the business models
is an effective tool in organising the components in such a way that it enables
viability [20]. Therefore, the BMOs should be able to visualise the business
model architecture. An overview of the criteria concerning the business model
architecture is shown in Table 5.

Design of Viable Business Models. The design of viable business models is

an iterative process. Therefore, the ontology should also allow for the manipu-
lation of the business model to achieve viability [20]. Moreover, it is crucial to
consider the capabilities of the underlying technologies (ICT technologies and
physical technologies) while designing business models. For example, let us con-
sider a business model where a prosumer produces electricity using a solar panel.
The produced electricity is delivered to the electricity grid, and the prosumer is
compensated for it. For this business model to work, two layers of technology are
necessary namely the physical technology layer, and the ICT layer. The physical
technology layer consists of technological infrastructure such as the solar panel,
the cables that carry the electricity, the meters that measure how much electric-
ity is delivered to the grid. The ICT layer consists of technologies, which collect
data and processes it into information. The information is necessary to support
the business processes and decision making such as the billing process and trade
decisions [30].
Modelling the underlying technological infrastructure alongside business
models is a good way of improving the business and technology alignment [4].
Modelling the required information services, ICT infrastructure, and the physical
technology infrastructure in essence is a cross-domain exercise, which involves
professionals from the business domain and technological domain. Therefore,
they need a common language in order to be effective. Additionally, talking about
ICT in terms of information services gives technologists and business profession-
als a common language. It facilitates the discussion about business models and
their underlying need for information and, therefore, ICT [31]. Hence, it is impor-
tant to conceptualise and model the underlying information services needed to
support the business model. Further, conceptualising the underlying information
services helps technologists to draft requirements for the ICT architecture. The
ICT architecture in turn supports the information services [31]. This also helps
improve the alignment between business models and ICT [32]. Therefore, the
BMOs should be able to model the underlying information services of a business
model. Furthermore, business models often embody multiple commodities and
include multiple stakeholders [1]. Consequently, the BMOs should also be able
464 A. D’Souza et al.

to model multiple commodities and multiple stakeholders. Business models span

multiple organisations; as a result, a systemic approach should be adopted to
design business models. The systemic approach entails designing business mod-
els at an organisational level, which involves analysing and modelling the value
created and contributed by each individual organisation. It also entails designing
business models at a network level, which entails configuring the organisations
and the value exchanges among them in a way that enables viability. Therefore,
the BMOs should be able to design business modelsat an organisational level as
well as at the business network level. Table 6 provides an overview of the criteria
concerning business model design.

Table 6. Assessment criteria concerning business model design.

No Criteria Sources
5. Design business models [20]
5.1. Ability to manipulate business models [20]
5.2. Model underlying information services [30, 31]
5.3. Model ICT necessary to support the information services [4, 30]
5.4. Model underlying physical technologies [33]
5.5. Model multiple commodities [1]
5.6. Model multiple stakeholders/roles [1]
5.7. Ability to design business models at organisational level [25]
5.8. Ability to design business models at the network level [25]

Evaluation of Viable Business Models. The BMO should facilitate the

evaluation of viability in terms of exchange value and use value. One of the ways
it could help evaluate the viability in terms of value is by generating reports on
the value captured by each stakeholder. Furthermore, technologists can evaluate
the modelled informationservices and the corresponding ICT, and the physical
technologies for technological viability. Table 7 shows the assessment criteria
concerning viability.

Table 7. Assessment criteria concerning viability.

No Criteria Sources
6. Evaluation of business models for viability [11]
6.1. Evaluate use value captured by the stakeholders [20]
6.2. Evaluate exchange value captured by the stakeholders [20]
6.3. Visualise the information services [30, 31]
6.4. Visualise ICT [4]
6.5. Visualise the physical technologies [33]
 A Review and Evaluation of Business Model Ontologies 465

Business Rules. Business rules define constraints, conditions, and policies that
govern a business model. A business rule can be defined as a statement that
affects the value creation, value capture, value exchange, and the underlying
business model architecture of a business model [29]. Business rules internalise
the external requirements put on the business models, such as governmental
regulations and technological limitations [29]. In addition, the business rules
include the internal requirements on the business model, such as requirements
of the strategy on the business model [29]. In context of viability, the business
rules can hamper or facilitate the viability of a business model. Hence, it is
important that the BMO considers the business rules. An overview of the criteria
concerning business rules is shown in Table 8.

Table 8. Assessment criteria concerning business rules.

No Criteria Sources
7. Embodies business rules [29]

3.4 Level of Analysis

Level of analysis corresponds to the level of abstraction or granularity at which
the business models are conceptualised. BMOs conceptualise business models
on a continuum, which ranges from a high level of granularity to a low level
of granularity. The BMOs that conceptualise and analyse business models at a
high level represent less information, and the business models are usually gener-
alisable. Contrarily, BMOs that conceptualise and analyse business models at a
low level represent more information, while the business models are specific to
an organisation or a business network. Consequently, they are not generalisable
[22]. It is clear from the evaluation criteria derived thus far, that the design
and evaluation of viable business models requires large amounts of information.
Therefore, the BMOs should model the business models at a relatively low level
[22]. Table 9 shows the assessment criteria concerning level of analysis.

Table 9. Assessment criteria concerning level of analysis.

No Criteria Sources
8. Model BMOs at low level of granularity [22]

4 BMO Assessment
The criteria presented in the previous section can be used to assess how well
BMOs support the design and evaluation of viable business models. The crite-
ria can be applied qualitatively to assess the characteristics of the BMOs. The
466 A. D’Souza et al.

challenge of assessing BMOs based on a set of criteria can also be framed as

a classic multi-criteria decision analysis (MCDA) problem [23, pp. 1–2]. There-
fore, we have subjected the criteria to MCDA conditions. The following section
elaborates on how the conditions impact the list of criteria.

4.1 Restructuring the Criteria

As mentioned previously, for the criteria to be usable they have to meet the fol-
lowing conditions [23, pp. 55–58]: value relevance, understandable, measurable,
non-redundant, judgementally independent, complete and concise, operational,
and simple (without sacrificing the complexity).
We reviewed the criteria in light of the above conditions. To satisfy the con-
dition of non-redundancy, we have eliminated the criteria of business network
and transactions (1.2.3 and 1.2.4). The business network concept emerges under
the category of design of business models, where we assess whether the business
model is conceptualised at the organisational level and at the business network
level. Similarly, the idea of transactions (value exchange) appears under the
concept of value exchange.

4.2 Selected BMOs

Our search led to six well established BMOs that focus on value. The following
BMOs will be assessed using the criteria described above.
e3-Value. The e3-Value adopts a value constellation (business network app-
roach), where business models span multiple organisations. e3-Value aims at
conceptualising business models and evaluating them for viability [20]. Further,
it aims to create a common understanding of the business models among col-
laborating firms (multi-stakeholder environment) by explicitly visualising the
business models. It aims to improve the alignment between business and ICT.
e3-Value has its roots in computer science and management science [5].
Value Network Analysis (VNA). VNA is rooted in the principles of living
systems. It views business models as a pattern of exchanges between stakehold-
ers. It focuses on both the tangible (e.g. money, and products) and the intangible
(e.g. knowledge) value exchanges among stakeholders [34]. VNA aims to incor-
porate a systemic view (business network) of business models, and the intangible
values into the mainstream business model analysis.
Business Modelling Canvas (BMC). The BMC views business models in
terms of 9 building blocks. The BMC conceptualises business models on the level
of a single organisation and not on the level of a business network. However, the
BMC does identify key partners [35]. The BMC is rooted in information systems
and management science. Their main goal is to help companies conceptualise
how they create, deliver and capture value [5].
Value Stream Mapping (VSM). VSM is based on the concept of lean man-
ufacturing. It conceptualises the flow of value in a value stream. VSM adopts a
 A Review and Evaluation of Business Model Ontologies 467

supply chain approach to map the demand back from customers to raw mate-
rials. Their main goal is to help managers shift their attention from individual
processes to a larger perspective. It is an attempt to shift the focus from individ-
ual processes to the system of interconnected processes required to deliver the
product to the customer [36].
Resource Event Agent (REA). REA is a domain specific (accounting domain)
modelling ontology, which focuses on conceptualising economic resources, eco-
nomic events, economic agents, and the relationships among them. These are
conceptualised from the perspective of a single organisation [7]. It is rooted in
information science and management science. It aims to design flexible account-
ing systems that are better integrated with other enterprise systems and decision
support systems [7].
e-Business Modelling Schematics (EBMS). The EBMS adopts a business
network approach to business models, aiming at e-business initiatives. It adopts
a focal organisation perspective to describe business models that span multiple
organisations. It is rooted in management science and information science. EBMS
was conceived with the aim of helping business executives to conceptualise and
analyse new e-business initiatives [31].

4.3 Assessment

Each of the BMOs is assessed against the criteria. The assessment is carried
out such that if the BMOs fully support the criteria a  sign is assigned, and
if it does not or partially support the criteria the criteria a sign is assigned.
We adopt this method of evaluating the BMOs, because even if the BMOs par-
tially support the criteria it will not lead to an accurate conceptualisation and
analysis of viability. Therefore, it could lead to unreliable design and evalua-
tion of viable business models. Table 10 shows how the six BMOs perform on
the viability criteria derived from literature. It is clear that not all the BMOs
conceptualise business models in the same way. Furthermore, it is evident that
certain important viability criteria are ignored. None of the BMOs conceptualise
value capture and evaluation of business models in terms of use value. Similarly,
none of them conceptualise and evaluate the underlying information services,
ICT, and physical technologies. The business model architectures are only rep-
resented to a certain extent, but not satisfactorily. Therefore, we have rated them
as not supporting the criteria “represent the business architecture”.
Based on Table 10, we observe that none of the BMOs perform satisfacto-
rily on all criteria. However, e3-value satisfies most of the criteria except for
the 9 criteria, namely model captured value in terms of use value, represent
the business architecture, model underlying information services, model under-
lying ICT, model underlying physical technologies, evaluate use value captured
by the stakeholders, visualise the information services, visualise the ICT, and
visualise the physical technologies. The reason why some of the BMOs perform
well against the criteria and some do not could be attributed to the reason that
468 A. D’Souza et al.

Table 10. Assessment of BMOs.

not all of them were exclusively conceived to represent business models. Fur-
thermore, even the ones that were conceived to represent business models were
not designed from the perspective of designing and evaluating viable business
models, except for e3-value. This shows that the viability perspective has been
largely ignored in context of BMOs.
 A Review and Evaluation of Business Model Ontologies 469

5 Conclusions
Viable business models are vital to businesses. However, due to the increased
complexity, the design and evaluation of viable business models has become
very hard. Therefore, we need appropriate tools to reliably design and evaluate
viable business models. BMOs are effective tools for designing and evaluating
business models.
In the past, different frameworks have been proposed to compare BMOs.
However, none of these frameworks compare BMOs from the perspective of
design and evaluation of viable business models. We have addressed this gap
by assessing BMOs on their capabilities to support the design and evaluation of
viable business models. A list of criteria is derived from literature for evaluating
BMOs from a viability perspective. We have subsequently applied these criteria
to evaluate six well-established BMOs, and identified a BMO which is best suited
for the design and evaluation of business models from a viability perspective.
Our analysis reveals that none of the BMOs satisfy all the criteria. Fur-
thermore, each of the BMOs conceptualises business models differently. Our
findings suggest that e3-value is the most appropriate BMO for designing and
evaluating business models from a viability perspective. However, it fails on 9
criteria: model captured value in terms of use value, represent the underlying
business architecture, model underlying information services, model underlying
ICT, model underlying physical technologies, evaluate use value captured by the
stakeholders, visualise information services, visualise ICT, and visualise physical
technologies.
Furthermore it is hard to assess the viability of business models, as current
BMOs have a number of deficits that are particularly important for complex,
distributed business settings. Consequently, these identified deficits provide clear
areas for improvement of each assessed BMO. However, the derived criteria are
relying on our conceptualisation of the term viability and influenced by the
assumption that business models rely on technology (ICT and physical technol-
ogy) for execution.
Our analysis shows that the gap between BMOs and strategic management
is large. Accordingly, future research should work towards enhancing BMOs to
fully support viability as an explicit design focus of business models for dynamic
and complex settings. However, this increases the risk of creating overly complex
BMOs, which are very hard to use and understand. Care should be taken not
to overcomplicate BMOs. Future research should also explore the possibility of
combining the strengths of different BMOs and other modelling ontologies, for
example by using BMC, e3-value, and business process modelling complemen-
tarily to design and evaluate business models from a viability perspective, and
to improve the alignment between business models and EAs. In addition, future
research can also involve the direct application of enhanced BMOs to design
and evaluate the highly complex business models in a distributed setting, and
to improve their alignment with the supporting distributed EAs.
Another interesting direction for future research is to extend the enter-
prise architecture modelling techniques (in particular business process modelling
470 A. D’Souza et al.

techniques) to incorporate the viability perspective in order to better align strat-

egy, business models, and enterprise architecture. It is logical to incorporate
the viability perspective into business process modelling techniques, because it
helps conceptualise and design business processes that align with business mod-
els in terms of value. By doing this, we minimise the risk of designing business
processes that do not add value to the business models. Consequently, this will
lead to better and more effective EISs.

Acknowledgements. This research has been financed by a grant of the Energy

Delta Gas Research (EDGaR) programme. EDGaR is co-financed by the Northern
Netherlands Provinces, the European Fund for Regional Development, the Ministry of
Economic Affairs and the Province of Groningen, the Netherlands. NICTA is funded
by the Australian Government through the Department of Communications and the
Australian Research Council.

References
1. Adhikari, R.S., Aste, N., Manfren, M.: Multi-commodity network flow models for
dynamic energy management-smart grid applications. Energy Procedia 14, 1374–
1379 (2012)
2. Lankshorst, M.: Enterprise Architecture at Work-Modelling, Communication and
Analysis. Springer, Heidelberg (2009)
3. Chesbrough, H., Vanhaverbeke, W., West, J.: Open Innovation: Researching a New
Paradigm. Oxford University Press, New York (2006)
4. Kraussl-Derzsi, Z.: OPerationalized ALignment: Assessing Feasibility of Value Con-
stellations Exploiting Innovative Services. Vrije Universiteit, Amsterdam (2011)
5. Gordijn, J., Osterwalder, A., Pigneur, Y.: Comparing two business model ontologies
for designing e-business models and value constellations. In: Proceedings of the 18th
Bled eConference, pp. 6–8, Bled (2005)
6. Lasa, I.S., Laburu, C.O., De Castro Vila, R.: An evaluation of the value stream
mapping tool. Bus. Process Manag. J. 14, 39–52 (2008)
7. McCarthy, W.E.: The rea accounting model: a generalized framework for account-
ing systems in a shared data environment. Account. Rev. 57, 554–578 (1982)
8. Ross, J.W., Weill, P., Robertson, D.C.: Enterprise Architecture as Strategy: Cre-
ating a Foundation for Business Execution. Harvard Business Press, Boston (2006)
9. Engelsman, W., Wieringa, R.: Goal-oriented requirements engineering and enter-
prise architecture: two case studies and some lessons learned. In: Regnell,
B., Damian, D. (eds.) REFSQ 2011. LNCS, vol. 7195, pp. 306–320. Springer,
Heidelberg (2012)
10. Osterwalder, A., Pigneur, Y.: An e-business model ontology for modeling
e-business. In: 15th Bled Electronic Commerce Conference, pp. 17–19, Bled (2002)
11. Pateli, A.G., Giaglis, G.M.: A research framework for analysing eBusiness models.
Eur. J. Inf. Syst. 13, 302–314 (2004)
12. Al-Debei, M.M., Avison, D.: Developing a unified framework of the business model
concept. Eur. J. Inf. Syst. 19, 359–376 (2010)
13. Lambert, S.: A Review of the Electronic Commerce Literature to Determine the
Meaning of the Term ‘Business Model’. Flinders University of South Australia,
School of Commerce (2003)
 A Review and Evaluation of Business Model Ontologies 471

14. Osterwalder, A., Pigneur, Y.: Clarifying business models: origins, present, and
future of the concept. Commun. Assoc. Inf. Syst. 16(1), 1–25 (2005)
15. Porter, M.: Competitive Advantage: Creating and Sustaining Superior Perfor-
mance. Free Press, New York (1985)
16. Gordijn, J., Akkermans, H., van Vliet, H.: Business modelling is not process mod-
elling. In: Mayr, H.C., Liddle, S.W., Thalheim, B. (eds.) ER Workshops 2000.
LNCS, vol. 1921, p. 40. Springer, Heidelberg (2000)
17. Sharma, S., Gutiérrez, J.A.: An evaluation framework for viable business models
for m-commerce in the information technology sector. Electron. Mark. 20, 33–52
(2010)
18. De Reuver, M., Haaker, T.: Designing viable business models for context-aware
mobile services. Telemat. Inform. 26, 240–248 (2009)
19. Keen, P., Qureshi, S.: Organizational transformation through business models: a
framework for business model design. In: Hawaii International Conference on Sys-
tem Sciences, 2006, Proceedings of the 39th Annual HICSS 2006, vol. 8, p. 206b.
IEEE (2006)
20. Gordijn, J., Akkermans, J.M.: Value-based requirements engineering: exploring
innovative e-commerce ideas. Requir. Eng. 8, 114–134 (2003)
21. Mäkinen, S., Seppänen, M.: Assessing business model concepts with taxonomical
research criteria: a preliminary study. Manag. Res. News 30, 735–748 (2007)
22. Lambert, S.: A conceptual model analysis framework: Analysing and comparing
business model frameworks and ontologies. Ph.D. thesis, International Business
Management Association (IBMA) (2010)
23. Belton, V., Stewart, T.: Multiple Criteria Decision Analysis: An Integrated App-
roach. Springer, Heidelberg (2002)
24. Tapscott, D., Lowy, A., Ticoll, D.: Digital capital: Harnessing the Power of Business
Webs. Harvard Business Press, Watertown (2000)
25. Zott, C., Amit, R., Massa, L.: The business model: recent developments and future
research. J. Manag. 37, 1019–1042 (2011)
26. Lepak, D.P., Smith, K.G., Taylor, M.S.: Value creation and value capture: a mul-
tilevel perspective. Acad. Manag. Rev. 32, 180–194 (2007)
27. Bowman, C., Ambrosini, V.: Value creation versus value capture: towards a coher-
ent definition of value in strategy. Br. J. Manag. 11, 1–15 (2000)
28. Amit, R., Zott, C.: Value creation in e-business. Strateg. Manag. J. 22, 493–520
(2001)
29. Eriksson, H.E., Penker, M.: Business modeling with UML. Wiley, Chichester (2000)
30. Davenport, T.H., Prusak, L.: Information Ecology: Mastering the Information and
Knowledge Environment. Oxford University Press, Cary (1997)
31. Weill, P., Vitale, M.: What it infrastructure capabilities are needed to implement
e-business models. MIS Q. Exec. 1, 17–34 (2002)
32. Henderson, J.C., Venkatraman, N.: Strategic alignment: leveraging information
technology for transforming organizations. IBM Syst. J. 32, 4–16 (1993)
33. Baden-Fuller, C., Haefliger, S.: Business models and technological innovation. Long
Range Plan. 46, 419–426 (2013)
34. Allee, V.: A value network approach for modeling and measuring intangibles. In:
Transparent Enterprise, Madrid (2002)
35. Osterwalder, A., Pigneur, Y.: Business Model Generation: A Handbook for Vision-
aries, Game Changers, and Challengers. John Wiley & Sons, Hoboken (2010)
36. Rother, M., Shook, J.: Learning to See: Value Stream Mapping to Create Value
and Eliminate Muda.-Version 1.3. Learning Enterprise Institute, Cambridge (2003)
 Enterprise-Specific Ontology-Driven Process Modelling

Nadejda Alkhaldi1 ✉ , Sven Casteleyn1,2, and Frederik Gailly3

( )

1
Vrije Universiteit Brussel, Pleinlaan 2, 1050 Brussels, Belgium
[email protected], [email protected]
2
Universitat Jaume I, Av. De Vicent Sos Baynat s/n, 12071 Castellon, Spain
3
Ghent University, Tweekerkenstaart 2, 9000 Ghent, Belgium
[email protected]

Abstract. Different process models are created within an enterprise by different

modelers who use different enterprise terms. This hinders model interoperability
and integration. A possible solution is formalizing the vocabulary used within the
enterprise in an ontology and put this ontology as bases for constructing process
models. Given that an enterprise is an evolving entity, the ontology needs to
evolve to properly reflect the domain of the enterprise. This paper proposes an
enterprise-specific ontology-driven process modelling method which tackles the
two aforementioned issues by assisting the modeller in creating process models
using terminology from the ontology and simultaneously supporting ontology
enrichment with feedback from those models. When the modeller creates a model,
matching mechanisms incorporated in the method are working together to suggest
a list of ontological concepts that have a high potential to be useful for a particular
modelling element. When the model is created, its quality is first evaluated from
different perspectives to make sure that it can be used within the enterprise, and
second to discover whether its feedback can be useful for the ontology. When the
feedback is extracted, the proposed method incorporates guidelines on how to use
this feedback.

Keywords: Business process modeling · Enterprise ontology · Ontology-driven

modelling · BPMN · UFO

1 Introduction

When different models within an enterprise are created by different modelers, integrating
those models is hard. A possible solution for this integration problem is providing
modelers with a shared vocabulary formalized in an ontology [1] and [2]. Over the last
30 years, different domain ontologies have been developed which describe the concepts,
relations between concepts and axioms of a specific domain. In a business context, a
particular type of domain ontologies are so-called enterprise ontologies. They describe
the enterprise domain and consequently provide enterprise domain concepts that can be
reused by different enterprises. Example of enterprise ontologies include the Enterprise
Ontology [3], TOVE [4] and the Resource Event Agent enterprise ontology [5]. Two
different approaches have been proposed to incorporate enterprise ontologies into the
modeling process. Some authors consider enterprise ontologies to be reference models

© Springer International Publishing Switzerland 2015

J. Cordeiro et al. (Eds.): ICEIS 2014, LNBIP 227, pp. 472–488, 2015.
DOI: 10.1007/978-3-319-22348-3_26
 Enterprise-Specific Ontology-Driven Process Modelling 473

that support the creation of different kind of models. For instance, [6] suggests devel‐
oping the Generic Enterprise Model as an ontology that is later used as a reference for
creating both data and process models. Other authors developed an enterprise-specific
modelling language which is based on the concepts, relations and axioms described in
the enterprise ontology [7].
In this paper we focus on using enterprise-specific ontologies (ESO) during the
development of business process models. Enterprise-specific ontologies are domain
ontologies that differ from enterprise ontologies in the fact that their Universe of
Discourse is a specific enterprise, rather than the enterprise domain. They may have their
origin in an established domain ontology or in an enterprise ontology, but their main
goal is describing the concepts, relations and axioms that are shared within a particular
enterprise. Enterprise-specific ontologies are getting increasingly important in the
context of data governance and knowledge representation [1]. Supporting tools, such as
IBM InfoSphere1or Collibra Enterprise Glossary2 allow enterprises to specify their own
enterprise glossary/ontology. Such an enterprise-specific ontology, once available, can
subsequently be deployed to help enterprise modellers in creating compatible, enter‐
prise-specific models, such as requirements, data or process models. This paper focuses
on business process models. Additionally the ESO ontology needs to be maintained and
enriched while the enterprise evolves. Enriching ESO from the process models is very
practical because it will reflect processes that were introduced recently within the enter‐
prise, or processes that where adjusted.
The work described in this paper is a process model-specific instantiation of a frame‐
work for ontology-driven enterprise modeling aimed to facilitate model construction
based on an enterprise-specific ontology on one hand, and support enterprise-specific
ontology creation and evolution based on feedback from the modelling process on the
other hand. This framework also proposes criteria to evaluate the quality of resulting
models to ensure that their feedback is potentially useful. To illustrate our work, we use
the Unified Foundational Ontology as core ontology, OWL as ontology representation
language and BPMN as business process modelling language.

2 Enterprise-Specific Ontology Engineering

and Process Modelling Method

The method that is proposed in this paper in an instantiation of the meta-method which
can be used for different enterprise modeling languages and which was proposed in
previous work [8]. As displayed in Fig. 1, the ontology and modeling method consists
of two parallel cycles, which in turn consist of different phases. For a general description
of the different phases we refer to [8]. In this paper we will focus on describing the
different phases specifically for process modeling using BPMN.

1
http://www-01.ibm.com/software/data/infosphere/.
2
http://www.collibra.com/.
474 N. Alkhaldi et al.

Fig. 1. Enterprise specific ontology and modeling method.

2.1 Ontology Setup Phase

This is the first phase of the Ontology Engineering cycle. It includes pre-processing of
the selected enterprise-specific ontology (ESO) so that matching mechanisms can be
applied in the later phases to derive useful suggestions. Pre-processing in this case
implies mapping concepts from the selected ESO to a core ontology. A core ontology
is an ontology that describes universally agreed upon, high level concepts and relations,
such as objects, events, agents, etc. [9]. Constructs of the selected modelling language
will also be mapped to this core ontology in the next phase of the proposed method. The
core ontology thus forms a bridge between the ESO and the selected modeling language,
and incorporates shared semantics.
The Unified Foundational Ontology (UFO) was selected as a core ontology in this
work for three reasons: 1/ the benefits of grounding domain ontologies in UFO are well
motivated [10], and several such UFO-grounded domain ontologies are available, e.g.,
[11] 2/ UFO is specifically developed for the ontological analysis of modelling
languages, and 3/ BPMN was chosen as modelling language in this work and analysis
of BPMN using UFO is available in literature [12].
 Enterprise-Specific Ontology-Driven Process Modelling 475

UFO has different layers of which here only UFO-U is used, as this is sufficient for
finding construct-based matches between the selected modeling language (explained
later) and ESO. However, it can be further investigated if using (full) UFO is beneficial
to refine the proposed algorithms. The top level element in UFO-U is a Universal. It
represents a classifier that classifies at any moment of time a set of real world individuals
and can be of four kinds: Event type, Quality universal, Relator universal and Object
type.
For the purpose of this demonstration, we have selected as ESO an existing ontology
in financial domain.3 This ontology contains static concepts related to finance, such as
Branch, Customer, Loan, Insurance, etc., which can be used as a reference for models
constructed in the financial domain. A sample of the mapping between the ESO and
UFO-U is presented in Table 1.

Table 1. Mappings between ESO and UFO-U.

ESO UFO-U ESO UFO-U

concept concept
AddedValue Quality_ Liability Relator Universal Mediates
Universal Customer and mediates
Branch
Adminstrative Role_Type Loan Relator Universal Mediates
Customer and mediates
Branch
Asset Mixin Type Mortgage‐ Relator Universal Mediates
Loan Customer and mediates
Branch
Branch Base_Type Castomer Mixin Type

2.2 Ontological Analyses of the Modeling Language

The selected modeling language also needs to be analyzed because the core ontology
forms a intermediary through which modeling constructs are mapped to the concepts of
ESO. Although our meta-method support any kind of core ontology, it is important that
the used modelling languages are analysed using the selected core ontology because the
suggestion generation process relies hereupon. Table 2 represents mappings between
BPMN and UFO-U.

3
http://dip.semanticweb.org/documents/D10.2eBankingCaseStudyDesignandSpecificationo‐
fApplicationfinal.pdf.
476 N. Alkhaldi et al.

Table 2. Mappings between BPMN and UFO-U.

BPMN construct UFO-U BPMN construct UFO-U

Pool Object type End event Event Type
Lane Object type Event noun Base type, Mixin
type, Relator
universal
Task Noun Relator universals Condition Exclu‐ Quality universal
or quality sive Gateway
universal
Noun Sub Process Relator universals Data object Relator universal,,
or quality Base type
universal
Start event Event Type Message flow label Relator universal
Intermediate event Event Type

2.3 Ontology Storage and Suggestion Generation Phase

Every time the modeler places a modeling element on the canvas, several matching
mechanisms are cooperating in order to rank the ESO concepts to display the most
relevant of them on the top of a suggestions list. Given the potentially extensive amount
of ESO concepts, relevance ranking of suggestions is a critical feature.
Depending on the type of modelling construct that is added, the position of the
construct relevant to other elements (i.e., its neighborhood) in the model, and the label
entered by the modeller, the order of the suggestion list is prioritized so that ontology
concepts with a higher likelihood to be relevant in the current context appear first. To
achieve this, four different suggestion generation mechanisms are used. These mecha‐
nisms are partly inspired by ontology matching techniques [13], but are specifically
focused to fit within our framework, where the semantics of the modelling language can
be exploited.
Every matching technique calculates a relevance score (between 0 and 1) for each
ESO concept, which is stored. Subsequently, the overall relevance score is calculated
using a weighted average of all individual scores. This corresponds to the formula below:

Where:
: the score and weight of string match
: the score and weight of synonym match
: the score and weight of construct match
: the score and weight of neighborhood based match
 Enterprise-Specific Ontology-Driven Process Modelling 477

The weights for each matching mechanism are thus configurable. In our demonstration
(see Sect. 3), we assigned a higher weight to string matching as we expect that, within a
particular enterprise context, a (quasi) exact string match has a high possibility of repre‐
senting the intended (semantic) concept. The lowest weight is assigned to construct
matching, because it typically matches very broadly, and thus delivers a large amount of
suggestions. Further experimentation with the distribution of weights over the individual
relevance scores should be performed to determine an optimal overall score calculation.
This is considered future work. As a final result, the suggestion list, a descending ordered
list of ESO concepts according to relevance, is generated and presented to the modeller. In
the next four subsections the different mechanisms are described in more detail. The
implementation of the mechanisms can be consulted via our Github repository: https://
github.com/fgailly/CMEplusBPMN.

2.3.1 String Matching Mechanism

The goal of the string matching algorithm is to find ESO concepts whose label is syntac‐
tically similar to the label of modeling elements entered by the modeller. If these two
strings are syntactically the same, there is a high possibility that they have the same
semantics, especially as both reside within the same enterprise and business context.
Currently Jaro-Winkler distance [14] is used to calculate the edit distance between
the given BPMN element label, and the label of each concept in the enterprise-specific
ontology. The Jaro-Winkler distance was chosen because this hybrid technique takes
into account that the text entered by the modeller can contain spelling errors, and addi‐
tionally favours matches between strings with longer common prefixes (i.e., a substring
test, which is very useful in our context because matching is executed each time a char‐
acter is added to the label). Jaro-Winkler distances are between 0 (no similarity) and 1
(exact match), and are thus immediately useable as a relevance score.

2.3.2 Synonym Matching Mechanism

The synonym matching mechanism aims to detect synonyms of the given BPMN
element label (or part of it) in the ESO. To realize this, WordNet [15] is used. WordNet
is an online lexical database that organizes English nouns, verbs, adjectives and adverbs
into sets of synonyms (so-called synsets). It is thus ideal to find synonyms. For each
synonym of the modeling element label, the previously described string matching algo‐
rithm is performed on all ESO concepts, thereby generating a relevance score between
0 (no match) and 1 (exact synonym match).

2.3.3 Construct Matching Mechanism

This matching mechanism operates based on the mapping performed in “ontology
set up” and “ontological analyses of the modeling language” phases described previ‐
ously. Consequently, BPMN modeling constructs can be mapped to ESO concepts
through the UFO core ontology. During the matching process, this mechanism
assigns a score of 1 to all ESO concepts mapped to the same UFO-U construct as the
modeling language construct created. All the other ESO concepts are assigned rele‐
vance score of 0. In our implementation, the mappings between the ESO and UFO,
478 N. Alkhaldi et al.

and BPMN and UFO, are each represented in an OWL file. OWL reasoning is then
used to perform the mappings between the ESO concepts and BPMN constructs
based on the two aforementioned OWL files.

2.3.4 Neighborhood – Based Matching Mechanism

Neighborhood-based mechanism calculates relevance scores for ESO concepts based
on the location of the newly added modeling element, the type of modeling element that
is added, and the relationships between the ESO concepts corresponding to the modeling
elements surrounding the newly added element. The neighborhood of a BPMN element
is determined by the connectivity objects (i.e. sequence flow, message flow, association),
and which pool (or lane) the BPMN element is located in. In other words, for every
element we can determine which pool or lane it is a part of, and which other element(s)
is/are connected to this element using either a sequence, message flow or association.
Next, the relationships (which are specified in terms of the UFO-U relationships through
the ESO-UFO-U mappings) between the ESO concepts are exploited. According to [10]
there are two types of relations: formal and material. A formal relation between entities
holds directly, without any further intervening individuals. A material relation has
material structure by itself. It includes relations such as working at, being treated at, etc.
Entities related by this type of relation are mediated by individuals called relators. In
Sect. 3 we will demonstrate how the UFO-U relators can be used to suggest concepts
from the ESO.
Finally, using both the relative position of the new element and the material relations
between the ESO concepts, the element neighborhood-based mechanism can now derive
relevance scores for ESO concepts in relation to some BPMN modeling elements (for
examples, see Sect. 3):
1. To create a pool construct when another pool already exists, the suggestions (rele‐
vance score 1) are UFO-U object types that are related by a material relationship
with the ESO concept with which the existing pool(s) was/were annotated (i.e., the
ontology annotations of the pool(s)).
2. To create a lane construct within a pool, the suggestions (relevance score 1) are UFO-
U role types that are related by a material relationship with the ontology annotation
of the pool(s)
3. To create a message construct that results in transmitting a message between a task
or event of a pool and another pool, the suggestions (relevance score 1) are UFO-U
relators mediating material relations connecting objects that annotate respectively
the noun of the task and the ontology annotation of the pool.
4. To create a conditional gateway, there are two ways to derive suggestions (both
receive relevance score 1):
• ESO concepts annotated by the task label preceding the gateway. This can work very
well for tasks that are performing evaluation or calculation, after which the gateway
is used to make a decision based on the results. In this case, the condition on the
gateway will use the same concept as used in the task. This concept is most likely
to be a quality, especially if the task at hand is performing calculations. Nevertheless,
it can also be a relator, such as for example verifying if the contract is ok or not.
 Enterprise-Specific Ontology-Driven Process Modelling 479

• Qualities associated with the UFO-U object type annotation of the pool where the
gateway is located. Or UFO-U qualities associated with UFO-U object types partic‐
ipating in material relations with Object type annotation of the pool where the
gateway is located.
5. For creation of a task construct, the suggested concepts are most likely to be related
through material relations to the pool where the task is located. The suggestions can
be either quality types of the concept annotating that pool or relators mediating those
material relations.

2.4 Enterprise Model Creation Phase

During this phase the modeler proceeds with creating the process model utilizing
suggestions derived by the matching mechanisms described in the previous section. With
every modeling element placed on the canvas the modeler is advised to select ESO
concept from the suggestions list to annotate the modeling element. This annotation
implies maintaining a link between the modeling element and the corresponding concept
of the ESO. With the help of the annotation, the modeling element is semantically
connected to the ESO concept even if they have different labels. In our implementation,
the annotation is realized by creating an OWL file where a URI of the OWL ESO concept
is added to the corresponding modeling element using the OWL annotation mechanism.
A portion of the annotation OWL file is presented below. It shows that pool construct
“Participant_1” has a label “Customer” and is annotated by “Customer” concept of the
bank ontology (our ESO).

<ClassAssertion>
<Class IRI = ”http://www.mis.ugent.be/ontologies/bpmn.owl#Pool“/>
<NamedIndividual IRI = ”#Participant_1”/>
</ClassAssertion>
<AnnotationAssertion>
<AnnotationProperty abbreviatedIRI = ”rdfs:isDefinedBy”/>
<IRI > #Participant_1 </IRI>
<Literal datatypeIRI = ”http://www.w3.org/2001/XMLSchema#string“>
http://www.mis.ugent.be/ontologies/bank#Customer
</Literal>
</AnnotationAssertion>

2.5 Enterprise Model Evaluation Phase

Within the Enterprise-Specific Ontology and Modeling framework, the model quality
evaluation phase has two main goals: (1) it ensures that the model can be used within
the enterprise and (2) that the feedback extracted from the creation of the model is useful
and potentially worth incorporating into the ESO. Literature provides a plethora of
frameworks for quality evaluation of different kinds of models. For an overview of how
process model quality can be evaluated we refer to [16]. A well-known scheme for
480 N. Alkhaldi et al.

classifying quality dimensions is the Lindland et al. framework [17] which makes a
distinction between syntactic, semantic and pragmatic quality dimensions. Syntactic
quality implies correspondence between the model and the modelling language.
Semantic quality measures how compliant the model is to the domain. And pragmatic
quality is the correspondence between the model and the user interpretation of it.
The first goal of the model evaluation phase can be satisfied by focusing on measures
that fall within the syntactic and pragmatic dimension because for the model to be used
within the enterprise it needs to be correct and understandable. The second goal can be
achieved by focusing on measures that fall within the semantic quality dimension. More
specific the developed process model has to reflect the process it was designed for.
Hence, the three quality dimensions postulated by [17] are well suited for this phase of
the meta-method.

2.5.1 Syntactic Quality

This dimension stipulates that the created model must be syntactically correct in order
to be used within an enterprise. If the model has syntactic flows, its process cannot be
correctly implemented. Syntactic quality is achieved using syntactic correctness crite‐
rion. According to [18] verification of the syntactic quality of a process models focuses
on two properties: static property and behavioral property. Static property is the related
to the elements of the model and how they are used and connected. For example, in
BPMN it is not allowed to have sequence flow between two pools. The static syntactic
quality can be verified by the modeling tool itself while the model is being created.
Behavioral property relates to the behavior of the process model. For example, the
process cannot reach a deadlock and proper process completion is guaranteed. This is
evaluated automatically by computer programs after the model is created.

2.5.2 Semantic Quality

Semantic quality of a process model is typically evaluated by means of completeness
and validity measures. Completeness is defined in [16] as a degree to which a model has
all the necessary and relevant information. Following the work of [19] completeness of
a process model can be measured by: 1. Counting the number of items in BPMN model
that do not correspond to the description of the actual process 2. counting the number
of requirements that are present in model description, but are not reflected in the model
itself. It is advisable to document the process to be constructed so that this document
can serve as a reference for model evaluation. But when no description is available, the
quality evaluation metrics can be executed by another stakeholder familiar with the
process, or by the modeler himself. According to [18] the model is valid when all its
statements are correct and relevant to the problem. In order to verify validity, one must
know the meaning of modeling elements and the process that the model is representing.

2.5.3 Pragmatic Quality

Pragmatic quality is about the correspondence between the process model and users’
comprehension of it. This quality dimension is not relevant when the modeler himself
is the only user of the model because he obviously will understand the model he made.
 Enterprise-Specific Ontology-Driven Process Modelling 481

But if the model will be presented to other stakeholders, it is very important that they
understand it.
The literature contains several propositions on how to evaluate understandability.
[16] proposes using structural complexity metrics suggested by [20] (which is specific
to BPMN) as an indication of the degree of model understandability. If the model is
complex, it is likely to be less understandable. In the context of the proposed method it
is possible to use a complexity metric to make sure that the model is not exceeding
complexity limits.

2.6 Community-Based Ontology Feedback Evaluation Phase

The method only proceeds to this phase if the model satisfies the expected semantic
quality, meaning that it correctly reflects the process it was designed for, and all the
statements in the model are valid within the process. When the process model is
complete, an OWL file representing this model and the ESO concepts selected for the
model annotation is stored. This file is processes in order to extract any possible feedback
which is potentially useful and can be incorporated into the ESO. A possible feedback
is a listing of the elements from the model that were not annotated by ESO concepts. As
they were not annotated, the reason might be that there is no equivalent for them in the
ESO. This feedback is made available for other community members and is subject to
discussion, until finally a consensus is reached whether or not the proposed change(s)
should be included (such as new concepts added) in the new version of the ontology.
The community will discuss the proposed concepts such as their usage, definition and
usefulness within the enterprise. Community members are other people working in the
same business domain. As community members are typically not co-located at the same
physical location, and we are aiming to progressively reach a consensus about what is
needed by the community, the Delphi approach [21] is used. This approach is perfectly
suited to capture collective knowledge and experience of experts in a given field, inde‐
pendently of their location, and to reach a final conclusion by consensus. More specif‐
ically, consensus is reached by commenting on the feedback in 3 cycles. Three cycles
were chosen because studies show that most changes in responses occur in the first 2
rounds [21]. In every cycle comments are assigned a score, and when all three cycles
are accomplished, a decision is taken whether to incorporate feedback or discard it. Only
in case the community is not able to reach a consensus, the final decision is made by
community members with a high level of trust. A system for assigning trust credits to
community members is foreseen.
If negotiation upon the feedback progresses slowly, the process may be terminated
without accomplishing predefined number of cycles. In this case highly trusted, author‐
ized community members are responsible to make a decision.

2.7 Ontology Update and Evolution Phase

After the feedback verification is performed, ontology expert incorporate its results into
the enterprise ontology. It is worth mentioning that ontology experts do not interfere in
feedback verification. Their mission is limited solely to incorporating the final results
482 N. Alkhaldi et al.

into the ontology in a syntactically correct way. Once a considerable amount of feedback
is incorporated, a new ontology version is proposed. The new ontology incorporates
new concept/relationships, updates lacking/incomplete ones, and/or removes irrelevant
once, as the domain evolves or new insights are reached by the expanding community.

3 Demonstration

This section illustrates the suggestions generation phase of the method explained in the
previous sections by means of a lab demonstration in which the modeller constructs a
process model in BPMN notation in the financial domain using an existing financial
domain ontology4 as enterprise-specific ontology.
Once the ESO is grounded in the core ontology (as mentioned, this only needs to be
done once, and can subsequently be re-used for any model created within the enterprise),
the modeller can start creating the process model. He selects a construct to be added,
places it on the canvas and starts typing the construct’s desired name. As he selects the
construct, and as he is typing, the mechanisms described in the previous section derive
suggestions from the ESO and present them to the modeller. If an ESO concept in the
suggestion list appropriately corresponds to the intention of the modeller for this partic‐
ular BPMN construct, he selects this concept, and the BPMN construct is (automatically)
annotated with the chosen ESO concept.
The process model to be created in our lab demonstration represents the loan appli‐
cation assessment process in a bank, and is taken from [22]. By using an existing spec‐
ification, we avoid bias towards our method. The process starts when the loan officer
receives a loan application from one of the bank’s customers. This loan application is
approved if it passes two checks: the first check is the applicant’s loan risk assessment,
which is done automatically by the system after a credit history check of the customer
is performed by a financial officer. The second check is a property appraisal check
performed by the property appraiser. After both checks are completed, the loan officer
assesses the customer’s eligibility. If the customer is found to be not eligible, the appli‐
cation is rejected. Otherwise, the loan officer starts preparing the acceptance pack. He
also checks whether the applicant requested a home insurance quote. If he did, both the
acceptance pack and the home insurance quote are sent to the applicant. If the insurance
was not requested, only the acceptance pack is sent. The process finally continues with
the verification of the repayment agreement.
Figure 2 represents the BPMN model of the loan application process. Constructs that
are surrounded by a thick red square are annotated with ESO concepts.

4
http://dip.semanticweb.org/documents/D10.2eBankingCaseStudyDesignandSpecificationo‐
fApplicationfinal.pdf.
 Enterprise-Specific Ontology-Driven Process Modelling 483

Fig. 2. BPMN model describing loan application.

Adding Branch Pool: The modeller selects the pool construct to be created. Based on
the construct matching mechanism all ESO concepts corresponding to UFO-U object
type are given a relevance score of 1 for this matching mechanism. Among those
concepts the modeller can find Branch which is classified as UFO-U base type. If there
is already a “Customer” pool, based on the construct neighbourhood matching mecha‐
nism, this case corresponds with rule 1. As the already existing pool is the Customer
pool, the mechanism looks for ESO concepts related to the Customer concept through
material relationships. There is only one concept satisfying this requirement: Branch.
As a result, the Branch concept is listed in the beginning of the suggestion list, as it
scored for both the construct and neighbourhood matching mechanisms (and no other
concept scored equal or higher). Note that in this scenario, string and synonym matching
cannot contribute to the overall relevance score yet, as the modeller did not (yet) type
any label.

Adding Message Flow “Loan Application”: According to the construct matching

mechanism, message flow corresponds to the relator universal. Therefore, all ESO
concepts corresponding to the UFO-U relator universal will be selected. Those concepts
are: Channel, loan, mortgage loan, current mortgage loan, future mortgage loan, invoice,
484 N. Alkhaldi et al.

liability, payment. For the element neighbourhood- based matching technique this situa‐
tion resolves under rule 2.
In our enterprise-specific ontology, all relator universals are mediating the same two
concepts Branch and Customer. Therefore, the results delivered by this suggestion
generation technique are the same as the results delivered by construct matching. In this
case, the previously mentioned suggestions all have equal overall score, and can thus
not be prioritized. We therefore present them alphabetically. The modeller may select
a concept from the list (i.e., “loan”), or, in case the list is too long, start typing any desired
label (e.g., “loan” or “credit”). This triggers the string- and synonym-based matching
mechanisms, both of which prioritize the concept Loan, which consequently appears on
top of the suggestion list, and is selected by the modeller to annotate the loan application
message flow.

Adding Reject Application Task: The modeller selects the BPMN task construct, and
subsequently the construct matching mechanism assigns a high relevance score to all
ESO concepts that correspond to UFO-U quality and relator universals as suggestions
for the task noun. A list of relator universals is mentioned in the previous example;
a list of quality universals is very exhaustive and is thus not mentioned here. The second
mechanism, element neighbourhood based matching, applies rule 3: the task at hand is
located in the Branch pool, so this matching mechanism suggests all the ESO concepts
corresponding to UFO-U relator universals related to Branch concept in the ESO, and
UFO-U object types mediated by those relators (all with relevance score 1). The Loan
concept is a relator universal, and therefore received relevance score of both matching
mechanisms; it therefore appears on the top of the suggestions list.

Adding “Home Insurance Quote is Requested” Gateway: In the last scenario, the
modeller draws an inclusive decision gateway on the canvas. Based on construct
matching mechanism, all quality universals will be assigned a priority score of 1. The
element neighbourhood-based mechanism classifies this situation under rule 4, which
suggests ESO concepts that were used to annotate a task construct preceding the
gateway. In this case it is the “check if home insurance quote is requested” task, which
is annotated with the Home Insurance concept. This concept thus receives relevance
score 1 for the gateway, and is prioritized in the suggestion list. It perfectly matches our
needs.
To start the discussion, we note that the scenarios elaborated here were chosen to
illustrate the more complicated cases. As a result, string- and synonym-based matching
mechanism are underrepresented. Evidently, when no or few BPMN elements are
already on the canvas, neighbourhood-based matching will be unable to sufficiently
differentiate between potential suggestions (as in scenario 2), and string- and synonym-
matching will become important. Equally, when the modeller has a certain label already
in mind, string and synonym matching will dominate the suggestion list, as the modeller
is typing the label he had in mind. Having made this comment, we note that in general,
it was possible to derive suggestions based on the construction and element neighbour‐
hood matching mechanisms for all model constructs for which the related concepts
existed in the ontology. In fact, as can be seen in the scenarios, construct and
 Enterprise-Specific Ontology-Driven Process Modelling 485

neighbourhood based matching complement each other well. The majority of the
concepts required by the model, but missing from the ontology were also correctly
classified under the assumptions of neighbourhood-based matching mechanism, and
would have been assigned a high relevance score if they would have been present in the
ontology. However, there was one case where the neighbourhood-based matching
mechanism was not very accurate. While creating the last message flow “Home insur‐
ance quote”, based on the second assumption of the neighbourhood-based matching
mechanism and the construct matching mechanism, relator universals must be
suggested. But in reality, it was annotated with a quality universal HomeInsurance,
instead of a relator. Further fine-tuning of the suggestion generation mechanism should
avoid this type of mismatches.
The lab demonstration was used here to demonstrate viability of our method, to detail
the different steps and provide a concrete case. It shows that the suggestion generation
algorithm indeed provides useful suggestions to the modeller, and allows (automatic)
annotations of the model, thereby semantically grounding them and facilitating model
integration. It needs to be mentioned that the lab demonstration was done using a single
modeller, and that therefore the aforementioned positive indications of using our method
cannot be statistically proven. We are currently performing a more elaborate empirical
validation, where a group of test users is divided in three different groups: one group is
given an ESO and our method, the second group is only given the ESO but without
support of our method, and the last group is not given an ESO and thus needs to model
without any ontology or method support. The experiment is specifically designed to
show the impact of our model on modelling efficiency, consistency in the use of termi‐
nology, and the semantic grounding of the resulting models.
After the model is created, it is time to perform model quality evaluation. Syntactic
quality is measured by counting the amount of violations of BPMN syntax. From the
static perspective, the model in Fig. 2 is syntactically correct which is expected as the
modeling tool itself prevents some basic violations. From the behavioral aspect it is
important to look into different scenarios of model completion. This model will always
reach a valid completion independently of which paths are selected at the gateway. There
are no tasks in the model that can never be executed. To evaluate semantic quality, we
need to look back to the model description in the beginning of this demonstration section.
Because within the context of this method, this description represents the domain to
which the model needs to correspond. There are two requirements in the description that
are not reflected in the process model in Fig. 2. The first requirement is: “applicant’s
loan risk assessment, which is done automatically by the system”, and the second
requirement is: “credit history check of the customer is performed by a financial officer”.
Those requirements need to be incorporated before the feedback from the model is taken.
To evaluate the pragmatic quality, structural complexity is calculated based on
measures presented in [16] presented in Table 3. According to [23] those measures are
directly related to understandability. The thresholds suggested for those values are:
number of nodes between 30 and 32, Gateway mismatch is between 0 and 2, depth is 1,
connectivity coefficient is 0.4. With those values the model is considered to be 70%
understandable. The values obtained in evaluating the model in Fig. 2 are close to the
proposed threshold, and therefore the model is potentially understandable.
486 N. Alkhaldi et al.

Table 3. Structural complexity measurements of the model in Fig. 2.

Metric Result
Number of nodes in the model 15
Gateway mismatch (sum of gateway pairs that do not match 1
with each other)
Depth (maximum nesting of structure blocks) 0
Connectivity coefficient (Ratio of the total number of arcs 13/15 = 0.9
in a process model to its total number of nodes)

4 Conclusions and Future Work

This paper presented an overview of an enterprise-specific ontology-driven process

modeling method. On the one hand, this method improves semantic consistency of
process models by relying in ESO in model construction. On the other hand, it supports
the evolvement of the ESO according to practical needs of the enterprise by taking
feedback from the process models. While constructing models, the modeler is aided by
a list of suggestions extracted from the ESO. ESO concepts are ranked in a suggestions
list using four matching mechanisms. Two of them, the string and synonym matching
mechanisms, are based on the label of the newly created BPMN element, which is
systematically compared with concepts in the ESO. The other two, namely construct
matching and neighbourhood-based matching, depend on the type of the BPMN
construct and the position (relative to other modelling elements) where it is added.
Another benefit of the proposed method is that it facilitates maintenance and
improvement of the ESO by means of feedback from the process models. This feedback
is only accepted if the resulted model properly reflects the process it is representing. The
proposed method incorporates guidelines on model quality evaluation. When a model
is created, an OWL file containing all modeling elements is stored. This file is processed
by ontology expert and all the elements representing potential feedback. This feedback
is subject to community discussion and if approved, it will be incorporated in the new
ontology version.
Future work will follow different directions. Concerning model creation based on
the ESO, first, we are currently performing further empirical validation of the benefits
of our method. Second, suggestions towards the ontology (e.g., missing concepts) based
on the modelling process, and subsequent community-based ontology evolution, needs
to be explored. Finally, we also plan to apply the method for other modelling languages
(i.e. i*, KAOS), and using other core ontologies. Concerning ESO maintenance and
amelioration, we are planning to set up a forum where the community will discuss model
feedback. Clear definition of guidelines for discussion and voting is also very essential.
 Enterprise-Specific Ontology-Driven Process Modelling 487

References

1. Bera, P., Burton-Jones, A., Wand, Y.: Guidelines for designing visual ontologies to support
knowledge identification. MIS Q. 35(4), 883–908 (2011)
2. Di Francescomarino, C., Tonella, P.: Supporting ontology-based semantic annotation of
business processes with automated suggestions. In: Halpin, T., Krogstie, J., Nurcan, S.,
Proper, E., Schmidt, R., Soffer, P., Ukor, R. (eds.) Enterprise, Business-Process and
Information Systems Modeling. LNBIP, vol. 29, pp. 211–223. Springer, Heidelberg (2009)
3. Uschold, M., King, M., Moralee, S., Zorgios, Y.: The enterprise ontology. Knowl. Eng. Rev.
Spec. Issue Putting Ontol. Use 13(1), 31–89 (1998)
4. Fox, M.S.: The TOVE project: towards a common-sense model of the enterprise. In: Belli,
F., Radermacher, F.J. (eds.) Industrial and Engineering Applications of Artificial Intelligence
and Expert Systems, pp. 25–34. Springer, Germany (1992)
5. Geerts, G.L., McCarthy, W.E.: An accounting object infrastructure for knowledge based
enterprise models. IEEE Intell. Syst. Their Appl. 14, 89–94 (1999)
6. Fox, M., Gruninger, M.: On ontologies and enterprise modelling. In: Kosanke, K., Nell, J.G.
(eds.) Enterprise Engineering and Integration, pp. 190–200. Springer, Heidelberg (1997)
7. Sonnenberg, C., Huemer, C., Hofreiter, B., Mayrhofer, D., Braccini, A.: The REA-DSL: A
domain specific modeling language for business models. In: Mouratidis, H., Rolland, C. (eds.)
CAiSE 2011. LNCS, vol. 6741, pp. 252–266. Springer, Heidelberg (2011)
8. Gailly, F., Casteleyn, S., Alkhaldi, N.: On the symbiosis between enterprise modelling and
ontology engineering. In: Ng, W., Storey, V.C., Trujillo, J.C. (eds.) ER 2013. LNCS, vol.
8217, pp. 487–494. Springer, Heidelberg (2013)
9. Doerr, M., Hunter, J., Lagoze, C.: Towards a core ontology for information integration. J.
Digit. Inf. 4(1) (2006)
10. Guizzardi, G., Wagner, G.: What’s in a relationship: an ontological analysis. In: Li, Q.,
Spaccapietra, S., Yu, E., Olivé, A. (eds.) ER 2008. LNCS, vol. 5231, pp. 83–97. Springer,
Heidelberg (2008)
11. Barcellos, M.P., Falbo, R., Dal Moro, R.: A well-founded software measurement ontology.
In: 6th International Conference on Formal Ontology in Information Systems (FOIS 2010),
vol. 209, pp. 213–226 (2010)
12. Guizzardi, G., Wagner, G.: Can BPMN be used for making simulation models? In: Barjis, J.,
Eldabi, T., Gupta, A. (eds.) EOMAS 2011. LNBIP, vol. 88, pp. 100–115. Springer, Heidelberg
(2011)
13. Euzenat, J., Shvaiko, P.: Ontology Matching. Springer, Berlin (2013)
14. Winkler, W.: String comparator metrics and enhanced decision rules in the Fellegi-Sunter
model of record linkage. In: Proceedings of the Section on Survey Research Methods,
American Statistical Association (1990)
15. Miller, G.: WordNet: a lexical database for English. Commun. ACM 38(11), 39–41 (1995)
16. Sánchez-González, L., García, F., Ruiz, F., Piattini, M.: Toward a quality framework for
business process models. Int. J. Coop. Inf. Syst. 22(01), 1350003 (2013)
17. Lindland, O., Sindre, G., Sølvberg, A.: Understanding quality in conceptual modeling. IEEE
Softw. 11(2), 42–49 (1994)
18. Reijers, H.A., Mendling, J., Recker, J.: Business Process Quality Management. In: vom
Brocke, J., Rosemann, M. (eds.) Handbook on Business Process Management 1, pp. 167–
185. Springer, Berlin (2010)
19. Moody, D.L.: Metrics for evaluating the quality of entity relationship models. In: Ling,
T.-W., Ram, S., Li Lee, M. (eds.) ER 1998. LNCS, vol. 1507, pp. 211–225. Springer,
Heidelberg (1998)
488 N. Alkhaldi et al.

20. Rolón, E., Ruiz, F., Garcia, F., Piattini, M.: Applying software metrics to evaluate business
process models. CLEI-Electron. J. 9(1) (2006)
21. Gupta, U.G., Clarke, R.E.: Theory and applications of the delphi technique: a bibliography
(1975–1994). Technol. Forecast. Soc. Chang. 53, 185–211 (1996)
22. Dumas, M., La Rosa, M., Mendling, J., Reijers, H.: Fundamentals of Business Process
Management. Springer, Heidelberg (2013)
23. Sánchez-González, L., Garcia, F., Mendling, J., Ruiz, F.: Quality assessment of business
process models based on thresholds. In: Meersman, R., Dillon, T.S., Herrero, P. (eds.) OTM
2010. LNCS, vol. 6426, pp. 78–95. Springer, Heidelberg (2010)
 Author Index

Alkhaldi, Nadejda 472 Haddad, Matheus N. 131

Augustin, Iara 320 Hoos, Eva 222
Huitema, G.B. 453
Baranauskas, Maria Cecília Calani 341, 397
Bernardo, Ivelize Rocha 397 Junior, Omir Correia Alves 285
Borges, Michela 397 Júnior, Paulo Afonso Parreira 244
Bouzeghoub, Amel 320
Buchdid, Samuel B. 341 Kähkönen, Tommi 39
Kramer, Stefan 222
Caetano, Nuno 149
Casanova, Marco A. 302 Laureano, Raul M.S. 149
Casteleyn, Sven 472 Lenz, Hans-J. 14
Chino, Daniel Y.T. 91 Lichtnow, Daniel 320
Coelho da Silva, Ticiana L. 75 Liu, Kecheng 3
Coelho, Pedro Henrique Gouvêa 167 Lopes, Giseli Rabello 302
Cortez, Paulo 149
Cota, Luciano P. 131 Machado, Alencar 320
Machado, Javam C. 75
D’Souza, A. 453 Maculan, Nelson 131
da Silva, Alberto Rodrigues 265 Magalhães, Regis Pires 75
da Silva, Camilla Almeida 379 Maglyas, Andrey 39
de Arruda Barreira, Luciane Fernanda 167 Maia, Marcelo A. 187
de Barros, Adriano Valladão 167 Mitschang, Bernhard 222
de Farias, Victor A.E. 75
de Macêdo, José A.F. 75 Naranjo, David 434
de Oliveira, José Palazzo Moreira 320 Neto, Antonio C. Araújo 75
del Alamo, Jose M. 417
Ding, Sun 204 Paes Leme, Luiz André P. 302
do Amaral, Jorge Luís Machado 167 Penteado, Rosângela Dellosso 244
do Amaral, José Franco Machado 167 Pequeno, Valéria 57
do Vale A. Gomes, Raphael 302 Pereira, Roberto 341
Pernas, Ana Marilza 320
Fernandes, António Ramires 379 Pontieri, Luigi 108
Fiorese, Adriano 285 Preda, Marius 361
Folino, Francesco 108
Rocher, Pierre-Olivier 361
Gailly, Frederik 472 Romani, Luciana A.S. 91
Galhardas, Helena 57
Goncalves, Renata R.V. 91 San Miguel, Beatriz 417
Gravier, Christophe 361 Sánchez, Mario 434
Gröger, Christoph 222 Santanchè, André 397
Grohmann, Ana Paula 379 Silva, Rodrigo F.G. 187
Guarascio, Massimo 108 Smolander, Kari 39
490 Author Index

Soares, Michel S. 187 Vidal, Vânia M. Ponte 57

Souza, Marcone J.F. 131 Vieira, Rafael Giordano 285
Subercaze, Julien 361 Villalobos, Jorge 434

Tan, Chekfoung 3 Wives, Leandro Krug 320

Tan, Hee Beng Kuan 204 Wortmann, J.C. 453
Traina, Agma J.M. 91
Traina Jr., Caetano 91
Yelmo, Juan C. 417
van Beest, N.R.T.P. 453
Velthuijsen, H. 453 Zhang, Hongyu 204