Algorithms for Intelligent Systems

Series Editors: Jagdish Chand Bansal · Kusum Deep · Atulya K. Nagar

Ibrahim Aljarah
Hossam Faris
Seyedali Mirjalili Editors

Evolutionary
Data Clustering:
Algorithms and
Applications
Algorithms for Intelligent Systems

Series Editors
Jagdish Chand Bansal, Department of Mathematics, South Asian University,
New Delhi, Delhi, India
Kusum Deep, Department of Mathematics, Indian Institute of Technology Roorkee,
Roorkee, Uttarakhand, India
Atulya K. Nagar, School of Mathematics, Computer Science and Engineering,
Liverpool Hope University, Liverpool, UK
This book series publishes research on the analysis and development of algorithms
for intelligent systems with their applications to various real world problems. It
covers research related to autonomous agents, multi-agent systems, behavioral
modeling, reinforcement learning, game theory, mechanism design, machine
learning, meta-heuristic search, optimization, planning and scheduling, artificial
neural networks, evolutionary computation, swarm intelligence and other algo-
rithms for intelligent systems.
The book series includes recent advancements, modification and applications
of the artificial neural networks, evolutionary computation, swarm intelligence,
artificial immune systems, fuzzy system, autonomous and multi agent systems,
machine learning and other intelligent systems related areas. The material will be
beneficial for the graduate students, post-graduate students as well as the
researchers who want a broader view of advances in algorithms for intelligent
systems. The contents will also be useful to the researchers from other fields who
have no knowledge of the power of intelligent systems, e.g. the researchers in the
field of bioinformatics, biochemists, mechanical and chemical engineers,
economists, musicians and medical practitioners.
The series publishes monographs, edited volumes, advanced textbooks and
selected proceedings.

More information about this series at http://www.springer.com/series/16171

Ibrahim Aljarah Hossam Faris
• •

Seyedali Mirjalili
Editors

Evolutionary Data
Clustering: Algorithms
and Applications

123
Editors
Ibrahim Aljarah Hossam Faris
King Abdullah II School King Abdullah II School
for Information Technology for Information Technology
The University of Jordan The University of Jordan
Amman, Jordan Amman, Jordan

Seyedali Mirjalili
Center for Artificial Intelligence
Research and Optimization
Torrens University Australia
Brisbane, QLD, Australia
Griffith University
Brisbane, QLD, Australia

ISSN 2524-7565 ISSN 2524-7573 (electronic)

Algorithms for Intelligent Systems
ISBN 978-981-33-4190-6 ISBN 978-981-33-4191-3 (eBook)
https://doi.org/10.1007/978-981-33-4191-3

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature
Singapore Pte Ltd. 2021
This work is subject to copyright. All rights are solely and exclusively licensed 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, expressed or implied, with respect to the material contained
herein or for any errors or omissions that may have been made. The publisher remains neutral with regard
to jurisdictional claims in published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Singapore Pte Ltd.
The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721,
Singapore
Preface

This book provides an in-depth analysis of the current evolutionary clustering

techniques, discussing the most highly regarded methods for data clustering. The
book provides literature reviews about single objective and multi-objective evolu-
tionary clustering algorithms. In addition, the book provides a comprehensive
review of the fitness functions and evaluation measures that are used in most of the
evolutionary clustering algorithms. Furthermore, the book provides a conceptual
analysis including definition, validation and quality measures, applications, and
implementations for data clustering using classical and modern nature-inspired
techniques. It features a range of proven and recent nature-inspired algorithms used
for data clustering, including particle swarm optimization, ant colony optimization,
grey wolf optimizer, salp swarm algorithm, multi-verse optimizer, Harris hawks
optimization, and beta-hill climbing optimization.
The book also covers applications of evolutionary data clustering in diverse
fields such as image segmentation, security, medical applications, EEG-based
person identification, and pavement infrastructure asset management.

Amman, Jordan Dr. Ibrahim Aljarah

Amman, Jordan Prof. Hossam Faris
Brisbane, Australia Dr. Seyedali Mirjalili
August 2020

v
Contents

Introduction to Evolutionary Data Clustering and Its Applications . . . . 1

Ibrahim Aljarah, Maria Habib, Hossam Faris, and Seyedali Mirjalili
A Comprehensive Review of Evaluation and Fitness Measures
for Evolutionary Data Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
Ibrahim Aljarah, Maria Habib, Razan Nujoom, Hossam Faris,
and Seyedali Mirjalili
A Grey Wolf-Based Clustering Algorithm for Medical
Diagnosis Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
Raneem Qaddoura, Ibrahim Aljarah, Hossam Faris,
and Seyedali Mirjalili
EEG-Based Person Identification Using Multi-Verse Optimizer
as Unsupervised Clustering Techniques . . . . . . . . . . . . . . . . . . . . . . . . . 89
Zaid Abdi Alkareem Alyasseri, Ammar Kamal Abasi,
Mohammed Azmi Al-Betar, Sharif Naser Makhadmeh, João P. Papa,
Salwani Abdullah, and Ahamad Tajudin Khader
Capacitated Vehicle Routing Problem—A New Clustering Approach
Based on Hybridization of Adaptive Particle Swarm Optimization
and Grey Wolf Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
Dao Vu Truong Son and Pham Nhat Tan
A Hybrid Salp Swarm Algorithm with b-Hill Climbing Algorithm
for Text Documents Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
Ammar Kamal Abasi, Ahamad Tajudin Khader,
Mohammed Azmi Al-Betar, Zaid Abdi Alkareem Alyasseri,
Sharif Naser Makhadmeh, Mohamad Al-laham, and Syibrah Naim
Controlling Population Diversity of Harris Hawks Optimization
Algorithm Using Self-adaptive Clustering Approach . . . . . . . . . . . . . . . 163
Hamza Turabieh and Majdi Mafarja

vii
viii Contents

A Review of Multiobjective Evolutionary Algorithms for Data

Clustering Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
Ruba Abu Khurma and Ibrahim Aljarah
A Review of Evolutionary Data Clustering Algorithms
for Image Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
Laila Al-Qaisi, Mohammad A. Hassonah, Mahmoud M. Al-Zoubi,
and Ala’ M. Al-Zoubi
Pavement Infrastructure Asset Management Using Clustering-Based
Ant Colony Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
Saqib Gulzar and Hasnain Ali
A Classification Approach Based on Evolutionary Clustering
and Its Application for Ransomware Detection . . . . . . . . . . . . . . . . . . . 237
Raneem Qaddoura, Ibrahim Aljarah, Hossam Faris, and Iman Almomani
Editors and Contributors

About the Editors

Ibrahim Aljarah is an associate professor of BIG Data Mining and Computational

Intelligence at the University of Jordan-Department of Information Technology,
Jordan. Currently, he is the Director of the Open Educational Resources and
Blended Learning Center at The University of Jordan. He obtained his PhD in
computer science from the North Dakota State University, USA, in 2014. He also
obtained the master degree in computer science and information systems from the
Jordan University of Science and Technology – Jordan in 2006. He obtained the
bachelor degree in Computer Science from Yarmouk University - Jordan, 2003. He
participated in many conferences in the field of data mining, machine learning, and
Big data such as CEC, GECCO, NTIT, CSIT, IEEE NABIC, CASON, and
BIGDATA Congress. Furthermore, he contributed in many projects in USA such as
Vehicle Class Detection System (VCDS), Pavement Analysis Via Vehicle
Electronic Telemetry (PAVVET), and Farm Cloud Storage System(CSS) projects.
He has published more than 60 papers in refereed international conferences and
journals. His research focuses on Data Mining, Data Science, Machine Learning,
Opinion Mining, Sentiment Analysis, Big Data, MapReduce, Hadoop, Swarm
intelligence, Evolutionary Computation, and large-scale distributed algorithms.

Hossam Faris is a Professor in the Information Technology Department at King

Abdullah II School for Information Technology at The University of Jordan,
Jordan. Hossam Faris received his B.A. and M.Sc. degrees in computer science
from the Yarmouk University and Al-Balqa’ Applied University in 2004 and 2008,
respectively, in Jordan. He was awarded a full-time competition-based scholarship
from the Italian Ministry of Education and Research to peruse his Ph.D. degrees in
e-Business at the University of Salento, Italy, where he obtained his Ph.D. degree in
2011. In 2016, he worked as a postdoctoral researcher with the GeNeura team at
the Information and Communication Technologies Research Center (CITIC),

ix
x Editors and Contributors

University of Granada, Spain. His research interests include applied computational

intelligence, evolutionary computation, knowledge systems, data mining, semantic
web, and ontologies.

Seyedali Mirjalili is an Associate Professor and the director of the Centre for
Artificial Intelligence Research and Optimization at Torrens University Australia.
He is internationally recognized for his advances in Swarm Intelligence and
Optimization, including the first set of algorithms from a synthetic intelligence
standpoint - a radical departure from how natural systems are typically understood -
and a systematic design framework to reliably benchmark, evaluate, and propose
computationally cheap robust optimization algorithms. He has published over 200
publications with over 20,000 citations and is in the list of 1% highly-cited
researchers by Web of Science. Seyedali is a senior member of IEEE and an
associate editor of several journals including Neurocomputing, Applied Soft
Computing, Advances in Engineering Software, Applied Intelligence, and IEEE
Access. His research interests include Robust Optimization, Engineering
Optimization, Multi-objective Optimization, Swarm Intelligence, Evolutionary
Algorithms, Machine Learning, and Artificial Neural Networks.

Contributors

Ammar Kamal Abasi School of Computer Sciences, Universiti Sains Malaysia,

George Town, Pulau Pinang, Malaysia
Salwani Abdullah Faculty of Information Science and Technology, Center for
Artificial Intelligence, Universiti Kebangsaan Malaysia, Bangi, Selangor, Malaysia
Ruba Abu Khurma King Abdullah II School for Information Technology, The
University of Jordan, Amman, Jordan
Mohammed Azmi Al-Betar Department of Information Technology, MSAI,
College of Engineering and Information Technology, Ajman University, Ajman,
United Arab Emirates;
IT Department, Al-Huson University College, Al-Balqa Applied University, Irbid,
Jordan
Hasnain Ali Formerly, Indian Institute of Technology Delhi, New Delhi, India
Ibrahim Aljarah King Abdullah II School for Information Technology, The
University of Jordan, Amman, Jordan
Mohamad Al-laham Department of Management Information Systems, Amman
University College, Al-Balqa Applied University (BAU), Amman, Jordan
Editors and Contributors xi

Iman Almomani King Abdullah II School for Information Technology, The

University of Jordan, Amman, Jordan;
Security Engineering Lab, Prince Sultan University, Riyadh, Saudi Arabia
Laila Al-Qaisi Information Systems and Networks Department, Faculty of
Information Technology, The World Islamic Sciences and Education University,
Amman, Jordan
Zaid Abdi Alkareem Alyasseri ECE Department, Faculty of Engineering,
University of Kufa, Najaf, Iraq;
Faculty of Information Science and Technology, Center for Artificial Intelligence,
Universiti Kebangsaan Malaysia, Bangi, Selangor, Malaysia
Ala’ M. Al-Zoubi School of Science, Technology and Engineering, University of
Granada, Granada, Spain
Mahmoud M. Al-Zoubi Department of Commutation and Information
Technology, Yarmouk Water Company, Irbid, Jordan
Hossam Faris King Abdullah II School for Information Technology, The
University of Jordan, Amman, Jordan
Saqib Gulzar Formerly, Indian Institute of Technology Delhi, New Delhi, India
Maria Habib King Abdullah II School for Information Technology, The
University of Jordan, Amman, Jordan
Mohammad A. Hassonah King Abdullah II School for Information Technology,
The University of Jordan, Amman, Jordan
Ahamad Tajudin Khader School of Computer Sciences, Universiti Sains
Malaysia, George Town, Pulau Pinang, Malaysia
Majdi Mafarja Department of Computer Science, Birzeit University, Birzeit,
West Bank, Palestine
Sharif Naser Makhadmeh School of Computer Sciences, Universiti Sains
Malaysia, George Town, Pulau Pinang, Malaysia
Seyedali Mirjalili Center for Artificial Intelligence Research and Optimization,
Torrens University Australia, Fortitude Valley, Brisbane, QLD, Australia
Syibrah Naim Technology Department, Endicott College of International Studies
(ECIS), Woosong University, Daejeon, South Korea
Razan Nujoom King Abdullah II School for Information Technology, The
University of Jordan, Amman, Jordan
João P. Papa Department of Computing, São Paulo State University - UNESP,
Bauru, Brazil
xii Editors and Contributors

Raneem Qaddoura Information Technology, Philadelphia University, Amman,

Jordan
Dao Vu Truong Son International University, Vietnam National University -
HCMC, Ho Chi Minh City, Vietnam
Pham Nhat Tan International University, Vietnam National University - HCMC,
Ho Chi Minh City, Vietnam
Hamza Turabieh Information Technology Department, CIT College, Taif
University, Taif, Saudi Arabia
Introduction to Evolutionary Data
Clustering and Its Applications

Ibrahim Aljarah, Maria Habib, Hossam Faris, and Seyedali Mirjalili

Abstract Clustering is concerned with splitting a dataset into groups (clusters) that
represent the natural homogeneous characteristics of the data. Remarkably, cluster-
ing has a crucial role in numerous types of applications. Essentially, the applications
include social sciences, biological and medical applications, information retrieval
and web search algorithms, pattern recognition, image processing, machine learn-
ing, and data mining. Even that clustering is ubiquitous over a variety of areas.
However, clustering approaches suffer from several drawbacks. Mainly, they are
highly susceptible to clusters’ initial centroids which allows a particular dataset to
easily fall within a local optimum. Handling clustering as an optimization problem is
deemed an NP-hard optimization problem. However, metaheuristic algorithms are a
dominant class of algorithms for solving tough and NP-hard optimization problems.
This chapter anticipates the use of evolutionary algorithms for addressing the prob-
lem of clustering optimization. Therefore, it presents an introduction to clustering
and evolutionary data clustering, reviews thoroughly the applications of evolutionary
data clustering and its implementation approaches.

Keywords Data clustering · Optimization · Evolutionary computation ·

Nature-inspired algorithms · Swarm intelligence · Metaheuristics unsupervised
learning · Machine learning · Data mining

I. Aljarah (B) · M. Habib · H. Faris

King Abdullah II School for Information Technology, The University of Jordan, Amman, Jordan
e-mail: [email protected]
M. Habib
e-mail: [email protected]; [email protected]
H. Faris
e-mail: [email protected]
S. Mirjalili
Center for Artificial Intelligence Research and Optimization, Torrens University Australia,
Fortitude Valley, Brisbane, QLD, Australia
e-mail: [email protected]

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer 1
Nature Singapore Pte Ltd. 2021
I. Aljarah et al. (eds.), Evolutionary Data Clustering: Algorithms and Applications, Algo-
rithms for Intelligent Systems,
https://doi.org/10.1007/978-981-33-4191-3_1
2 I. Aljarah et al.

1 Introduction

Clustering is a typical data mining technique with plentiful applications in diverse

fields of science. For instance, it is popular in the case of grouping a cohort of patients
based on similar symptoms or promoting the performance of a web search engine
by finding documents of similar topics [2, 4, 5, 76]. Clustering is an unsupervised
type of learning that groups unlabeled and m-dimensional data instances based on
common characteristics. Clustering, in literature, is defined in several ways, but
is generally identified as the process of recognizing the natural and homogeneous
groups within data. While in which, the objects of every single cluster are all similar,
but also distinct from other objects of other clusters. Quantifying the similarity or
dissimilarity among data is achieved through the utilization of proximity measures
such as distance measures. However, determining a quality measure of the obtained
clusters leads to formulating the clustering problem as an optimization problem.
Where a set of data that has (n) points needs to be clustered into a set of (k) groups
by optimizing certain fitness or objective function. Such objective functions might
be the sum squared error, homogeneity, purity, or others [38, 69, 70].
For a decade ago, clustering has been considered as a hard optimization problem
[11]. For instance, searching for the optimal set of clusters requires calculating the
minimum distances of all points from the initially selected centroids. Finding the
minimum distances of all points is computationally expensive and infeasible to solve
in a sensible amount of time. Primarily, optimization problems are divided into single-
objective problems or multi-objective problems. In literature, however, the clustering
problem is formulated as either a single-objective optimization problem or a multi-
objective optimization problem. This chapter focuses on the problem of clustering
as a single-objective optimization problem. Essentially, there are two popular ways
for solving optimization problems; the deterministic algorithms or the stochastic
algorithms. Deterministic algorithms always follow the same strategy to solve the
problem, which results in having the same solutions each time. Even though the
stochastic algorithms tend to produce approximate global optimum. Yet the stochastic
algorithms have the ability to generate more efficient global optimal solutions, even
though they are not always guaranteed. Primarily, stochastic algorithms adopt a
randomization component that leads to having different solutions at each run of the
algorithm [7, 9, 10, 67, 68].
A remarkable kind of stochastic algorithms is the metaheuristics. Metaheuristic
algorithms can search at local-level and global-level, which is known as exploitation
and exploration, respectively. A successful metaheuristic algorithm is the one that bal-
ances exploration and exploitation. Nature-inspired algorithms are well-established
examples of metaheuristic algorithms, where they are designed in a way that mim-
ics physical, biological, or chemical phenomena found in nature. Such examples of
nature-inspired algorithms are the Genetic Algorithm (GA) [34], that are inspired by
Darwinian principles of evolution, and the Particle Swarm Optimization (PSO) [29],
which is inspired by the social behavior of birds flocking.
Introduction to Evolutionary Data Clustering and Its Applications 3

This chapter introduces the definition of clustering, its main approaches, as well
as its limitations. Upon that, it presents an alternative outstanding approach that is
the evolutionary data clustering. Evolutionary data clustering has shown substantial
abilities in overcoming the drawbacks of classical data clustering methods, where it
mainly integrates the evolutionary algorithms as an optimization algorithm. The aim
is to address the clustering problem of the premature convergence by skipping the
need to determine the initial centroids, or the prior setting of the number of clusters.
The rest of the chapter is organized as follows. Section 2 is an introduction to data
clustering. Section 3 is an analytical discussion of common clustering approaches.
Section 4 presents and analyzes an alternative approach of clustering, which is evo-
lutionary clustering. Section 5 shows a thorough review of evolutionary clustering in
a wide range of applications. Finally, we summarize our work and suggest possible
future work in Sect. 6.

2 Clustering

Clustering or cluster analysis is a fundamental data mining technique and is the pro-
cess of dividing a collection of data instances into groups. Each group is a cluster of
similar objects which are dissimilar to other objects of other clusters. Each cluster is
defined by a central point and a proximity metric that measures the level of similarity
or dissimilarity of the candidate data points. Clustering analysis results in a set of
clusters with the objective of having the most compact clusters of similar points and
having the dissimilar clusters as the most separated from each other. Given a large
amount of data, it is impractical to cluster the data manually by human’s capacity,
rather, there are specialized computational algorithms for clustering. However, differ-
ent clustering algorithms result in different sets of clusters. Therefore, quantifying
the quality of the produced clusters is an essential step throughout the clustering
process [3, 6, 31, 32]. Figure 1 presents the potential output of clustering methods
relying on the used algorithm and the application field.
As clustering categorizes the data into clusters, we can say that it performs a
kind of implicit classification. Yet it is distinct from the classification process, where
mainly the data is unlabeled. As a result, it is known as an unsupervised type of
learning. A typical clustering algorithm is described as follows; assuming a dataset
D which has a set of n data instances {x1 , x2 , ..., xn }, where each data instance has
d dimensions. The clustering analysis results in a set of clusters C, with K number
of clusters; so that C = {C1 , C2 , ..., C K }. However, a set of constrains should be
satisfied
1. The combination of all clusters {C1 ∪ C2 ∪ ... ∪ C K } equals to D.
2. The intersection of any two clusters is an empty set {C1 ∩ C2 = φ}.
3. The cardinality of any cluster is a non-empty set {|C j | = φ}.
Based on the previous definition of clustering; each data instance must be attached
to one cluster that is known as an exclusive type of clustering. However, a broader
4 I. Aljarah et al.

(a) (b) (c)

(d) (e)

Fig. 1 A representation of different clustering outcomes depending on either the clustering algo-
rithm or the application area. a shows the data before clustering. In b and c diverse groups of data
may be obtained. While, d and e exhibit how different algorithms result in different clusters

type of clustering is the fuzzy clustering. Taking an example, the task of classifying
a library book of bioinformatics to one class of materials, where it belongs to the
biology section, as well as to the computer science section. So, in fuzzy clustering,
each data instance might belong to several clusters to some extent based on a pre-
defined membership function. An exceptional case of fuzzy clustering is when the
membership function has either 1 (belong) or 0 (does not belong), which results in
having each data point belonging to one cluster [78].
Since the eminent advantage of clustering is to group the similar objects; cluster-
ing has numerous applications in various fields. For instance, in marketing we may
be interested in finding customers with similar shopping behaviors, or in academic
institutions, we may analyze the students’ academic performance by grouping the
students who have similar studying behaviors. In addition, clustering has vast appli-
cations areas such as image segmentation and outlier detection, and the detection
of tumor or fraudulent activities. Furthermore, clustering is considered a significant
technique for revealing hidden and unknown patterns within data. Although cluster-
ing is adopted in various fields, it still faces several challenges [78].
In order to achieve the objective of clustering; this requires clustering to handle
large and different types of data such as data with complex graphs and networks.
Further, the data itself might be high-dimensional data which perhaps be highly
dispersed and skewed; hence increases the computational cost. Besides that, the
Introduction to Evolutionary Data Clustering and Its Applications 5

clusters might be in a complex, random form that is a non-convex shape, which

makes the process of finding the optimal set of clusters more challenging and time-
consuming [38].

3 Approaches of Clustering

This section presents an overview of basic clustering methodologies that are broadly
categorized into partitioning and hierarchical approaches.

3.1 Partitioning

In partitioning approaches, the data is divided into several nonoverlapping groups,

where each group is a cluster with at least one element. Mainly, partitioning algo-
rithms perform an exclusive clustering, which means that each object in the dataset
belongs only to one cluster. Partitioning algorithms calculate the proximity scores
among the objects by the utilization of distance measurements. In literature, there
are several common partitioning algorithms; including K-means [55], K-medoids
[64], PAM [47], CLARA [47], CLARANS [59]. However, among all, k-Means is
the most common partitioning-based clustering method. Hence, the following sub-
section describes the k-means algorithm.

3.1.1 k-Means Clustering

The k-Means clustering is a heuristic approach for grouping a dataset into a set
of clusters. The k-means algorithm starts by determining the number of clusters k
that is preferred for dividing the data. Assuming a dataset D with n objects that
have m dimensions and k number of clusters; then the k-means will result in a set of
clusters C = {C1 , C2 , ..., Ck }, where each cluster is a unique set of objects, while any
cluster C j must belong to the dataset D. The aim of clustering generally and k-means
particularly is to gather the most similar objects in one cluster while, simultaneously,
assures that they are dissimilar to any other object resides in other clusters. In other
words, the objective is to increase the intra-cluster (within the cluster) similarity and
decrease the inter-cluster (outside the cluster) similarity.
Handling k-means as a single-objective optimization problem requires determin-
ing an objective function to be optimized. One popular objective function of k-means
is the sum squared error. Hence, treating k-means as an optimization problem means
to best minimize the sum squared error.
6 I. Aljarah et al.

3
3

2
15 15 15
2

1
1
5 5 5

5 15 5 15 5 15

(a) (b) (c)

Fig. 2 An illustration of k-means procedure, which includes selecting the initial centroids, forming
the clusters, and iterating until the centroids are stable

Primarily, each cluster C j is represented by a center point c j . Thus, assessing the

quality of a cluster is performed by calculating the sum squared error of the distances
of all points (n) that belong to C j from the centroid point c j , as defined in Eq. 1.

|C j |

k 
SS E = dist (n i , c j )2 (1)
j=1 i=1

In which, SS E is the sum squared error of all points in the dataset, dist is the
Euclidean distance between the ith point n of xth dimension and the centroid c j . The
center c j is defined by Eq. 2, in case of m-dimensional space.

 m

dist (n i , c j ) =  (n i x − c j x )2 (2)
x=1

The procedure of k-means algorithm starts by determining the number of clusters k

and by initializing the clusters to random centroid values. For instance, at subfigure
(a) in Fig. 2, a three random centroids are selected for categorizing the data into
three clusters; hence, c1 is representing cluster (1), c2 represents cluster (2), and c3
substitutes cluster (3).
By calculating the Euclidean distance for each point from the three centroids;
each candidate point is assigned to the cluster that has the minimum distance to it.
As shown in subfigure (b), where at this iteration, the centroid of each cluster is
calculated by finding the mean of all points in the corresponding cluster. Hence, at
each iteration ever since, k-means calculates the distances of all points and find the
means for centroids until the centroids are stable, then the algorithm stops. Subfigure
(c) shows the results of k-means at the final iteration, where the centroids are stable.
Algorithm 1 presents a pseudo-code of the k-means clustering algorithm.
Introduction to Evolutionary Data Clustering and Its Applications 7

Algorithm 1 K-means clustering algorithm pseudo-code

1: procedure k- means (dataset D, clusters- number k)
2: Select randomly initial clusters centers ck
3: for each instance n i ∈ D do
4: Search for the nearest cluster center ck
5: Assign each n i instance to the nearest cluster Sk
6: Recompute the clusters’ centers based on the allocated instances
7: Repeat steps (4–6) until centers are stable
8: end for
9: end procedure

Even that k-means is very popular and a well-regarded algorithm, but it experi-
ences different challenges. Hence, one of the obvious drawbacks of k-means is that
it can easily get stuck in a local optimum during the search for optimal clustering.
Since it is highly susceptible to the initial, random centroids of the initial clusters.
Furthermore, k-means is sensitive to the presence of outliers, which might reflect
false centroids that leads to improperly build the clusters. Further, it is particularly
applied for detecting spherical-like clusters, as well as it is best suited with relatively
small-sized datasets. Addressing clustering problems as optimization problems and
reaching global optimality is very costly and exhaustive using partitioning-based
approaches. Thus, optimizing the minimum sum square error of clustering for k ≥ 2
is considered an NP-hard problem [11, 38].

3.2 Hierarchical

Essentially, hierarchical clustering approaches group the data into different levels
in a hierarchical way. The output of hierarchical clustering are clusters that consist
of sub-clusters visualized by dendrogram or binary tree. Hierarchical clustering is
achieved depending on proximity metrics. An example of hierarchical partitioning
is the grouping of the employees of a company into managers, officers, and trainee.
Mainly, hierarchical approaches are divided into agglomerative and divisive. The
agglomerative methods are bottom-up processes that start by handling each data
point as a standalone cluster, then repeatedly combine them into larger clusters.
Whereas the divisive methods are top-down processes. They start with one cluster of
all data points and then iteratively separates them into smaller clusters. The divisive
algorithms proceed until each point represents a single cluster, which is computa-
tionally very expensive. The most used hierarchical methods are BIRCH [85], CURE
[35], ROCK [36], and Chameleon [45].
8 I. Aljarah et al.

The procedure of agglomerative clustering algorithms follows a set of consecutive

steps. A general agglomerative clustering method starts by determining the number
of clusters k, calculate the proximity matrix of all clusters, then find the minimum
distances between clusters. Upon that, the closest clusters are merged together. There-
after, iteratively, the proximity matrix is calculated again, the distances are calculated
again, and the clusters are merged until there are one cluster [38].
Despite that, hierarchical approaches overcome some drawbacks of partition-
ing methods involving the initialization of the first initial centroids. But hierarchi-
cal methods still encounter different obstacles. The main challenges of hierarchical
approaches are the high execution time; due to performing many merging and splitting
points. In addition to the requirement of selecting the optimal merging and splitting
points; which if not chosen correctly, then the quality of the obtained clusters will
deteriorate.
Generally, there are several further clustering methods based on density, grid,
fuzzy theory, graph theory, and many others [28]. However, all these methods are
facing difficulties either in generalizing to cope with the big data of various densities,
or dealing with certain kinds of clusters, or handling outliers, or even the requirement
of the prior defination of some parameters as the number of clusters. Hence, diverse
alternative approaches were proposed for clustering. Evolutionary clustering are
salient alternative methodologies for clustering in compare with the partitioning and
hierarchical approaches. Evolutionary algorithms perform higher-level heuristic for
reaching the optimal clustering.

4 Evolutionary Clustering

Typically, finding high-quality solutions (clusters) is considered one of hardest opti-

mization problems, where reaching global optimality is not trivial using classical
clustering methods. Traditional clustering algorithms experience problems of falling
in local optimum due to the improper selection of the initial centroids or the expected
number of clusters. A more resilient strategy is desired for performing a local search,
as well as going further and search at a global scale. Stochastic algorithms are
an intriguing type of solutions for tackling optimization problems in a reasonable
amount of time. Stochastic search algorithms are heuristic or metaheuristic types of
algorithms, where the heuristic algorithms can generate very good solutions, but does
not necessarily ensure that the obtained solutions are the optimal solutions. Whereas,
metaheuristic algorithms are higher-level search algorithms that search at local and
global scales resulting in high-quality solutions in a reasonable amount of time [83].
Nature-inspired algorithms are involved under the realm of metaheuristic algo-
rithms that are best described as the pioneering algorithms for hard optimization prob-
lems. Roughly speaking, evolutionary algorithms are classified under the umbrella
of nature-inspired algorithms, where the nature-inspired algorithms mainly consist
of evolutionary algorithms, swarm intelligence algorithms, and other algorithms that
are inspired by physical, chemical, or social phenomena [83]. The integration of evo-
Introduction to Evolutionary Data Clustering and Its Applications 9

lutionary algorithms into the traditional clustering methods avoids the main draw-
backs of the stagnation into local optima or the premature convergence toward global
optimality.
Evolutionary algorithms are stochastic and inspired by the natural evolution prin-
ciples in nature and Darwinian theories. As evolutionary algorithms compromise
between diversification (search at global scale) and intensification (search at local
scale); this hinders them from the getting stuck in the local optimum. Conventionally,
evolutionary algorithms depend on an initial population of solutions that incorpo-
rate iteratively essential components to share information during the search pro-
cess. The evolutionary components include the selection, mutation, and crossover
(combination), which stand on the principle of the survival of the fittest. A popular
well-regarded evolutionary algorithm is the GA algorithm.
On the contrary, swarm intelligence algorithms are inspired by the collective
social behavior of birds, insects, or animals. Where they imitate, for instance, the
birds flocking, fish schooling, or the foraging behavior of bacteria. Both evolutionary
algorithms and the swarm intelligence algorithm adopt a fitness or objective function,
which aims for evaluating the quality of solutions until a stopping criterion is satisfied.
Evolutionary clustering as a term implies the application of evolutionary algo-
rithms for performing clustering. Recently, evolutionary clustering has been con-
sidered as one of the most popular kinds of clustering approaches in various fields
of science and industry. Evolutionary clustering is involved in various clustering
tasks such as to find the optimal number of clusters, to best optimize the quality
of the potential clusters, or for performing feature selection. One of the early pro-
posed evolutionary clustering approaches is the genetic algorithm-based clustering,
by Maulik and Bandyopadhyay in 2000 [56], where its objective was to find the most
appropriate number of clusters by taking advantage of the searchability of the GA
algorithm. The next subsection demonstrates the basic principle of clustering using
the GA algorithm.

4.1 GA Algorithm for Clustering

The main idea of using the GA algorithm for clustering is to exploit the searchability
of GA in finding the optimal set of centers. GA based clustering algorithm follows
the same steps of the original GA algorithm. GA algorithm starts by initializing a
population of individuals, then enter a loop for performing the selection, combination,
mutation, and then the evaluation of individuals. GA algorithm stops whenever a
stopping criterion is met. The pseudo-code steps of GA algorithm are represented in
Algorithm 2.
10 I. Aljarah et al.

Fig. 3 A representation of Chromosome Gene

GA individual for clustering.
The data has 2-dimensions,
in which, the centroid of the 3.3 5.0 1.5 10.2 7.0 9.6
first cluster is
c1 = (3.3, 5.0), the second
center is c2 = (1.5, 10.5)
and the third center is 3.3 5.0 1.5 10.2 7.0 9.6
c3 = (7.0, 9.6)

C1 C2 C3

Algorithm 2 Conventional GA algorithm pseudo-code

1: procedure GA (PoP- size, crossover- rate, mutation- rate)
2: Initialize the population of random individuals
3: Evaluate each individual
4: while the stopping criteria is not reached do
5: Select the best individuals for the next generation
6: Crossover (recombine) the best parent individuals
7: Mutate genes of offspring
8: Evaluate each new individual
9: Increment the loop counter
10: end while
11: end procedure

In order to prepare the GA for performing clustering, the individuals should be

encoded properly, and a fitness criterion is defined. Hence, as the problem is to
find the best clusters, the individuals represent the potential centroids. Taking into
consideration the dimensionality of the problem m, each individual represents the
candidate centroids of all clusters. Therefore, if the number of clusters is k, then
the length of the individual (chromosome) is m × k. Each gene in a chromosome
represents the value of one dimension of a center point. Thus, each chromosome is
a vector of continuous values. Figure 3 shows a representation for an individual.
Initially, all individuals in the population are assigned the random center points
from the dataset. In order to evaluate an individual, each unlabeled point is categorized
into one of the clusters based on a fitness value. Herein, the fitness function is the
Euclidean distance between the data instance and the respective centroid. The smaller
the Euclidean distance, the higher the similarity value; therefore, GA algorithm seeks
for the minimum distance value for each point from the corresponding centers. After
the assignment of the data instances, the values of the individuals were replaced
by the mean values of the new points. In other words, if cluster 1 has two data
instances; n 1 = (1.0, 2.5) and n 2 = (3.0, 2.0); then the new centroid is c1 = ((3.3 +
1 + 3)/3, (5.0 + 2.5 + 2)/3), which is c1 = (2.4, 3.2).
Introduction to Evolutionary Data Clustering and Its Applications 11

In the GA-clustering algorithm, all the genetic operators including the selection,
crossover, and mutation are performed similarly as in GA algorithm. The individual
that results in the minimum sum of distances is preserved throughout all generations
as the optimal solution.

5 Applications of Evolutionary Clustering

As clustering is applied to various fields of science. This section presents a review

of the clustering applications in data mining and machine learning, web search,
bioinformatics, and business intelligence.

5.1 Data Mining and Machine Learning

The implementation of evolutionary algorithms for solving the clustering problem

had accomplished several decades ago. An early implementation of evolutionary
algorithms for clustering was originally mostly dependent on the GA algorithm. In
[57], the authors used the GA algorithm for optimizing the sum of squared Euclidean
distances for k-means, in an attempt for reaching the optimal set of clusters. In which,
the proposed algorithm remarkably escaped the problem of premature convergence
throughout all the conducted experiments. In [74], a different evolutionary clustering
method is developed, which is based on the evolutionary programming algorithm.
The developed algorithm aimed to search for the optimal number of clusters, besides
identifying the best representative centroids. However, it is designed to handle spher-
ical or crisp-like types of clusters.
In [79], the authors offered an evolutionary clustering method using the parti-
cle swarm optimization algorithm, which was implemented mainly to search for
the initial clustering centroids, as well as to promote the clustering process toward
optimality. Where the proposed algorithm outperformed k-means in terms of inter-
cluster and intra-cluster distances, in addition to the quantization error. Whereas in
[13], four evolutionary variants of k-means proposed that adopted diverse genetics
operators for the objective of having the best grouping of data. In which, the authors
used the Rand Index and the Simplified Silhouette as fitness functions. Nevertheless,
as k-means faces a premature convergence; authors in [60] design a hybrid evolu-
tionary algorithm for clustering, which consists of particle swarm optimization and
simulated annealing for searching for the optimal clustering, where it accomplished
better convergence ability than the k-means algorithm.
Further, in [75], a novel clustering algorithm is designed that is relied on Firefly
Algorithm (FA). Whereas, [44] presented a new evolutionary clustering algorithm
that is based on Artificial Bee Colony optimization algorithm, which was utilized for
seeking the superior clustering set. Even that the authors used the proposed ABC-
Clustering for classification purposes, but it showed efficient performance results.
12 I. Aljarah et al.

In addition, [43] proposed an ABC algorithm for optimizing the fuzzy clustering
algorithms by minimizing the objective function of the sum squared error. [4] devel-
oped a parallel PSO algorithm for clustering that showed effective scalingability with
large datasets. Whereas, [14] proposed a parallel Bat algorithm for clustering large-
scale data. [6] introduced a clustering approach based on the Glowworm swarm
algorithm for determining the optimal centers of clusters, which showed superior
results in terms of purity and entropy. Nonetheless, [27] presented the integration of
an improved adaptive genetic algorithm into fuzzy c-means, in order to optimize the
initial cluster centers by optimizing the sum squared error based on the Euclidean
distance.
Authors in [76] implemented static and dynamic clustering approaches based on a
recent evolutionary algorithm which is the Multi-verse Optimizer (MVO). In which,
MVO was adopted for searching for the optimal grouping of data that is named as
the static clustering method. As well as, MVO has been used as a dynamic clustering
approach for finding the optimal number of clusters. The proposed evolutionary
clustering algorithms outperformed the Clustering Genetic Algorithm (CGA) and
the Clustering Particle Swarm Optimization (CPSO), in regards to purity and entropy
evaluation measures.
Interestingly, authors in [82] utilized a multi-objective evolutionary algorithm for
tackling the problem of the optimal number of clusters. Hence, the Non-dominated
Sorting Genetic Algorithm (NSGA-II) is used to perform a multi-clustering that
takes advantage of the parallelism of NSGA-II. In which, the number of clusters and
the sum squared distances were used as objective functions. Further, in [8], a novel
evolutionary clustering method is designed for approaching the problem of stagnation
in local optimal, which showed superior performance results. The novel method is
a combination of Grey Wolf Optimizer (GWO) and Tabu search, where the fitness
function is the sum of squared errors. Whereas, [23] utilized a many-objective ABC
algorithm for clustering software modules, which achieved robust results against
different multi-objective evolutionary algorithms as the NSGA-III and MOEA/D.
Moreover, [10] investigated the effectiveness of a Multi-verse optimizer for
addressing the problem of clustering. Where it surpassed the efficiency of PSO,
GA, and Dragonfly algorithms regarding purity, homogeneity, and completeness.
[58] proposed a hybrid of memetic algorithm and adaptive differential evolution
(DE) mutation strategy. The objective of the proposed algorithm is overcoming the
premature convergence of traditional clustering. Apparently, it exhibited remarkable
results over the classical clustering algorithms. While [12], designed a modified dif-
ferential evolution algorithm for data clustering. The designed algorithm attempted
to address the slow convergence, increase solutions diversity, and balance between
the exploration and exploitation abilities. Where it obtained high compactness of the
clustering and improved the convergence speed as well.
Introduction to Evolutionary Data Clustering and Its Applications 13

5.2 Image Processing and Pattern Recognition

The authors in [61], used the evolutionary algorithms in the field of image clustering,
which applied a PSO particularly on synthetic, Magnetic Resonance Imaging (MRI)
and satellite images. The objective of the particle swarm optimization algorithm is
to search for the optimal set of centroids for a predefined number of clusters. The
proposed algorithm outperformed k-means, fuzzy c-means clustering, k-harmonic
means, and GA-based clustering algorithms in terms of quantization error. Further,
[39] implemented a fuzzy clustering approach based on Ant Colony Optimization
(ACO) for image segmentation.
As image segmentation plays a vital role in image processing and computer
vision applications; [84] developed a combination of biogeography-based optimiza-
tion algorithm and fuzzy c-means for image segmentation. The designed approach
exhibited very good performance over PSO and ABC based clustering algorithms.
In [86], a multi-objective evolutionary fuzzy clustering method is utilized for image
segmentation, where the compactness and separation were the objective functions.
Additionally, [51] proposed an improved GWO based on DE and fuzzy c-means for
segmenting synthetic aperture radar (SAR) images. In which, DE-GWO is employed
for searching the initial number of clusters and the optimal centroids. In the field of
computer vision, [20] proposed a deep clustering approach of k-means and convolu-
tional neural networks for learning and grouping visual features, where the authors
used the ImageNet and YFCC100M datasets for testing.
Nonetheless, [53] the authors deployed a clustering approach for driving pattern
investigation for an electric vehicle. Whereas, [21] applied a clustering approach
for railway delay pattern recognition. Furthermore, [33] implemented a hierarchical
clustering approach for recognizing clinically useful patterns of patient-generated
data. And [26] anticipated the application of GA-clustering approach for pattern
identification of errors or outages of smart power grids.

5.3 Web Search and Information Retrieval

Recently, the number of web pages has increased rapidly. Therefore, clustering doc-
uments (text) is important for boosting the performance of web search engines.
[1] presented an evolutionary clustering algorithm-based on Krill herd optimization
algorithm for web text clustering. In which, the proposed algorithm outperformed k-
means in terms of purity and entropy. Remarkably, [42] developed a method to solve
the problem of abstract clustering. Which elevates the information retrieval process
of a biomedical literature database (MEDLINE). In which, the authors utilized a
hybrid of GA, vector space model that represents the text and agglomerative clus-
tering algorithm. Where the agglomerative clustering is used to generate the initial
population and to find the similar texts based on similarity metrics. [62] designed an
algorithm called (EBIC) that concerns with finding a complex pattern of information
14 I. Aljarah et al.

within complex data. Particularly, in gene expression datasets. Mainly, the proposed
algorithm stands on the implementation of evolutionary parallel biclustering algo-
rithms.
Further, [72] designed a clustering-based genetic algorithm for community detec-
tion over social networks. The detection of modular communities is essential for
investigating the behavior of complex environments. As well as, aids in the develop-
ment of recommender systems. The authors utilized a modularity metric for quanti-
fying the quality of the candidate clusters. While, [22] demonstrated the application
of a heterogeneous evolutionary clustering method for the prediction of the rating of
a collaborative filtering approach. More interestingly, [19] introduced a novel algo-
rithm for language and text identification, which is called genetic algorithms image
clustering for document analysis (GA-ICDA). In [50], the authors implemented the
binary PSO for large-scale text clustering, in which, the binary PSO is used for per-
forming feature selection. Furthermore, [54] designed a bio-inspired fuzzy clustering
ensemble approach for better personalized recommendations through collaborative
filtering.
Aspect-based summaries have a significant role in opinion mining. [66] proposed
an evolutionary clustering algorithm for aspect-summarization. [73] examined the
usage of multi-objective based DE algorithm for clustering, in the context of scientific
documents clustering. In which, the Self-Organizing Map (SOM) is implemented
with DE for searching the optimal set of clusters, while the Pakhira-Bandyopadhyay-
Maulik index and Silhouette index were used as a fitness functions for optimization.

5.4 Bioinformatics

Digital image processing is an integral part of several scientific areas. In particular,

the case of the detection of cancer. Where [40] proposed a hybrid of firefly algorithm
and k-means clustering for the detection of brain tumors. In which, the algorithm
objective is to perform brain image segmentation with Otsu’s method as a fitness
function for FA and the inter-cluster distance for k-means. The aim of k-means is
to search for the optimal centroids. Hence, it obtained very good results in terms of
different image segmentation quality measures. Halder et al. [37] proposed a genetic
algorithm-based spatial fuzzy c-means for medical image segmentation. Further, [41]
anticipated the integration of multi-objective evolutionary algorithms for clustering
complex networks. Thus, the authors proposed a novel multi-objective evolutionary
algorithm-based on decomposition and membrane structure for community detec-
tion. Evidently, the authors stated that the proposed algorithm is convenient for
clustering disease-gene networks and DNA binding protein networks identification.
Moreover, [52] interpreted the implementation of multi-objective evolutionary clus-
tering in case of RNA sequencing, where the high-dimensionality and sparseness of
data are the major challenges. Nonetheless, [65] deployed a multi-objective algo-
rithm that is hybridized with intensification and diversification components for gene
clustering. In which, quantifying genetic similarities has benefits in advancing the
Introduction to Evolutionary Data Clustering and Its Applications 15

science of diseases and metabolic disorders. The cluster-quality index, compactness,

and separation were all utilized for assessing the fitness function. In [15], the authors
presented the potential of stochastic biclustering algorithms in the field of microarray
data analysis, which has significant importance in finding the similar genes of similar
gene expression data.
Bara’a et al. [17] applied a novel multi-objective evolutionary algorithm, for
detecting the functional modules of protein-protein interaction networks. In which,
the authors adopted a heuristic component called protein complex attraction and
repulsion for assimilating the topological features of the networks. Additionally,
[16] proposed a hybrid of cuckoo search with Nelder–Mead algorithm for biclus-
tering microarray, gene expression data. Where the cuckoo search algorithm seeks
for the optimal clusters. The proposed algorithm showed reasonable improvements
over other swarm intelligence and biclustering algorithms. While [71], designed two
approaches of multi-objective optimization, which are based on PSO and differential
evolution for clustering gene expression data in the context of cancer classification.
Regarding the proposed algorithms, the Xie–Beni index and FSym-index were used
as objective functions.

5.5 Business Intelligence and Security

Chou et al. [24] proposed a hybrid of GA and fuzzy c-means for bankruptcy predic-
tion. In which, the fuzzy c-means is integrated as a fitness function in order to seek
for the best set of features that improve the prediction accuracy of GA algorithm.
Additionally, [49] developed a two-stage clustering method for order clustering. It
aims for minimizing the production time and the machine idle time. The proposed
approach depends on neural networks and a hybrid of GA and PSO, by utilizing the
sum of Euclidean distances as a fitness function. In which, the designed approach
outperformed the GA, PSO, GA-PSO, and other variants. Further, Vehicular Ad-hoc
Networks (VANETs) is a sort of intelligent transportation systems. Whereas, devel-
oping a routing protocol is fundamental for VANETs since its stochastic topology.
In [30], the authors designed a GWO-clustering approach for best controlling the
routing and stability of such scalable networks. Recommender systems play a sig-
nificant role in e-commerce and social networks. Since they seek for customizing
the recommendations for each user based on their preferences. Berbague et al. [18]
provided an evolutionary clustering approach for enhancing the procedure of rec-
ommender systems. The proposed approach is a combination of GA and k-means.
In which, the fitness function is the summation of both the group-precision and
centers-diversity. Whereas, [46] designed a hybrid of cuckoo search and k-means
for an efficient collaborative filtering approach for a movie recommender system.
Mashup services is a technology to enhance the application of the Internet of things.
In [63], the authors presented a GA-based clustering approach which aims to cluster
mashup services over a cloud of Internet of things. In which, a structural similarity
method (SimRank) is used, while the GA is utilized to find the optimal number of
16 I. Aljarah et al.

clusters. Furthermore, [25] designed an evolutionary approach with fuzzy c-means

for clustering solar energy; in order to establish a solar power plant. The proposed
approach compares each of PSO, DE, and GA for the best clustering, in terms of
Calinski-Harabasz, Davies-Bouldin, and Silhouette indices. Hence, the performance
results were superior over the fuzzy c-means.
Enhancing the security of computational services is of critical importance.
Recently, Intrusion Detection Systems (IDSs) have attracted crucial awareness world-
wide. Numerous research studies were conducted to promote the effectiveness of IDS
and reduce its false alarm rate. In [80], the authors developed a hybrid approach of
artificial neural networks and fuzzy clustering in order to improve the efficiency of
such IDS. [48] presented a hybrid of FA and k-means for intrusion detection. In
which, the FA algorithm is utilized in order to improve the slow convergence of
k-means. Whereas, [81] designed an improved clustering approach based on adap-
tive GA and fuzzy c-means algorithms for the detection of public security events.
In which, the GA algorithm searches for the optimal centroids. Additionally, [77]
proposed a GA based on k-means for network intrusion detection. Where the GA
algorithm is utilized to search for the superior clusters regarding a combination of
inner-cluster distance and inter-cluster distance.

6 Conclusion

This chapter introduced a common unsupervised data mining technique; which is

data clustering. The chapter includes a definition of clustering or cluster analysis. The
main approaches for performing clustering (partitioning and hierarchical), alongside
the most used algorithms such as the k-means clustering. In addition, presented
a description of their major challenges and limitations. Nonetheless, the chapter
investigated the formulation of clustering as NP-hard optimization problem. Where
seeking for the optimal clustering with at least 2 clusters, is deemed as a tough
optimization problem.
Therefore, the chapter suggested the integration of an alternative approach, which
is the evolutionary algorithms for clustering. Evolutionary algorithms showed great
efficiency in finding global optimal solutions. Evolutionary clustering uses evolu-
tionary algorithms such as the GA algorithm for searching the optimal clustering.
The optimal clustering involves the search of the best centroids or the optimal num-
ber of clusters by optimizing certain fitness function. Evolutionary clustering has
a wide range of applications in various fields. So, in this chapter, the implementa-
tion of evolutionary clustering is covered in the area of data mining and machine
learning. In image processing, web search, bioinformatics, and business intelligence
applications.
Introduction to Evolutionary Data Clustering and Its Applications 17

References

1. Abualigah, Laith Mohammad, Ahamad Tajudin, Khader, Mohammed Azmi, Al-Betar, and
Mohammed A. Awadallah. 2016. A krill herd algorithm for efficient text documents clustering.
In 2016 IEEE symposium on computer applications and industrial electronics (ISCAIE), pp.
67–72. IEEE.
2. Al-Madi, Nailah, Ibrahim, Aljarah, and Simone A. Ludwig. 2014. Parallel glowworm swarm
optimization clustering algorithm based on mapreduce. In 2014 IEEE Symposium on Swarm
Intelligence, pp. 1–8. IEEE.
3. Al Shorman, Amaal, Hossam, Faris, and Ibrahim, Aljarah. 2020. Unsupervised intelligent
system based on one class support vector machine and grey wolf optimization for iot botnet
detection. Journal of Ambient Intelligence and Humanized Computing, 11(7):2809–2825.
4. Aljarah, Ibrahim, and Simone A. Ludwig. 2012. Parallel particle swarm optimization clustering
algorithm based on mapreduce methodology. In 2012 Fourth World Congress on Nature and
Biologically Inspired Computing (NaBIC), pp. 104–111. IEEE.
5. Aljarah, Ibrahim, and Simone A. Ludwig. 2013. Mapreduce intrusion detection system based
on a particle swarm optimization clustering algorithm. In 2013 IEEE congress on evolutionary
computation, pp. 955–962. IEEE.
6. Aljarah, Ibrahim, and Simone A. Ludwig. 2013. A new clustering approach based on glowworm
swarm optimization. In 2013 IEEE congress on evolutionary computation, pp. 2642–2649.
IEEE.
7. Aljarah, Ibrahim, and Simone A. Ludwig. 2013. Towards a scalable intrusion detection system
based on parallel pso clustering using mapreduce. In Proceedings of the 15th annual conference
companion on Genetic and evolutionary computation, pp. 169–170.
8. Aljarah, Ibrahim, Majdi, Mafarja, Ali Asghar, Heidari, Hossam, Faris, and Seyedali, Mir-
jalili. 2019. Clustering analysis using a novel locality-informed grey wolf-inspired clustering
approach. Knowledge and Information Systems 1–33.
9. Aljarah, Ibrahim, Majdi, Mafarja, Ali Asghar, Heidari, Hossam, Faris, and Seyedali, Mir-
jalili. 2020. Clustering analysis using a novel locality-informed grey wolf-inspired clustering
approach. Knowledge and Information Systems 62(2):507–539.
10. Aljarah, Ibrahim, Majdi, Mafarja, Ali Asghar, Heidari, Hossam, Faris, and Seyedali, Mirjalili.
2020. Multi-verse optimizer: theory, literature review, and application in data clustering. In
Nature-Inspired Optimizers, pp. 123–141. Berlin: Springer
11. Aloise, Daniel, Amit Deshpande, Pierre Hansen, and Preyas Popat. 2009. Np-hardness of
euclidean sum-of-squares clustering. Machine Learning 75 (2): 245–248.
12. Alswaitti, Mohammed, Mohanad Albughdadi, and Nor Ashidi Mat Isa. 2019. Variance-based
differential evolution algorithm with an optional crossover for data clustering. Applied Soft
Computing 80: 1–17.
13. Alves, Vinícius S., Ricardo J.G.B. Campello, and Eduardo R. Hruschka. 2006. Towards a fast
evolutionary algorithm for clustering. In 2006 IEEE international conference on evolutionary
computation, pp. 1776–1783. IEEE.
14. Ashish, Tripathi, Sharma, Kapil, and Bala, Manju. 2018. Parallel bat algorithm-based clustering
using mapreduce. In Networking communication and data knowledge engineering, pp. 73–82.
Berlin: Springer.
15. Ayadi, Wassim, Mourad, Elloumi, and Jin-Kao, Hao. 2018. 14 systematic and stochastic biclus-
tering algorithms for microarray data analysis. Microarray Image and Data Analysis: Theory
and Practice 369.
16. Balamurugan, R., A.M. Natarajan, and K. Premalatha. 2018. A new hybrid cuckoo search
algorithm for biclustering of microarray gene-expression data. Applied Artificial Intelligence
32 (7–8): 644–659.
17. Bara’a, A. Attea, and Qusay Z. Abdullah. 2018. Improving the performance of evolutionary-
based complex detection models in protein—protein interaction networks. Soft Computing
22(11):3721–3744.
18 I. Aljarah et al.

18. Berbague, Chems Eddine, Nour, El Islem Karabadji, and Hassina, Seridi. 2018. An evolutionary
scheme for improving recommender system using clustering. In IFIP International Conference
on Computational Intelligence and Its Applications, pp. 290–301. Berlin: Springer.
19. Brodić, Darko, Alessia, Amelio, and Zoran N. Milivojević. 2018. Language discrimination by
texture analysis of the image corresponding to the text. Neural Computing and Applications
29(6):151–172.
20. Caron, Mathilde, Piotr, Bojanowski, Armand, Joulin, and Matthijs, Douze. 2018. Deep clus-
tering for unsupervised learning of visual features. In Proceedings of the European conference
on computer vision (ECCV), pp. 132–149.
21. Cerreto, Fabrizio, Bo Friis, Nielsen, Otto Anker, Nielsen, and Steven S. Harrod. 2018. Applica-
tion of data clustering to railway delay pattern recognition. Journal of Advanced Transportation
2018.
22. Chen, Jianrui, Hua Wang, Zaizai Yan, et al. 2018. Evolutionary heterogeneous clustering for
rating prediction based on user collaborative filtering. Swarm and Evolutionary Computation
38: 35–41.
23. Chhabra, Jitender Kumar et al. 2018. Many-objective artificial bee colony algorithm for large-
scale software module clustering problem. Soft Computing, 22(19):6341–6361.
24. Chou, Chih-Hsun, Su-Chen Hsieh, and Chui-Jie Qiu. 2017. Hybrid genetic algorithm and fuzzy
clustering for bankruptcy prediction. Applied Soft Computing 56: 298–316.
25. de Barros Franco, David Gabriel, and Maria Teresinha Arns, Steiner. 2018. Clustering of solar
energy facilities using a hybrid fuzzy c-means algorithm initialized by metaheuristics. Journal
of Cleaner Production 191:445–457.
26. De Santis, Enrico, Antonello Rizzi, and Alireza Sadeghian. 2018. A cluster-based dissimilarity
learning approach for localized fault classification in smart grids. Swarm and Evolutionary
Computation 39: 267–278.
27. Ding, Yi, and Fu Xian. 2016. Kernel-based fuzzy c-means clustering algorithm based on genetic
algorithm. Neurocomputing 188: 233–238.
28. Dongkuan, Xu, and Yingjie Tian. 2015. A comprehensive survey of clustering algorithms.
Annals of Data Science 2: 165–193.
29. Eberhart, Russell, and James, Kennedy. 1995. A new optimizer using particle swarm theory.
In MHS’95. Proceedings of the sixth international symposium on micro machine and human
science, pp. 39–43. IEEE.
30. Fahad, Muhammad, Farhan, Aadil, Salabat, Khan, Peer Azmat, Shah, Khan, Muhammad,
Jaime, Lloret, Haoxiang, Wang, Jong Weon, Lee, Irfan, Mehmood, et al. 2018. Grey wolf
optimization based clustering algorithm for vehicular ad-hoc networks. Computers & Electrical
Engineering 70:853–870.
31. Hossam Faris, Ibrahim Aljarah, and Ja’far Alqatawna. 2015. Optimizing feedforward neural
networks using krill herd algorithm for e-mail spam detection. In 2015 IEEE Jordan conference
on applied electrical engineering and computing technologies (AEECT), pp. 1–5. IEEE.
32. Faris, Hossam, Ibrahim, Aljarah, Seyedali, Mirjalili, Pedro A. Castillo, and Juan Julián Merelo,
Guervós. 2016. Evolopy: An open-source nature-inspired optimization framework in python.
In IJCCI (ECTA), pp. 171–177.
33. Feller, Daniel J., Marissa, Burgermaster, Matthew E. Levine, Arlene, Smaldone, Patricia G.
Davidson, David J. Albers, and Lena, Mamykina. 2018. A visual analytics approach for pattern-
recognition in patient-generated data. Journal of the American Medical Informatics Associa-
tion, 25(10):1366–1374.
34. Goldberg, David E., and John H. Holland. 1988. Genetic algorithms and machine learning.
Machine Learning, 3(2):95–99.
35. Guha, Sudipto, Rajeev, Rastogi, and Kyuseok, Shim. 1998. Cure: an efficient clustering algo-
rithm for large databases. In ACM Sigmod Record, vol. 27, pp. 73–84. ACM.
36. Guha, Sudipto, Rajeev Rastogi, and Kyuseok Shim. 2000. Rock: A robust clustering algorithm
for categorical attributes. Information Systems 25: 345–366.
37. Halder, Amiya, Avranil, Maity, and Ananya, Das. 2019. Medical image segmentation using
ga-based modified spatial fcm clustering. In Integrated Intelligent Computing, Communication
and Security, pp. 595–601. Berlin: Springer.
Introduction to Evolutionary Data Clustering and Its Applications 19

38. Han, Jiawei, Jian Pei, and Micheline Kamber. 2011. Data mining: concepts and techniques.
Elsevier.
39. Han, Yanfang, and Pengfei Shi. 2007. An improved ant colony algorithm for fuzzy clustering
in image segmentation. Neurocomputing 70 (4–6): 665–671.
40. Hrosik, Romana CAPOR, Eva, Tuba, Edin, Dolicanin, Raka, Jovanovic, and Milan, Tuba.
2019. Brain image segmentation based on firefly algorithm combined with k-means clustering.
Stud. Inform. Control, 28:167–176.
41. Ying, Ju, Songming Zhang, Ningxiang Ding, Xiangxiang Zeng, and Xingyi Zhang. 2016. Com-
plex network clustering by a multi-objective evolutionary algorithm based on decomposition
and membrane structure. Scientific Reports 6: 33870.
42. Karaa, Wahiba Ben Abdessalem, Amira S. Ashour, Dhekra Ben, Sassi, Payel, Roy, Noreen,
Kausar, and Nilanjan, Dey. 2016. Medline text mining: an enhancement genetic algorithm
based approach for document clustering. In Applications of intelligent optimization in biology
and medicine, pp. 267–287. Berlin: Springer.
43. Karaboga, Dervis, and Celal Ozturk. 2010. Fuzzy clustering with artificial bee colony algorithm.
Scientific Research and Essays 5 (14): 1899–1902.
44. Karaboga, Dervis, and Celal Ozturk. 2011. A novel clustering approach: Artificial bee colony
(abc) algorithm. Applied Soft Computing 11 (1): 652–657.
45. Karypis, George, Eui-Hong Sam, Han, and Vipin, Kumar. 1999. Chameleon: Hierarchical
clustering using dynamic modeling. Computer, (8):68–75.
46. Rahul Katarya and Om Prakash Verma. 2017. An effective collaborative movie recommender
system with cuckoo search. Egyptian Informatics Journal 18 (2): 105–112.
47. Kaufman, Leonard, and Peter J. Rousseeuw. 2009. Finding groups in data: an introduction to
cluster analysis, vol. 344. New York: Wiley.
48. Kaur, Arvinder, Saibal K. Pal, and Amrit Pal, Singh. 2018. Hybridization of k-means and
firefly algorithm for intrusion detection system. International Journal of System Assurance
Engineering and Management 9(4):901–910.
49. Kuo, R.J., and L.M. Lin. 2010. Application of a hybrid of genetic algorithm and particle swarm
optimization algorithm for order clustering. Decision Support Systems 49 (4): 451–462.
50. Kushwaha, Neetu, and Millie Pant. 2018. Link based bpso for feature selection in big data text
clustering. Future Generation Computer Systems 82: 190–199.
51. Li, M.Q., L.P. Xu, Na, Xu, Tao, Huang, and Bo, Yan. 2018. Sar image segmentation based
on improved grey wolf optimization algorithm and fuzzy c-means. Mathematical Problems in
Engineering 2018.
52. Li, Xiangtao, and Ka-Chun, Wong. 2019. Single-cell rna sequencing data interpretation by
evolutionary multiobjective clustering. IEEE/ACM transactions on computational biology and
bioinformatics.
53. Li, Xuefang, Qiang Zhang, Zhanglin Peng, Anning Wang, and Wanying Wang. 2019. A data-
driven two-level clustering model for driving pattern analysis of electric vehicles and a case
study. Journal of Cleaner Production 206: 827–837.
54. Logesh, R., V. Subramaniyaswamy, D. Malathi, N. Sivaramakrishnan, and V. Vijayakumar.
2019. Enhancing recommendation stability of collaborative filtering recommender system
through bio-inspired clustering ensemble method. Neural Computing and Applications, pp.
1–24.
55. MacQueen, James, et al. 1967. Some methods for classification and analysis of multivariate
observations. In Proceedings of the fifth Berkeley symposium on mathematical statistics and
probability, vol. 1, pp. 281–297. Oakland, CA, USA.
56. Maulik, Ujjwal, and Sanghamitra Bandyopadhyay. 2000. Genetic algorithm-based clustering
technique. Pattern Recognition 33 (9): 1455–1465.
57. Murthy, Chivukula A., and Nirmalya, Chowdhury. 1996. In search of optimal clusters using
genetic algorithms. Pattern Recognition Letters, 17(8):825–832.
58. Mustafa, Hossam M.J., Masri, Ayob, Mohd Zakree Ahmad, Nazri, and Graham, Kendall.
2019. An improved adaptive memetic differential evolution optimization algorithms for data
clustering problems. PloS One 14(5):e0216906.
20 I. Aljarah et al.

59. Ng, Raymond T., and Jiawei, Han. 2002. Clarans: A method for clustering objects for spatial
data mining. IEEE Transactions on Knowledge & Data Engineering (5):1003–1016.
60. Niknam, Taher, Babak Amiri, Javad Olamaei, and Ali Arefi. 2009. An efficient hybrid evolution-
ary optimization algorithm based on pso and sa for clustering. Journal of Zhejiang University-
SCIENCE A 10 (4): 512–519.
61. Omran, Mahamed, Andries Petrus, Engelbrecht, and Ayed, Salman. 2005. Particle swarm
optimization method for image clustering. International Journal of Pattern Recognition and
Artificial Intelligence 19(03):297–321.
62. Orzechowski, Patryk, Moshe, Sipper, Xiuzhen, Huang, and Jason H, Moore. 2018. Ebic:
an evolutionary-based parallel biclustering algorithm for pattern discovery. Bioinformatics
34(21):3719–3726.
63. Pan, Weifeng, and Chunlai Chai. 2018. Structure-aware mashup service clustering for cloud-
based internet of things using genetic algorithm based clustering algorithm. Future Generation
Computer Systems 87: 267–277.
64. Park, Hae-Sang, and Chi-Hyuck Jun. 2009. A simple and fast algorithm for k-medoids clus-
tering. Expert systems with Applications 36: 3336–3341.
65. Parraga-Alava, Jorge, Marcio Dorn, and Mario Inostroza-Ponta. 2018. A multi-objective gene
clustering algorithm guided by apriori biological knowledge with intensification and diversifi-
cation strategies. BioData Mining 11 (1): 16.
66. Priya, V., and K, Umamaheswari. 2019. Aspect-based summarisation using distributed clus-
tering and single-objective optimisation. Journal of Information Science 0165551519827896.
67. Qaddoura, Raneem, Hossam Faris, and Ibrahim Aljarah. 2020. An efficient clustering algorithm
based on the k-nearest neighbors with an indexing ratio. International Journal of Machine
Learning and Cybernetics 11 (3): 675–714.
68. Qaddoura, Raneem, Hossam, Faris, Ibrahim, Aljarah, and Pedro A., Castillo. 2020. Evocluster:
An open-source nature-inspired optimization clustering framework in python. In International
Conference on the Applications of Evolutionary Computation (Part of EvoStar), pp. 20–36.
Berlin: Springer.
69. Qaddoura, R., H. Faris, and I. Aljarah. 2020. An efficient evolutionary algorithm with a near-
est neighbor search technique for clustering analysis. Journal of Ambient Intelligence and
Humanized Computing, 1–26.
70. Qaddoura, R., H. Faris, I. Aljarah, J. Merelo, and P. Castillo. 2020. Empirical evaluation of
distance measures for nearest point with indexing ratio clustering algorithm. In Proceedings
of the 12th International Joint Conference on Computational Intelligence - Volume 1: NCTA,
ISBN 978-989-758-475-6, pp. 430–438. https://doi.org/10.5220/0010121504300438.
71. Saha, Sriparna, Ranjita Das, and Partha Pakray. 2018. Aggregation of multi-objective fuzzy
symmetry-based clustering techniques for improving gene and cancer classification. Soft Com-
puting 22 (18): 5935–5954.
72. Said, Anwar, Rabeeh Ayaz, Abbasi, Onaiza, Maqbool, Ali, Daud, and Naif Radi, Aljohani.
2018. Cc-ga: A clustering coefficient based genetic algorithm for detecting communities in
social networks. Applied Soft Computing, 63:59–70.
73. Saini, Naveen, Sriparna Saha, and Pushpak Bhattacharyya. 2019. Automatic scientific docu-
ment clustering using self-organized multi-objective differential evolution. Cognitive Compu-
tation 11 (2): 271–293.
74. Sarkar, Manish, B. Yegnanarayana, and Deepak, Khemani. 1997. A clustering algorithm using
an evolutionary programming-based approach. Pattern Recognition Letters, 18(10):975–986.
75. Senthilnath, J., S.N. Omkar, and V. Mani. 2011. Clustering using firefly algorithm: performance
study. Swarm and Evolutionary Computation 1 (3): 164–171.
76. Shukri, Sarah, Hossam Faris, Ibrahim Aljarah, Seyedali Mirjalili, and Ajith Abraham. 2018.
Evolutionary static and dynamic clustering algorithms based on multi-verse optimizer. Engi-
neering Applications of Artificial Intelligence 72: 54–66.
77. Sukumar, J.V. Anand, I. Pranav, M.M. Neetish, and Jayasree, Narayanan. 2018. Network intru-
sion detection using improved genetic k-means algorithm. In 2018 international conference on
advances in computing, communications and informatics (ICACCI), pp. 2441–2446. IEEE.
Introduction to Evolutionary Data Clustering and Its Applications 21

78. Theodoridis, Sergios, and Konstantinos, Koutroumbas. 2006. Clustering: basic concepts. Pat-
tern Recognition, 483–516.
79. Van der Merwe, D.W., and Andries Petrus, Engelbrecht. 2003. Data clustering using particle
swarm optimization. In The 2003 Congress on Evolutionary Computation, 2003. CEC’03.,
vol. 1, pp. 215–220. IEEE.
80. Wang, Gang, Jinxing Hao, Jian Ma, and Lihua Huang. 2010. A new approach to intrusion detec-
tion using artificial neural networks and fuzzy clustering. Expert Systems with Applications 37
(9): 6225–6232.
81. Wang, Heng, Zhenzhen, Zhao, Zhiwei, Guo, Zhenfeng, Wang, and Xu, Guangyin. 2017. An
improved clustering method for detection system of public security events based on genetic
algorithm and semisupervised learning. Complexity 2017.
82. Wang, Rui, Shiming Lai, Wu Guohua, Lining Xing, Ling Wang, and Hisao Ishibuchi. 2018.
Multi-clustering via evolutionary multi-objective optimization. Information Sciences 450: 128–
140.
83. Yang, Xin-She. 2010. Nature-inspired metaheuristic algorithms. Luniver Press. 2010.
84. Zhang, Minxia, Weixuan, Jiang, Yu, Xiaohan Zhou, Xue, and Shengyong, Chen. 2019. A hybrid
biogeography-based optimization and fuzzy c-means algorithm for image segmentation. Soft
Computing 23(6):2033–2046.
85. Zhang, Tian, Raghu, Ramakrishnan, and Miron, Livny. 1996. Birch: an efficient data clustering
method for very large databases. In ACM Sigmod Record, vol. 25, pp. 103–114. ACM.
86. Zhao, Feng, Hanqiang, Liu, Jiulun, Fan, Chang Wen, Chen, Rong, Lan, and Na, Li. 2018.
Intuitionistic fuzzy set approach to multi-objective evolutionary clustering with multiple spatial
information for image segmentation. Neurocomputing, 312:296–309.
A Comprehensive Review of Evaluation
and Fitness Measures for Evolutionary
Data Clustering

Ibrahim Aljarah, Maria Habib, Razan Nujoom, Hossam Faris,

and Seyedali Mirjalili

Abstract Data clustering is among the commonly investigated types of unsuper-

vised learning; owing to its ability for capturing the underlying information. Accord-
ingly, data clustering has an increasing interest in various applications involving
health, humanities, and industry. Assessing the goodness of clustering has been
widely debated across the history of clustering analysis, which led to the emergence
of abundant clustering evaluation measures. The aim of clustering evaluation is to
quantify the quality of the potential clusters which is often referred to as clustering
validation. There are two broad categories of clustering validations; the external and
the internal measures. Mainly, they differ by relying on external true-labels of the
data or not. This chapter considers the role of evolutionary and swarm intelligence
algorithms for data clustering, which showed extreme advantages over the classical
clustering algorithms. The main idea of this chapter is to present thoroughly the
clustering validation indices that are found in literature, indicating when they were
utilized with evolutionary clustering and when used as an objective function.

Keywords Data clustering · Optimization · Evolutionary computation ·

Evaluation measures · Swarm intelligence · Metaheuristics · Fitness functions ·
Validation indices

I. Aljarah (B) · M. Habib · R. Nujoom · H. Faris

King Abdullah II School for Information Technology, The University of Jordan, Amman, Jordan
e-mail: [email protected]
M. Habib
e-mail: [email protected]; [email protected]
R. Nujoom
e-mail: [email protected]
H. Faris
e-mail: [email protected]
S. Mirjalili
Center for Artificial Intelligence Research and Optimization, Torrens University Australia,
Fortitude Valley, Brisbane, Australia
e-mail: [email protected]

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer 23
Nature Singapore Pte Ltd. 2021
I. Aljarah et al. (eds.), Evolutionary Data Clustering: Algorithms and Applications, Algo-
rithms for Intelligent Systems,
https://doi.org/10.1007/978-981-33-4191-3_2
24 I. Aljarah et al.

1 Introduction

Clustering is a well-regarded unsupervised learning process in data mining [3, 8, 54,

55]. Even that clustering is endlessly utilized in various applications, but it still faces
different challenges. For instance, with the plentiful clustering algorithms, selecting
the appropriate algorithm is not trivial, choosing the number of clusters, assessing the
clustering tendency, quantifying the optimal similarity among the data points with the
respective problem, and avoiding the randomly-created clusters are all such example
obstacles. Apart from that is how to evaluate the goodness of obtained clusters and
compare them to the results of other clustering algorithms. Essentially, assessing the
quality of clusters encompasses three aspects; the clustering evaluation, the clustering
tendency, and the clustering stability. Clustering evaluation measures the goodness of
clusters based on validation metrics. The clustering tendency concentrates on deter-
mining if the respective dataset has nonarbitrary structures which result in ambiguous
clusters. While the clustering stability denotes results susceptibility when varying
the algorithm parameters. Hence, this chapter focuses on the clustering evaluation
with deeper coverage within the context of evolutionary clustering [9, 11, 12, 140,
141].
Fundamentally, there are two essential concepts relating to the clustering prob-
lem. First is the clustering validation which is also known as the evaluation using
evaluation metrics or quality measures. Second is the clustering objective function
or the fitness function that is often used in the context of clustering optimization.
Primarily, the validation of clustering involves assuring that the clustering algorithm
does not create arbitrary clusters. In addition, it is used for setting the number of
clusters and for comparing the obtained results with other implemented cluster-
ing algorithms. Clustering Validation Indices (CVI) are measures for evaluating the
results of clustering analysis. Generally, CVI are categorized into external measures,
internal measures, and relative measures. The external measures depend on data that
is not used through the clustering process as the known class labels. They measure
how close the clustering is to the class labels; so, these measures are useful when the
ground truth is available [2, 6, 7, 172].
There are numerous indices were used as external measures, such as Fowlkes-
Mallows Scores [58], Rand Index [146], Adjusted Rand Index [78], Mutual Infor-
mation based scores [180], V-measure [154], Jaccard Index [80], and Czekanowski-
Dice Index [42]. Whereas, the internal measures are used to quantify the compactness
and separation of formed clusters. Where the compactness measures how much the
data points in a single cluster are close to each other, which usually refers to within-
cluster variance metrics. The separation finds how much the formed clusters are
well-disconnected from each other, which can be measured by finding the distances
between clusters’ centers. Figure 1 shows an illustration of compactness and separa-
tion concepts. In which, for example, C1 reach the optimal compactness by having
the minimum distance between all points and the center as between (a, b1 ) points.
While the clusters (C1 , C2 , C3 ) have the best separation by having the furthest dis-
A Comprehensive Review of Evaluation and Fitness Measures … 25

Fig. 1 An Illustration of
compactness and separation C1
of three imaginary clusters in
a b1
a two-dimensional Cartesian d(a,b )
1
coordinate system. All the C2
clusters (C1 , C2 , C3 ) are

d(b
well-separated from each

1
,b 3
)
other

b3
C3

tances, which could be the representative distance between centroids. As the distance
between (C1 , C3 ).
In literature, there are numerous numbers of internal measures, such measures are
including Sum of Squared Errors (SSE) [72], Davies-Bouldin Index [47], Calinski-
Harabasz Index [26], Silhouette Coefficient [156], the Partition Coefficient [23], Xie-
Beni Index [194], I Index [114], Average Between Group Sum of Squares (ABGSS)
[90], Intra-cluster Entropy [151], Overall Cluster Deviation [74], Cluster Separation
[151], Dunn Index [53], and Edge Index [170]. While the relative CVIs are a hybrid
of external and internal indices.
On the other hand, the aim of objective (fitness) functions is to assess the quality
of the clustering results during the clustering process [142, 143]. As clustering can
be formulated as an optimization problem, the objective function is the optimization
function. Traditional clustering algorithms as the k-means algorithm uses the sum of
squared errors as an objective function. Usually, the sum of squared errors function
uses the Euclidean distance to find the distances between each data point and the
corresponding centroid. In which, the algorithm forms clusters and stops whenever
it reaches the minimum sum of squared errors value. In the context of evolutionary
clustering, the evolutionary algorithms optimize the clustering process by optimizing
a fitness function aiming for searching the optimal solutions. Broadly speaking, evo-
lutionary objective functions for clustering can be classified into distance measures
and similarity or dissimilarity measures. There are a variety of utilized distance mea-
sures such as the Euclidean, Mahalanobis, Minkowski, and cosine distances. Also,
a popular similarity measures are the Jaccard and cosine similarity indices.
Over the decades, wide-range of clustering validation measures have been devel-
oped. Few of them have been utilized for evolutionary data clustering. Milligan and
Cooper [119] presented a comprehensive survey of CVI accounting for 30 measures.
Figure 2 shows a classification of internal and external clustering validation mea-
sures indicating whether they used within the area of evolutionary computation. In
the figure, the red circles denote that the respective index is utilized with evolution-
ary clustering. The black and green circles indicate the internal and external indices,
respectively. The blue one refers to newly proposed indices. The light orange indi-
26 I. Aljarah et al.

cates a widely used index, and the grey circle presents a measure for fitness functions.
Noticeably, the red circles with an overline dash symbol represents an index rarely
used with evolutionary clustering.
The rest of the chapter is organized as follows. Section 2 represents the evolu-
tionary clustering validation measures, encompassing external and internal measures
only integrated within evolutionary clustering approaches, and presents other indices
not used with evolutionary clustering, as described in Fig. 2. Section 3 presents the
objective functions that were only used as fitness measures. While Sect. 4 is a sum-
mary of the chapter.

2 Evolutionary Clustering Validation Indices (CVI)

This section presents the clustering validation measures used with evolutionary clus-
tering including both the external and the internal. In addition, it points out their
implementation in further applications.

2.1 External Measures

External validation measures are used when the ground truth labels are known. They
measure the extent to which cluster labels match externally supplied class labels.
Hence, it targets maximizing cluster homogeneity and completeness. Maximizing
the homogeneity means a maximized purity, while higher completeness indicates
better ability to assign points to their true clusters. The following subsections show
the external validation measures for evolutionary clustering.

2.1.1 Rand Index

Rand Index (RI) measures the similarity and consistency between the resultant groups
of two random clusterings of a dataset [146]. In which, the comparison is conducted
for all data points in each group (cluster) of each clustering (partition). In other
words, it examines if two data points are in the same cluster at all partitions or even
if they are in different clusters between partitions. Given two data points (y, y  ), the
two points (y, y  ) are paired points if they exist in the same cluster of a partition.
Equation 1 defines the formula of Rand index given two partitions (P1 , P2 ) and n
data points [146].
a+d a+d
RI = = n  (1)
a+b+c+d 2

where a is the number of paired points that located in the two partitions (P1 , P2 ).
Parameter d is the number of points that are not paired in any of the partitions
 Clustering Validation
Measures

Internal External Fitness Functions

_ R-Squared Within Group Sum of Squares Rand Index, Adjusted Rand Index Root-Mean-Squared Error
Ratkowsky-Lance Index
_ Score Index Between Group Sum of Squares Mutual Information Scale
Ray-Turi Index Euclidean
_ Sym-Index In-Group Proportion (IGP) Maximum Matching
Scott-Symons Index Mahalanobis
COP-Index Calinski-Harabasz Mirkin Metric
Wemmert-Gancarski Index _ _ Manhattan
Gap Statistics Trace_W Index Fisher‛s Discriminant Ratio V-Measure
Minkowski
VCN Measure SV Index Big Data Dunn's Index Fowlkes-Mallows Score Cosine
_
KCE Index OS Index Dunn's Index Wallace Index Weighted Within Group Sum
STR Index McClain & Rao Index Big Data Silhouette Index Sokal-Sneath Indices
NK Criterion Silhouette Index F-Measure, Recall, Precision Intra-Cluster Distance
SD Validity Index
BetaCV Index Purity
Root-Mean-Squared STD Davis-Bouldin Index Inter-Cluster Distance
Least Squared Error Index _ Entropy
_ Fukuyama-Sugeno Index Xie-Beni Index Negated-Average Association
Kwon's index Jaccard Index
CS-Index S-Dbw Validity Pearson Correlation
Zhang Index Czekanowski-Dice Index
_ C-Index Hartigan Index Hamming distance
Marriott index Kulczynski Index
_ Ball Hall Index Bayesian Information Criterion Jaccard Similarity
TraceCovW index. _ McNemar Index
_ Banfeld-Raftery Index Krzanowski-lai Index Gaussian Similarity
Rubin index Phi Index
Det Ratio Index WB Index Compactness & Separation Measure of Clusters
Frey index _ Rogers-Tanimoto Index
_ Wint Index Gamma Index Kendall-Tau Rank
Cubic Clustering Criterion Hubert Gamma Statistic
Tau Index Bezdek's Partition Coefficient Variance
Beale index Partition Index Russel-Rao Index
_ G_plus Index Multi-kernel Function
Duda index
A Comprehensive Review of Evaluation and Fitness Measures …

Ksq_DetW Index Intra-cluster Entropy Average Density

Pseudot2 index Log_Det_Ratio Index Overall Cluster Deviation
_ Spearman Coefficient Correlation
Scott index Log_SS_Ratio Index Negative Entropy Increment
_
Friedman index PBM Index Edge Index
Point-Biserial Index Cluster Separation
VSC Index Graded Distance Index
Fuzzy Hypervolume
PCAES Index

Evolutionary CVI Internal CVI Fitness Measure Newly-Proposed CVI External CVI Popular CVI
27

Fig. 2 A tree-like illustration of internal and external clustering validation measures, as well as the used fitness functions with evolutionary clustering
28 I. Aljarah et al.

(P1 , P2 ). While b is the number of points that are paired in one partition P1 and c is
the number of points that are paired just in the second partition P2 .
RI has a value in the range [0, 1]. Where 0 means the output of the clusterings
does not have any matches of the clustered data points. While RI with value 1 means
that all clusterings have the same output.
Rand index was used to evaluate a new evolutionary automatic clustering tech-
nique. The proposed algorithm is implemented by combining the k-means algorithm
and the teaching-learning-based optimization algorithm (TLBO) [95]. A major draw-
back of RI is the sensitivity to the number of clusters in the partitions. An alternative
variant of RI was proposed to overcome its weakness which is the Adjusted Rand
Index (ARI) [78].
RI was also used in a novel combinatorial merge-split approach for automatic
clustering using imperialist competitive algorithm [5]. In order to evaluate the qual-
ity of the clustering method, the new approach was compared with various meth-
ods including basic Imperialist Competitive Algorithm (ICA), Hybrid K Modify
Imperialist Competitive Algorithm (K-MICA), Improved Imperialist Competitive
Algorithm with mutation operator (IICA-G), and Explorative Imperialist Competi-
tive Algorithm (EXPLICA) on 10 datasets. The results indicated that the proposed
method was better than other algorithms. Another application of this index is that it
was used in [196] to evaluate the performance of an evolutionary clustering method.

2.1.2 Adjusted Rand Index

Adjusted Rand (AR) index [78] is an external validity index that measures the similar-
ity between two clusterings. It is the corrected for chance version of the Rand index;
which means that a random result has 0 score. Equation 2 shows the formulation of
AR Index.
R I − E[R I ]
AR = (2)
max(R I ) − E[R I ]

where R I is the Rand Index, E[R I ] is the Expected Rand index, and max[R I ] is the
Maximum Rand index.
Adjusted Rand index was used in clustering stability-based evolutionary K-Means
to measure the clustering accuracy [77]. To evaluate the evolutionary K-Means, its
results were compared to two consensus clustering algorithms; a clustering stability-
based algorithm and a multi-index clustering approach. The results indicated that
the proposed method is more robust to noise. The strength of this index is that it is
corrected for chance. Adjusted Rand index was also used in evolutionary spectral
clustering with adaptive forgetting factor as in [195], in order to evaluate the perfor-
mance when using different values of the forgetting factor. Experiments indicated
that the proposed method has results better than the usage of fixed forgetting factors.
AR index was also used in a merge-split approach for automatic clustering using
imperialist competitive algorithm [5].
A Comprehensive Review of Evaluation and Fitness Measures … 29

2.1.3 Mutual Information Based Scores

Mutual Information (MI) based scores [180] are external validity indexes that eval-
uate the similarity of two clusterings. They compute the amount of information
collected about a random variable by studying the other random variable. There are
two versions; the Normalized Mutual Information (NMI) and the Adjusted Mutual
Information (AMI). Equations (3–5) define MI score, NMI, and AMI, respectively
[189].
R  C
ni j n i j /N
M I (X, Y ) = log . (3)
i b j /N
N a 2
i=1 j=1

where R is the number of clusters in clustering X , C the number of clusters in

clustering Y , n i j the number of objects that are common to clusters X i and Y j , and
N the number of data items.

I (X, Y )
N M I (X, Y ) = √ . (4)
H (X )H (Y )

where X and Y are two clusterings, H (X ) the entropy of X , H (Y ) the entropy of Y ,

and I (X, Y ) is the Mutual Information between them.

I (X, Y ) − E{I (X, Y )}

AM Imax (X, Y ) = (5)
max{H (X ), H (Y )} − E{I (X, Y )}

where E{I (X, Y )} is the expected mutual information of X and Y , (Eq. 6) is the
expected mutual information.

min(ai ,b j )

R 
C  ni j N .n i j
E{I (X, Y )} = log( )× Q (6)
i=1 j=1 n i j =max(ai +b j −N ,0)
N ai b j

where
ai !b j !(N − ai )!(N − b j )!
Q= (7)
N !n i j !(ai − n i j )!(b j − n i j )!(N − ai − b j + n i j )!

Normalized Mutual Information index is an important entropy measure in infor-

mation theory. It calculates the similarity between two clusterings. It was used to
evaluate a multi-objective and evolutionary clustering method for dynamic networks.
It measured the performance of the proposed evolutionary clustering method on a
dynamic network of a fixed number of communities, and then on a dynamic network
of a variable number of communities [56]. AMI was also used to evaluate a clustering
stability-based evolutionary K-Means [77]. The weakness of NMI and AMI is that
they are not adjusted against chance.
30 I. Aljarah et al.

2.1.4 V-measure

V-measure [154] is an entropy-based external index that measures how homogeneity

and completeness are fulfilled. Homogeneity means that each cluster involves only
the data points that are members of a single class, and completeness means that all
data points of a given class are assigned to the same cluster. Homogeneity is defined
by Eq. 8. 
1 if H (C, K ) = 0
h= (8)
1 − HH(C|K
(C)
)
otherwise

where H (C|K ) is the conditional entropy of the class distribution given the proposed
clustering, and H (C) is the entropy of the classes. Equation 9 is the conditional
entropy, and Eq. 10 is the entropy of the classes.

|K | 
 |C|
ack ack
H (C|K ) = − log |C| (9)
k=1 c=1
N c=1 aack

|C| |K |ack |K |ack


H (C) = − k=1
log k=1
(10)
c=1
n n

where N is the number of data points, C the set of classes, K the set of clusters, ai j
is the number of data points that are members of class ci , and elements of cluster k j .
Completeness is symmetrical to homogeneity, it is described in Eq. 11.

1 if H (K , C) = 0
c= H (K |C) (11)
1− H (K )
otherwise

where H (K |C) is the conditional entropy of the proposed cluster distribution given
the class of the component data points, and H (K ) is the entropy of clusters. Equation
12 is the conditional entropy and Eq. 13 is the entropy of clusters.

|C| 
 |K |
ack ack
H (K |C) = − log |K | (12)
c=1 k=1
N k=1 aack

|K | |C|ack |C|ack

H (K ) = − c=1
log c=1
(13)
k=1
n n

V-measure considers homogeneity and completeness, a higher value of this index

means a better performance. Hence, Eq. 14 is the formulation of V-measure.

(1 + β) ∗ h ∗ c
Vβ = (14)
(β ∗ h) + c
A Comprehensive Review of Evaluation and Fitness Measures … 31

V-measure was used in the evaluation of various simulation scenarios, one is the
usage of a probabilistic method for inferring intra-tumor evolutionary lineage trees
from somatic single nucleotide variants of single cells. Where V-measure was used to
evaluate the clustering performance [155]. This index provides accurate evaluations
and solves problems such as the dependence on clustering algorithm or dataset.

2.1.5 Fowlkes-Mallows Index

Fowlkes-Mallows (FM) index [58] is an external validity measure that evaluates the
similarity between two clusterings. The range of this index is between 0 and 1, while
a higher index value means a better clustering. It is defined as the geometric mean
of precision and recall. Equation 15 is Fowlkes-Mallows index.

TP
FM = √ (15)
(T P + F P)(T P + F N )

where T P is the number of True Positive, F P is the number of False Positive, and
F N is the number of False Negative.
In [73], the authors suggested a multi-objective differential evolution approach
for clustering and feature selection. Further, [166] applied an evolutionary clustering
approach for breast cancer prediction, which utilized various internal and external
clustering validations including FM index. FM index was used to evaluate a swarm
intelligence-based clustering method in [197]. Several indices were used as fitness
functions, where the results stated that using silhouette statistic achieved the best
results in most of the datasets.

2.1.6 Jaccard Index

Jaccard Index [80] (also known as Jaccard Similarity Coefficient) is an external

validation index that measures similarity between two sets of data. Jaccard distance
is complementary to this index and it measures dissimilarity. Equation 16 represents
the formula of Jaccard Index [48].

TP
JI = (16)
T P + FN + FP

where T P is the number of True Positive, F P is the number of False Positive, and
F N is the number of False Negative. Jaccard distance is calculated by subtracting
Jaccard coefficient from 1.
Jaccard index was used for overlapping correlation clustering. In [14], the authors
implemented a Biased Random-Key Genetic Algorithm (GA) method for over-
lapping correlation clustering. In which, Jaccard similarity coefficient and set-
intersection indicator utilized as performance metrics. Furthermore, [33] suggested
32 I. Aljarah et al.

a local update strategy-based multi-objective evolutionary algorithm for commu-

nity (cluster) detection. where the Jaccard index is implemented for measuring the
similarity of nodes.

2.1.7 Czekanowski-Dice Index

Czekanowski-Dice index is an external validity measure that takes into consideration

both precision and recall. It has several names; Ochiai index, Sørensen-Dice index,
F1 Score, Dice’s coefficient, and Dice similarity coefficient. This index was inde-
pendently developed by Czekanowski [42], Dice [49], and Sorensen [176]. Equation
17 defines Czekanowski-Dice index [48]. Where T P is the number of True Positive,
F P is the number of False Positive, and F N is the number of False Negative.

2×TP
C − Dice = (17)
2 × T P + FN + FP

Czekanowski-Dice index was used in image segmentation using enhanced parti-

cle swarm optimization. It evaluated the performance of the proposed method and
has been compared with related studies using different datasets. The mean of Dice
coefficients over 10 runs was used as the criteria for performance comparison [181].
The main strength of this measure is that it reflects a realistic performance; since it
considers both Precision and Recall. This index was also used to evaluate the per-
formance of a new evolutionary clustering algorithm that is an improvement of the
multi-objective clustering with automatic determination of the number of clusters
(MOCK) algorithm [111].

2.1.8 Purity

Purity [204] is an external validation index that calculates to what degree a cluster
contains data points of the same class. So, it finds the majority class in the cluster
and computes the proportion of data points that belong to that class. The larger value
of this index means better clustering. Equation 18 defines the formula of Purity [8].

1
k
Purit y = max(|L i ∩ C j |) (18)
n j=1 i

where C j contains all data instances assigned to cluster j, n is the number of data
instances, k is the number of generated clusters, L i is the true assignments of the
data instances in cluster i.
Purity index was also used to evaluate the performance of a new clustering
approach that is based on glowworm swarm optimization [8], and to assess a cluster-
ing algorithm using a novel grey wolf-inspired clustering approach [10]. Furthermore,
A Comprehensive Review of Evaluation and Fitness Measures … 33

authors in [57] designed a clustering strategy that integrates background informa-

tion. In the proposed approach, the purity was used for evaluating the objective of
the clustering algorithm, whereas the Davies-Bouldin index was used to measure the
quality of the final obtained clusters.

2.1.9 Entropy

Entropy [204] is an external validation index that evaluates the distribution of the
classes within each cluster. The smaller value of this index means less diversity of
classes in the same cluster which indicates that the clustering is better. Equation 19
shows the definition of Entropy [8]. Where C j contains all data instances assigned to
cluster j, n is the number of data instances, k is the number of generated clusters, and
E(C j ) is the individual entropy of cluster C j which is defined by Eq. 20. In which,
L i is the true assignments of the data instances in cluster i and q is the number of
actual clusters.
k
|C j |
Entr opy = E(C j ) (19)
j=1
n

1  |C j ∩ L i |
q
|C j ∩ L i |
E(C j ) = − log( ) (20)
log q i=1 |C j | |C j |

Entropy index was used to evaluate the clustering quality of a new clustering approach
that is based on glowworm swarm optimization [8]. Three different fitness functions
were compared with different datasets. Where the results indicated that the new
clustering approach is efficient in comparison to well-known clustering algorithms.
Moreover, the entropy was utilized in a novel locality-informed Grey Wolf clustering
approach by [10]. In which, the proposed approach used the entropy for evaluating the
clustering quality which showed merits over other well-known clustering algorithms.
Nonetheless, the entropy validation measure was also used in [5] for automatic clus-
tering based on an imperialist competitive algorithm. [16] suggested a GA approach
for clustering, which tested three different functions as objective functions. Where
the utilized objective functions are the Euclidean distance, the Mahalanobis distance,
and the entropy.

2.1.10 Hubert Γˆ Statistic

Hubert Γ index is a statistical measurement that finds the correlation between two
matrices of size (N × N ), which is proposed in [79]. A higher value of Hubert Γˆ
index indicates a higher correlation similarity of matrices. Hubert Γˆ index is defined
by Eq. 21, where X and Y are two matrices and M is (N (N − 1)/2).
34 I. Aljarah et al.

N −1 N
1  
Γˆ = X (i, j) · Y (i, j) (21)
M i=1 j=i+1

A modified version of it has been proposed in [78] for tackling the validation of
clustering problems. Where it mainly replaced the two matrices by a proximity matrix
of the dataset P and the matrix Q. Each element in the proximity matrix represents the
distance between the respective two data points of (i,j) indices. While each element
in the Q matrix presents the distance of the two centroids that the corresponding two
data points belong to. The modified version is shown in Eq. 22. Accordingly, a higher
value of the modified Hubert Γˆ index denotes a higher clustering compactness. In
other words, given two data points in different clusters and the distance between
them is relatively close to the distance of their centroids that means higher compact
clusters.
N −1 N
1  
ˆ
Γ = P(i, j) · Q(i, j) (22)
M i=1 j=i+1

Several studies have used the Hubert Γˆ index or its modified version for clustering
validation. In [193], authors suggested a GA-based clustering approach for feature
selection, in which, authors utilized Hubert Γˆ statistics as fitness function. More-
over, [137] proposed a clustering approach based on Genetic Programming for the
determination of level-of-service in urban street, in which, Hubert statistic with other
metrics were used for the validation.

2.2 Internal Measures

Internal validation measures are used when the ground truth labels are not known.
These indices depend only on the information in the data. The following subsections
present the internal indices used with evolutionary clustering.

2.2.1 Calinski-Harabasz Index

Calinski-Harabasz (CH) index [26] is also known as the Variance Ratio Criterion, is
an internal evaluation measure that is based on the Between Cluster Scatter Matrix
(BCSM) and the Within Cluster Scatter Matrix (WCSM). A larger BCSM value
results in a better clustering, while a larger WCSM means a worse one. BCSM and
WCSM are defined by Eqs. (23 and 24), respectively [163]. Where n is the total
number of data points, k is the number of clusters, z i is the centroid of the current
cluster, z tot is the centroid of all the data points, x is a data point belongs to cluster
ci, and d(x, z i ) is the Euclidean distance between x and z i . Hence, depending on
BCSM and WCSM, Eq. 25 presents the mathematical formula of Calinski-Harabasz
index.
A Comprehensive Review of Evaluation and Fitness Measures … 35


k
BC S M = n i .d(z i , z tot )2 (23)
i=1


k 
WCSM = d(x, z i )2 (24)
i=1 x∈ci

BC S M n − k
C Hk = . (25)
k − 1 WCSM

Calinski-Harabasz index was used in unsupervised image clustering for feature

selection using differential evolution. It evaluated clustering in two experiments [68].
Results indicated that the differential evolution feature selection had better results
than the other used methods. This index is easy to compute, and its value is larger
when clusters are compact and well-separated. Additionally, this index was used in
clustering using Flower Pollination Algorithm (FPA) as an optimization criterion.
The results indicated that FPA-based clustering method had a better performance than
the other used methods [106]. Calinski-Harabasz index was also used in achieving
natural clustering by validating results of iterative evolutionary clustering Approach
[131], where a multi-objective genetic algorithm was used to decide on the number
of clusters.

2.2.2 P B M Index (I Index)

P B M Index, which is also known by I Index, is an internal validity measure that

evaluates the combination of compactness and separation. It has three key factors: k1 ,
E1
Ek
, and D K [133]. The larger value of index P B M means better clustering. Equation
26 illustrate the formula of P B M Index [122]. Whereas, Eq. 27 defines the formula of
intra-cluster distance E K , and Eq. 28 defines the formula of the inter-cluster distance
DK .  γ
1 E1
PBM = × × DK (26)
K EK


K 
n
EK = u k j D(z k , x j ) (27)
k=1 j=1

K
D K = max{D(z i , z j )} (28)
i, j=1

where E1 is a constant, u ki is the membership degree of the ith data point to the
kth cluster, and D(z k , xi ) is the distance of the ith data point xi from the kth cluster
center z k .
36 I. Aljarah et al.

I Index was used in modified differential evolution based fuzzy clustering for pixel
classification in remote sensing imagery. It evaluated the performance of the proposed
method which is a Modified Differential Evolution based Fuzzy Clustering (MoD-
EFC) and compared its performance with five other clustering algorithms. Those
algorithms were Differential Evolution based Fuzzy Clustering (DEFC), genetic
algorithm based fuzzy clustering (GAFC), simulated annealing based fuzzy clus-
tering (SAFC), Fuzzy C-means (FCM), and Average Linkage (AL). The results
indicated that the new method (MoDEFC) outperformed the other methods [115].

2.2.3 Dunn Index

Dunn Index is an internal evaluation measure that detects the clusters that are com-
pact and well-separated. It depends on the minimum inter-cluster distance and the
maximum cluster size [53]. A larger value of this index means a better clustering.
Equation 29 defines the formula of Dunn Index [122]. In literature, different improve-
ments have been conducted on Dunn index due to its sensitivity for noise. Kim et
al. in [24, 89] proposed several generalized Dunn indexes such as gD31↑, gD41↑,
gD51↑, and others.
 
δ(Ci , C j )
D N = min min (29)
1≤i≤K 1≤ j≤K , j =i max1≤k≤K { (Ck )}

where δ(Ci , C j ) is the distance between clusters Ci and C j . Equation 30 defines this
distance.
δ(Ci , C j ) = min {D(xi , x j )} (30)
xi ∈Ci ,x j ∈C j

And (Ci ) is the diameter of cluster Ci , where the diameter is shown by Eq. 31.

(Ci ) = max {D(xi , xk )} (31)

xi ,xk ∈Ci

Dunn index was used in automatic scientific document clustering to evaluate a new
multi-objective document clustering approach that is based on differential evolution
(DE) and self-organizing map (SOM). The proposed method was compared to several
clustering algorithms using this Index. Those algorithms were a multi-objective DE-
based clustering approach without using SOM-based operators (MODoc-clust), and
an archived multi-objective simulated annealing algorithm (AMOSA). The results
indicated that the proposed approach had a better performance [161]. Dunn index
is one of the most famous indices for clustering evaluation, but it requires high
computational power when we have a high-dimensional data and when the number
of clusters is large. This index was also used to select the best number of clusters in
an iterative evolutionary clustering method to solve the scalability issue [131]. Dunn
index was used in a comparison study of validity indices [197]. Swarm intelligence-
based algorithms were evaluated when using the Dunn index and other indices as
A Comprehensive Review of Evaluation and Fitness Measures … 37

fitness functions. Furthermore, in [107], authors suggested an improved version of

the Dunn index to handle the problem of clustering large data in a reasonable time,
which is called the Big Data Dunn (BD-Dunn) index.

2.2.4 Silhouette Coefficient

Silhouette Coefficient [156] is an internal validation index that takes a data point
and checks how much similar it is to its cluster in comparison to the other clusters.
In other words, it shows how strongly the data points in the cluster are gathered. If
most of the data points have a large value; this means a good clustering. Silhouette
Coefficient is determined by Eq. 32. Where (a) is the mean distance between a point
and the points in that cluster, and (b) is the mean distance between a point and the
points in the next nearest cluster.

b(i) − a(i)
s(i) = (32)
max{a(i), b(i)}

Silhouette Coefficient was used to evaluate a new evolutionary clustering algo-

rithm named adaptive sequential k-means (ASK). The new algorithm was compared
to c-means and k-means, where the results indicated that it accomplished better than
both [139]. The value of this index is better when the clusters are compact and
well-separated. This index was also used to evaluate the performance of a proposed
evolutionary clustering algorithm [111]. In [66], the Silhouette index was utilized for
fitness evaluation of Particle Swarm Optimization (PSO) clustering method. Remark-
ably, [107] suggested a modified version of Silhouette index for approaching the
problem of validating the clustering of big data.

2.2.5 Davies-Bouldin Index

Davies-Bouldin index is an internal evaluation measure that was introduced in 1979

[47]. It uses the average similarity between a cluster and its most similar one to reflect
how good the clustering is. The smaller DB index value means better clustering. In
Eq. 33, the similarity has been formed [122].

si + s j
Ri = max (33)
j, j =i di j

where si is the average distance between each point of the cluster and the centroid
of that cluster, and di j is the distance between-cluster centroids i and j. Equation 34
defines Davies-Bouldin coefficient.
38 I. Aljarah et al.

1 
N
DB = Ri (34)
N i=1

Davies-Bouldin Index was used in genetic clustering for automatic image clas-
sification as a measure of the validity of the clusters [20]. Further, it was used to
evaluate a proposed multi-objective document clustering approach [161]. In [130],
the authors implemented an evolutionary clustering approach based on GA for iden-
tifying traffic similarity. Where the k-means algorithm is used with Davies-Bouldin
index for fitness evaluation. Moreover, [66] used Davies-Bouldin index for assessing
the fitness of the PSO algorithm for clustering.

2.2.6 Xie-Beni Index

Xie-Beni index [194] is an internal validity measure that evaluates the compactness
and separation of clusters. It is defined as the ratio between compactness σ and the
minimum separation sep of the clusters. A smaller value of this index means better
clustering. Equation 35 and 36 define the compactness and separation. Equation 37
represents Xie-Beni index [122].


K 
n
σ= u 2ki D 2 (z k , xi ) (35)
k=1 i=1

sep = min{D 2 (z k , xi )} (36)

k =l

σ
XB = (37)
n × sep

where u ki is the membership degree of the ith data point to the kth cluster, and
D(z k , xi ) is the distance of the ith data point xi from the kth cluster center z k .
Xie-Beni index was used to evaluate a modified differential evolution based fuzzy
clustering algorithm [115]. Even that this index is simple to implement, but its weak-
ness is that separation decreases when the number of clusters are close to the number
of data points. Further, its used as a validity index in an automatic image pixel clus-
tering with an improved differential evolution approach [45]. Another application
of this index is that it was used to evaluate an improved differential evolution based
clustering algorithm in [46].

2.2.7 Bezdek’s Partition Coefficient

Partition Coefficient [23], it is an internal validity measure that describes the global
cluster variance and how much overlap is among clusters. The range of values for
A Comprehensive Review of Evaluation and Fitness Measures … 39

this index is ∈ [1/c, 1], where c is the cluster number. The best result of this validity
index is at its maximum value. Partition Coefficient is explained in Eq. 38 [169].
Where Ui j is the membership degree of the point x j in cluster I, and c is the cluster
number.
1  2
c N
PC(c) = U (38)
N i=1 j=1 i j

Partition Coefficient was used to evaluate a novel evolutionary fuzzy clustering

algorithm to solve the fuzzy C-means optimization problem. It was evaluated on three
datasets, and the results indicated that it had a better performance than the classical
fuzzy C-means algorithm [135]. Partition coefficient strength is that it identifies the
right number of clusters in most of the datasets, but it decreases as the cluster number
increases. Another drawback is that it is sensitive to the fuzzifier.

2.2.8 Average Between Group Sum of Squares (ABGSS)

ABGSS index is a measure that describes how well-separated are the clusters [90].
It represents the average distance between-clusters centroids and the centroid of all
the data points. Equation 39 defines the formula of ABGSS [122].
K
g=1 n g · D 2 (z g , z̄)
ABG SS = (39)
K
where n g is the number of points in cluster g and D(z g , z̄) is the distance between
the centroid of the cluster g and the center of the dataset z̄. Moreover, another index
that is related to ABGSS index is the Average Within Group Sum of Squares index
(AWGSS) [90]. This index describes how compact the clusters are instead of focusing
on clusters separation. Equation 40 defines AWGSS index.
n g

K
D 2 (z g , xi )
AW G SS = i=1
(40)
g=1
ng

where D(z g , xi ) is the distance between the centroid of the cluster g and a data point
xi .
However, similar previously established validity indices are the T race_W and
T raceCovW [118]. Mainly, T race_W computes the within-cluster sum of squares,
while T raceCovW is the covariance of the within-cluster sum of squares. Both
T race_W and T raceCovW have been used in [132], where a clustering approach
depending on a multi-objective GA algorithm was proposed for skylines detection
in databases.
40 I. Aljarah et al.

ABGSS index was used in a newly proposed multi-objective genetic algorithm for
clustering. Where used as a fitness function [90]. The new algorithm was compared
to the k-means algorithm, and the results indicated that it had a better performance.

2.2.9 Intra-cluster Entropy

Intra-cluster Entropy [151] is an internal validation index that computes the average
purity of clusters, while it does not rely on class labels. This index is different from
Sum of Squared Error measure. The larger value of this index is better, Eq. 41 presents
the Intra-cluster Entropy [122].


K
1
H= [(1 − H (Ci ).g(z i ))] k (41)
i=1

where
H (Ci ) = − [g(z i )log2 g(z i ) + (1 − g(z i ))log2 (1 − g(z i ))] (42)

And g(z i ) is the average similarity between a cluster center z i and the data points
belonging to cluster i. Equation 43 is the formula of this average similarity.

1
n
C O(z i , xi )
g(z i ) = 0.5 + (43)
n j=1 2

where C O is the cosine distance, defined by Eq. 44. Where d is the number of
features. d
xik .x jk
C O(xi , x j ) =  k=1  (44)
d d
x 2
k=1 ik . x 2
k=1 jk

Intra-cluster Entropy was used to evaluate the performance of a multi-objective

data clustering algorithm [151]. It was also used as a fitness function in an evolution-
ary multi-objective clustering for overlapping clusters detection [150]. The results
indicated that the proposed method successfully identified the overlapping clusters.
This index has a low computational cost and it is insensitive to outliers; which makes
it an attractive measure to utilize.

2.2.10 Overall Cluster Deviation (Cluster Connectedness)

Overall Cluster Deviation, which is also known as Cluster Connectedness, is an inter-

nal validation index that measures the compactness of clusters [74]. It is calculated as
the summed distances between the data points and their clusters center. The smaller
value of this index means a better clustering. Equation 45 defines the foundation of
A Comprehensive Review of Evaluation and Fitness Measures … 41

Overall Cluster Deviation [122].

 
Dev(C) = D(z k , xi ) (45)
Ck ∈C xi ∈Ck

where D(z k , xi ) is the distance of the ith data point xi from the kth cluster center z k .
Overall Cluster Deviation was used to evaluate the performance of an evolutionary
multi-objective data clustering algorithm [151]. The experimental results indicated
that the proposed algorithm optimizes both Intra-cluster entropy and Inter-cluster
distance.

2.2.11 Edge Index

Edge Index is an index that was introduced by Shirakawa and Nagao in 2009 [170].
This index calculates the summed distances between clusters. A smaller value of this
index means a better clustering and higher separation. It presented by Eq. 46 [122].
j
Where ξi, j = D(xi, x j) if  Ck : i ∈ Ck ∈ Ck , and ξi, j = 0 otherwise, F is the
set of N nearest neighbors of ith point, and N is a user-defined integer number.


n 
Edge(C) = − ξi, j (46)
i=1 j∈F

Edge Index was used in a proposed multi-objective evolutionary image segmenta-

tion algorithm; as an objective to be optimized [170]. Edge index is a very important
factor in image segmentation. The results indicated that this algorithm obtains var-
ious good solutions; hence, could be useful if applied to other kinds of images like
medical images.

2.2.12 Cluster Separation

Cluster Separation [151], which is also known as Inter-cluster Distance, is an index

that evaluates the average distance between-cluster centers. The larger value of this
index means a better clustering, as shown by Eq. 47 [122]. Where D(z i , z j ) is the
distance between the centroid of cluster i and the centroid of cluster j.

2 K K
Sep(C) = D 2 (z i , z j ) (47)
K (K − 1) i=1 j=1, j =i

Cluster Separation was used to evaluate a novel multi-objective evolutionary clus-

tering algorithm [151]. This index has a low computational cost and it is a reliable
index to implement since it is insensitive to outliers.
42 I. Aljarah et al.

2.2.13 Bayesian Information Criterion

Bayesian Information Criterion (BIC) [167], which is also known as Schwarz Infor-
mation Criterion. It is an internal validity index that was first introduced by Schwarz,
in 1978, to solve the problem of model selection. Therefore, this index is based on
the likelihood function and a larger value of it means a better clustering. Equation
48 is the BIC index [32].

K
1 1
BIC = {− n i log |i |} − N k(d + d(d + 1)) (48)
i=1
2 2

where K is the number of clusters, n i is the number of data points in cluster i, i is

the sample covariance matrix, N is the number of points in the dataset and d is the
number of dimensions. So d + 21 d(d + 1) represents the number of parameters for
each cluster.
In [39], a new evolutionary clustering method was proposed. Where the new
method relied on cuckoo search and k-means for optimizing the search on the web.
The developed algorithm utilized for document clustering, while integrated the BIC
measure for evaluating and optimizing the quality of results. Moreover, [38] con-
ducted web documents clustering based on genetic programming and a modified
BIC technique for fitness optimization. In [164], authors designed a method used in
the context of structural flaws detection of bridges. Hence, the methodology is a GA-
based method used to search for the optimal hyper-parameters for having the optimal
number of clusters. Where BIC is used for assessing the quality of GA solutions.

2.2.14 Sym-Index

Sym-Index [159], which is also known as Symmetry Distance-based index, is an

internal validation index that was introduced in 2007. It is based on the idea of finding
different-size symmetric clusters, as illustrated by Eq. 49. Where K is the number
of clusters and D K is the maximum Euclidean distance between two centroids. Dk
is defined by Eq. 50.
1 1
Sym(K ) = × × DK (49)
K εk

K
D K = max ||c¯i − c¯j || (50)
i, j=1

where c¯i is the centroid of cluster i, and εk is defined in Eq. 51.

knear

K 
ii=1 dii
εk = de (x̄ j , c̄i ). (51)
i=1 x̄ j ∈u i
knear
A Comprehensive Review of Evaluation and Fitness Measures … 43

where knear is the k nearest neighbor and de (x̄ j , c̄i ) is the Euclidean distance
between the point x̄ and the centroid c¯i .
Sym-Index was used in an evolutionary multi-objective automatic clustering
approach as an objective function [206]. Two algorithms were proposed and com-
pared with other well-known algorithms. The results indicated that the new algo-
rithms outperformed the other algorithms in terms of computational cost, time, and
the ability to divide the data into meaningful clusters without knowing the cluster
numbers. This index was also used as an objective function in a multi-objective fuzzy
clustering approach based on tissue-like membrane systems. The approach was com-
pared with other multi-objective clustering techniques, and the results indicated that
it had a better performance than the others [138]. In [160], authors proposed an evo-
lutionary clustering approach for symmetry-based automatic clustering. In which,
the proposed approach utilized PSO and DE algorithms, in which, the Sym-Index
was the objective function for the evolutionary algorithms.

2.2.15 CS Index

CS Index [35], which is also known as Chou, Su, and Lai Index. It is an internal
validity index that is based on the ratio of clusters diameters to the between-clusters
separation. The smaller value of this index means better clustering, as represented in
Eq. 52 [28]. Where d is a distance function, K is the number of clusters, and |Ck | is
the number of data points in cluster Ck .
K 1 
xi nCk max x j ∈Ck d(x i , x j )
1
K k=1 |Ck |
CS = 1 K
(52)
K k=1 mink,k =k  d(ck , ck  )

CS Index was used to evaluate a proposed differential evolution clustering algo-

rithm for determining the number of clusters relying on k-Means [173]. The new
method was compared with other k-means related studies using six UCI datasets.
Accordingly, the results indicated that the proposed method had the smallest value of
CS; which means better clustering. This index was also used in automatic clustering
using metaheuristic algorithms for content-based image retrieval [17]. The results
indicated that clustering using optimization algorithms had better results than clus-
tering using only main image retrieval algorithms. Moreover, CS Index was used as
an objective function in [83]. In which, authors used a Differential Memetic algo-
rithm for clustering that adopted CS index as a fitness function. Where the results
were better than other algorithms in terms of CS, Davies-Bouldin, and accuracy.

2.2.16 C-Index

C-Index [43] is an internal validity index that was introduced in 1970. This index is
based on the ratio between the minimum and the maximum distances between data
44 I. Aljarah et al.

points (Eq. 53) [48].

SW − Smin
C − I ndex = (53)
Smax − Smin

where SW is the sum of the within-cluster distances. Smin is the sum of the N W
smallest distances between all the pairs of data points, and Smax is the sum of the
N W largest distances between all the pairs of data points. Equation 54 shows the
formulation of N W .
K
n k (n k − 1)
NW = (54)
k=1
2

In literature, different research studies have used C-Index within the evolutionary
clustering. For instance, [40] developed two genetic-based clustering algorithms, in
which, the Calinski and Harabasz, C-index, and trace-W indices were compared for
having the optimal number of clusters. Moreover, [105] applied a multi-objective
genetic algorithm for clustering, in which various clustering validation indices were
utilized including the C-Index. While the objectives of the proposed strategy were
to minimize the number of clusters and the within-cluster variation. While in [158],
a canonical GA algorithm was suggested for clustering. Where the canonical GA
algorithm was used for optimizing the (C) validity index as an objective function for
the sake of obtaining the optimal clusters.

2.2.17 Ball Hall Index

Ball Hall (BH) index [19] is an internal validation index that measures the average
distance between each data point and the centroid of the cluster it belongs to. The
smaller value of this index means better clustering. Equation 55 demonstrates the
formula of Ball Hall index, where m i is the centroid of the cluster ci [145].

1 
K
BH = ||x − m i ||. (55)
N i=1 x∈C
i

Ball Hall index was used in particle swarm clustering for fitness evaluation with
computational centroids [145]. This index was compared with the other 13 indices
as fitness functions. Where it achieved better performance when the clustering task
became more complicated. This index was also used to choose the number of clusters
[131]. Furthermore, authors in [157] proposed a clustering method based on the GA
algorithm for grouping the graphical data. Where the ball hall index was used as a
fitness function for GA algorithm.
A Comprehensive Review of Evaluation and Fitness Measures … 45

2.2.18 BetaCV Measure

BetaCV measure is an indicator for the coefficient of variation. It finds the proportion
of the average intra-cluster distance to the average inter-cluster distance. The smaller
the BetaCV value the better the clustering quality. Given K clusters, W (x, y) is the
sum of all weights of the edges located in between x and y, Nout is the number of
unique edges between clusters and Nin is the number of unique edges within the
cluster. BetaCV is defined by Eq. 56.
k
Nout i=1 W (Ci , Ci )
BetaC V = k (56)
i=1 W (C i , C i )
Nin

In [127], an interpretation of the PSO algorithm for evolutionary clustering was

maintained. Where the BetaCV and Dunn indices were adopted for evaluating the
fitness of the proposed algorithm. The results exhibited merits against the other clus-
tering algorithms. Further, [126] performed a fuzzy rule interpolation based on the
GA clustering approach. In their proposed methodology several clustering validation
measures were deployed for evaluating the quality of GA individuals. The experi-
mented fitness functions were Dunn, Davies-Bouldin, BetaCV, and Ball Hall indexes.
The obtained outcomes showed that Dunn and Davies-Bouldin achieved the most
accurate results.

2.2.19 Other Evolutionary Clustering Quality Indices

Among all mentioned indices, yet there are more used in the context of evolutionary
clustering even that they are not widely used. For instance, Fisher’s discriminant
ratio is a measure of separability. A higher value of it indicates better separability.
The mathematical model of Fisher’s discriminant ratio is shown in Eq. 57, where k
is the number of clusters, c is the centroid and σ is the cluster’s variance.

k  k
||ci − c j ||
Fisher  s Ratio = (57)
i=1 j=1
σ i2 − σ 2j

Sinha and Jana [174] investigated the use of Fisher’s discriminant ratio for validating
the results of a GA-based clustering approach, which showed efficient results in
comparison with other clustering algorithms.
Another measure is the Hartigan index. Hartigan index is a well-regarded measure
for searching the optimal number of clusters that are proposed in [75]. It stands on
the within-cluster squared-sum of the distance between the points and the centroid.
In Eq. 58, the (n − k − 1) is the correction index, k is the clusters, n is the number
of items in the data, and W (k) is a matrix of the sum of squared distances.
46 I. Aljarah et al.

W (k)
H ar tigan = − 1 · (n − k − 1) (58)
W (k + 1)

In [30], a fuzzy clustering approach was proposed relying on PSO algorithm. In

which, several clustering validity measures were employed to find the optimal number
of clusters including the Hartigan index. According to their reported results, Hartigan
index achieved promising performance results.
The S_Dbw validity index not just quantify the separation and compactness of
clusters, but also the density. It was proposed in [69]. In which, both the separation and
compactness were represented by the inter and intra clusters variances, respectively.
While the density is determined by the number of neighbors within a hyper-sphere.
A lower value of S_Dbw index is preferable for better clustering quality. Equation
59 shows that the S_Dbw index is a summation of the scattering and density. Where
the scattering and density are formulated in (Eq. 60, Eq. 61), respectively. Where
σ(vi ) is the variance of a cluster, σ(x) is the variance of the dataset, u i j is the central
point of the link that connects the centroids vi and v j .

S_Dbw = scattering + densit y (59)

1  ||σ(vi )||
k
scattering = (60)
k i=1 ||σ(x)||
⎛ ⎞
1 k k
densit y(u )
densit y = ⎝ ij ⎠ (61)
k(k − 1) i=1 j=1 Max (densit y(vi ), densit y(v j ))

Authors in [206] implemented an automatic clustering approach based on multi-

objective evolutionary algorithm. Where different clustering evaluation measures
were utilized including the S_Dbw index.
The SD validity index [70] depends on the mean scattering and the separation of
clusters, as represented in Eq. 62. The scattering is as represented in Eq. 60, while
the distance is the distance between the centroids of clusters as given by Eq. 63. A
lower value of the index indicates better quality.

S D validit y = scattering + distance (62)

⎛ ⎞−1
Max(||v j − vi ||)  ⎝ 
k k
distance = ||vt − vz ||⎠ (63)
Min(||v j − vi ||) i=1 z=1,z =t

Even that SD validity index is not that widely used, but [105] showed an implementa-
tion of the multi-objective GA algorithm for clustering. In which, SD validity index
was utilized for examining the results alongside other measures.
A Comprehensive Review of Evaluation and Fitness Measures … 47

Furthermore, the R-squared index is often used when looking for determining
the number of clusters within the hierarchical clustering approaches. R-squared is
computed by finding the proportion of the sum of squares between clusters (SSB)
to the sum of squares of the whole dataset (SST). Where the SST is the total sum
of the SSB and the sum of squares
n within clusters (SSW) (Eq. 64). In which, the
sum of squares is given by i=1 (xi − x)2 . The value of R-squared is within the
range of [0, 1]; a value of 0 denotes that there is no difference between the clusters
[188]. R-squared has been utilized in [166] for measuring the quality of a clustering
strategy. In which, [166] presented a hybrid combination of Whale optimization and
Moth Flame optimization for feature selection-based on clustering. The proposed
approach is maintained in the situation of a breast cancer diagnosis.
SS B
R − squar ed = (64)
SS B + SSW

The COP-index is a relatively newly proposed index [67] that is designed to

find the best cluster of hierarchical clustering. COP-index depends on the intra-
cluster variance and inter-cluster variance (Eq. 65). P Y is apartition Y , X is
the dataset, d(x,y) is the Euclidean distance, intra(C)=1/|C| x∈C d(x, C), and
/ Max x j ∈C d(x i , x j ).
inter(C)= Min xi ∈C

1  intra(C)
C O P(P Y , X ) = |C| (65)
|Y | Y
inter (C)
C∈P

In [123], COP-index was used and compared with different clustering validation
measures to assess the clustering quality of greedy-based evolutionary algorithm for
clustering.
Nonetheless, Wint-index is a weighted inter and intra measurement developed
in [179]. It involves the calculation of within-cluster distance (intra(Ci )) and the
between-cluster
 distance (inter(Ci , C j )) (Eq. 66). The intra(Ci ) =2/n i (n i − 1)
x,y∈Ci d(x, y), inter(C i , C j )=1/n n
i j x∈Ci ,y∈C j d(x, y). Given C is a cluster, k
the number of clusters, and n is the data items.
 k k 
2k i=1
ni
n−n i j=1, j =i n j · inter (Ci , C j )
W int (k) = (1 − ) 1 − k (66)
n i=1 n i · intra(Ci )

Fukuyama-Sugeno (FS) index quantifies the compactness and separation of clus-

ters, it was created in 1989, in [62]. In Eq. 67, the first expression denotes compact-
ness, while the second is an indicator for separation. Given that x is the data, c is
the centroid, u is the degree of membership, and m is a controlling parameter of the
fuzziness of the membership.


n 
n 
n 
n
FS = u imj ||xi − c j || − u imj ||c j − c|| (67)
i=1 j=1 i=1 j=1
48 I. Aljarah et al.

Chen and Ludwig [31] implemented Wint index and Fukuyama-Sugeno index for
the evaluation of a PSO-based clustering approach. As well as, [51] utilized Wint
and Fukuyama-Sugeno indices for the validation of clustering algorithm based on
Artificial Bee Colony (ABC) and PSO.
Krzanowski-Lai (KL) index [92] stands on the squared distance of the points from
their corresponding centroid. At the optimal value of this index the optimal number of
clusters is reached. Equations (68 and 69) illustrates the formulation Krzanowski-Lai
index, in which m is the number of features.

di f f (k)
K L − index(k) = | | (68)
di f f (k + 1)


k−1  
k 
di f f (k) = (k − 1)2/m d 2 (x, xi ) − k 2/m d 2 (x, xi ) (69)
i=1 x∈Ci i=1 x∈Ci

In [121], a fuzzy-GA clustering algorithm was developed for identifying the level-
of-service of urban streets, where Krzanowski-lai index, c-index, Hargitan, weighted
inter-intra measure, and R-squared were all maintained.
Ray-Turi index [148] was suggested to find the optimal number of clusters. Ray-
Turi index is defined by Eq. 70, where the numerator presents the intra distance.
k−1 
1
n j=1 x∈Ck d(x, ck )
Ray − T uri = (70)
Min i = j d(ci , c j )

In [100], a k-means based PSO algorithm was developed for the detection of
discontinuities within the rock mass avoiding such problems in the engineering sector.
In the proposed approach the sum of all distances over all clusters was the fitness
function, while the Ray-Turi index was the clustering validation index. Moreover,
[63] used the inter-cluster distance, intra-cluster distance, and Ray-Turi index for
validating the clustering results of the dynamic DE clustering algorithm.
Fuzzy Hypervolume [185] (FHV) is a validity index for fuzzy clustering. A smaller
FHV value denotes a better quality of non-overlapping clusters. Interestingly, FHV
does not depend on the sizes or shapes of the clusters. Primarily, the FHV index
relies on the sum of the squared-root of the determinant of the covariance matrices
(Eq. 71). In Eq. 71, det is the determinant and Fi is the fuzzy covariance matrix.
Few research studies have used fuzzy hypervolume for evolutionary clustering
validation. However, in [36], an improved fuzzy c-mean algorithm was developed
based on PSO algorithm and fuzzy c-means++. In which various validity measures
were utilized for assessing the quality of the proposed method including the FHV.


k 
FHV = det (Fi ) (71)
i=1
A Comprehensive Review of Evaluation and Fitness Measures … 49

Furthermore, the Partition Coefficient and Exponential Separation (PCAES) index

was proposed by Wu and Yang [192]. PCAES index measures the clustering quality
by calculating a normalized fuzzy partition coefficient and an exponential separation

measurement. PCAES is explained by Eq. 72. In which, u M is min 1≤ i ≤c nj=1 u i2j ,
c 
βT = ( i=1 ||z i − z||2 )/c, and z = nj=1 (y j /n).


c 
n
u i2j 
c
PC AE S = − ex p(−min k =i ||z i − z k ||2 /βT ) (72)
i=1 j=1
uM k=1

PCAES was used with evolutionary algorithms. In [31], a fuzzy-based PSO clus-
tering algorithm was created for finding the optimal number of clusters, which
involved PCAES for evaluation. Additionally, it was employed in a fuzzy-based
chaotic PSO algorithm for clustering in [99].
Ratkowsky-Lance (RL) index [147] is a clustering quality measure that seeks
for the best number of clusters. In literature, RL is identified as shown in Eq. 73,
given that B is the sum of squares between the clusters for each data point, T is
the total sum of squares of each data point. Although that RL was not widely used
with evolutionary clustering algorithms, but [132] deployed it for the evaluation of
a multi-objective GA clustering approach.

average( TB )
RL = √ (73)
k

Moreover, Kwon’s index [97] is a fuzzy clustering validation index. It was essentially
proposed to alleviate the monotonically decreasing tendency with the increasing

number of clusters. Kwon et al. defined the index as in Eq. 74, where v = n1 nj=1 x j .
Kwon’s index has been utilized for assessing two-level multi-objective GA clustering
approach in [4].
n c c
 j=1 i=1 u i2j ||x j − vi ||2 + 1
c i=1 ||vi − v||2
K won s index = (74)
min i =k ||vi − vk ||2

Roughly speaking, clustering validation indices are vast, some are broadly-utilized
and experimented in the context of evolutionary clustering, while other are less
popular. Some of the rarely used are Marriot [109], Rubin [61], Cubic Clustering
Criteria [165], Duda [52], Scott [168], and Friedman index [61]. Equations (75–80)
present the conventional mathematical formulation for those indices, given that W
is the within-cluster matrix, B is the between-clusters matrix, T is the total sum of
squares matrix, R 2 = 1 − (T race(W )/T race(T )), p is the number of attributes,
and SS E W C2 is the sum of squared errors within-cluster when the data is grouped
in two clusters, while SS E 1 is the sum of squared errors when the data represented
by one cluster.
50 I. Aljarah et al.

In [132], Marriot, Rubin, Duda, Scott, and Friedman were used for the validation of
multi-objective GA algorithm for clustering. While [110] used the Cubic Clustering
Criteria (CCC) for evaluating a clustering method that depends on a constructive GA
algorithm for feature selection.

Marriot = k 2 W | (75)

|T |
Rubin = (76)
|W |

  np
1 − E(R 2 ) 2
CCC = ln · (77)
1 − R2 (0.001 + E(R 2 ))1.2

SS E W C2
Duda = (78)
SS E 1

|T |
Scott = nlog( ) (79)
|W |

Friedman = T race(W −1 B) (80)

Table 1 presents a summary of clustering validation indices used within the era of
evolutionary computation. It shows the indices names, their respective applications,
and their references.

2.3 Other Non-evolutionary CVIs

Over the past decades, massive research studies have been conducted particularly for
improving the evaluation of clustering. In previous sections and according to Fig. 2
in P. 27, an explanation of clustering quality measures has been conducted showing
the extent of the integration with evolutionary algorithms.
Nowadays, plenty of clustering indices have been designed, which return to the
early of the 60s. Yet, the evaluation of clustering is still an active research area.
Figure 2 shows several indices that are not used with evolutionary clustering. Some
of them were designed, especially for fuzzy clustering, whereas some others focus
on optimizing the quality while maintaining the optimal number of clusters. To the
best of our knowledge, Table 2 represents the clustering validation indices that are
not used in the field of evolutionary clustering.
A Comprehensive Review of Evaluation and Fitness Measures … 51

Table 1 A synopsis of evolutionary clustering quality measures; where they utilized and their
related articles
Index Type Application Related papers
CH index Int Data clustering, feature selection, determining number of [68, 106, 131]
clusters
PBM index Int Fuzzy image clustering [115]
Dunn index Int Document clustering, determining the number of clusters, [93, 131, 161, 197]
feature selection
Silhouette Int Acoustic emission clustering, multi-objective optimization, [66, 93, 111, 139, 187]
feature selection, medical data clustering
Davies-Bouldin Int Document clustering, traffic monitoring, feature selection, [20, 37, 66, 93, 161]
index image clustering
Xie-Beni index Int Remote sensing imagery and fuzzy clustering, image cluster- [45, 46, 115]
ing
Bezdek’s Int Fuzzy clustering [135]
P.Coefficient
ABGSS Int Multi-objective optimization [90]
Intra-cluster entropy Int Multi-objective clustering [150, 151]
Cluster deviation Int Multi-objective clustering [112, 151]
Edge index Int Image segmentation [170]
Cluster separation Int Multi-objective clustering [102, 151]
BIC Int Web search optimization, bridges flaws detection [38, 39, 164]
Sym-index Int Evolutionary multi-objective automatic clustering and fuzzy [138, 160, 206]
clustering
CS-index Int Determining the number of clusters, image retrieval [17, 83, 173]
C-index Int Data clustering, determining the number of clusters [40, 105, 158, 187]
Ball Hall index Int Data clustering, determining the number of clusters, cluster- [131, 145, 157]
ing graphical data
Improved Hubert Γ Int Feature selection, deciding the level-of-service [137, 193]
BetaCV Int Data clustering, fuzzy rule interpolation [126, 127]
Fisher’s ratio Int Data clustering [174]
Hartigan Int Fuzzy data clustering [30]
S_Dbw index Int Automatic data clustering [206]
SD validity Int Data clustering [105]
R-squared Int Medical feature selection, image clustering [37, 166]
COP-index Int Data clustering [123]
Wint index Int Data clustering [31, 51]
FS index Int Data clustering [31, 51]
KL index Int Level of service in urban cities [121]
Ray-Turi index Int Detecting discontinuities in rock, data clustering [63, 100]
FHV Int Fuzzy data clustering [36]
PCAES Int Fuzzy data clustering [31, 99]
RL index Int Data clustering [132]
Kwon’s index Int Data clustering [4]
Rand index Ext Automatic clustering, adaptive clustering, data clustering [5, 81, 95, 196]
Adjusted Rand index Ext Clustering stability, evolutionary spectral clustering, auto- [5, 77, 195]
matic clustering
Mutual information Ext Multi-objective clustering of dynamic networks, clustering [56, 77]
stability
V-measure Ext Medical data clustering [155]
Fowlkes M. Scores Ext Data clustering, feature selection, cancer diagnosis [73, 166, 197]
Jaccard index Ext Overlapping clustering, community detection [14, 33]
C. Dice index Ext Image segmentation, multi-objective clustering [111, 181]
Purity Ext Data clustering [8, 10, 57]
Entropy Ext Data clustering, automatic clustering [5, 10, 16]
52 I. Aljarah et al.

Table 2 Internal and external clustering validation indices not used within evolutionary clustering

Type Measure References Formula

Internal WB index [203] WB(M) = M · SSW/SS B
Gamma index [18] GI = (s(+) − s(−))/(s(+) + s(−))
  
Negative entropy incre- [98] NEI = 21 ck ∈C p(ck )log| ck | − 21 log| X | −

ment ck ∈C p(ck )log pck
N
i=1 (u i,1st Max −u i,2nd Max )
Graded distance index [85] GDI = N − c
N
bcd(C)+wcd(C)
Score index [162] S(C) = 1 − (1/(ee )
k
Sk Min j∈k,i = j d(vi ,v j )
SV index [200] SV(k,V,X) = i=1
Vk k =
i=1 Max x j ∈Ci d(x j ,vi )
 n 
i=1 x j ∈Ci ox j
OS index [200] OS(k,X) = OSkk = k
 i=1 Min j∈k,i = j  d(vi ,v j )
k=1 nk nk
avg q i=1 j=i+1 di j
McClain and Rao index [116] MC =   q  
q
avg j∈Cl di j
k=1
p
i∈Ck l=k+1
SS i
Root-Mean-Squared [152] RMSSTD = i=1 p
i=1 d f i
STD
k  
T r (W G k )
Banfeld-Raftery index [48] BR = k=1 n k log nk
s(+)−s(−)
Tau index [153] TI = [(n d (n d −1)/2−t)((n d (n d −1)/2)]1/2
G_plus index [153] G+ = 2s(−)/(n d (n d − 1))
Ksq_DetW index [48] KDI = K 2 (|W G|)
 
Log_Det_Ratio index [48] LDR = Nlog |W|TG| |
 
BG SS
Log_SS_Ratio index [48] LSR = log |W G SS|
Point-Biserial index [117] PBI = (d b − d w )( f w . f b /n 2d )1/2 /Sd
VSC index [149] V_SC (c,U) = Sep N (c, U ) + Comp N (c, U )
k 
Wemmert-Gancarski [48] WI = N1 k=1 Max{0, n k − i∈Ik R(Mi )}
index
B
Gap statistics [183] Gap = 1
B b=1 log Wqb − log Wq
bd( j)−wd( j)
VCN measure [205] VCN = Max(bd( j),wd( j))
KCE index [82] KCE = k × J
STR index [177] STR = [E(K ) − E(K − 1)] · [D(K + 1) − D(K )]
N
NK criterion [184] NK = i=1 f i (x, m i )
V ar N ,c (U,V )
Zhang index [202] Z_c(U,V) = Sep N (c,U )
d v j+1 −d v j
Frey index [60] F=
d s j+1 −d s j
    
Wm −(Wk +Wl ) −1
Beale index [21] BI = (Wk +Wl ) / nn mm −2 ∗ 22/ p − 1
Pseudot2 index [52] Pdot = Bkl / (W k +Wl )
n k +nl −2
In-Group Proportion [86] IGP = (#k|class(k)=class(k N )= j)/(#k|class(k) = j)
(IGP)
    
External Mirkin metric [120] M= k n 2k + k
 n 2 −2 k k
 n2 
k kk
Wallace index [191] WLI = W A→B = a+b ,
a
W B→A = a
a+c

2 (Pr ecision + Recall)

1
Kulczynski index [48] KuI =
McNemar index [48] McI = √yn−ny
yn+ny
yy·nn−yn·ny
Phi index [48] Phi = (yy+yn)(yy+ny)(yn+nn)(ny+nn)
yy+nn
Rogers-Tanimoto index [48] RTI = yy+nn+2(ny+yn)

Russel-Rao index [48] RRI = N1T i< j X 1 (i, j)X 2 (i, j)
yy+nn
Sokal-Sneath indices [48] SSI =
yy+nn+ 21 (ny+yn)
A Comprehensive Review of Evaluation and Fitness Measures … 53

3 Fitness Evaluation Indices

The objective of evolutionary clustering is to optimize the clustering process by

optimizing a fitness function. The aim of the fitness (objective) functions is to evaluate
the quality of the clustering. Broadly speaking, evolutionary objective functions for
clustering can be classified into distance measures and similarity measures. The
following sub-sections review thoroughly the used fitness functions with evolutionary
clustering approaches by presenting a description of each corresponding measure and
in which applications it used.

3.1 Root Mean Squared Error

Distance metrics calculate the space between two points in m-dimensional space.
While in clustering, distance measures find the space between two data instances of m
features. Distance functions in cluster analysis are a kind of similarity measurement.
They used to quantify how much the objects are alike during the clustering process.
Where the smaller the distance implies much similar data points. The Sum of Squared
Errors (SSE) is a measurement used to optimize the obtained clusters throughout
the search process. Equation 81 describes the formula of SSE. It represents the
summation of the squared distance of all data points and each respective centroid.
The clustering technique that results in the smallest SSE value is considered as the
best clustering approach.

|C j |

k 
SS E = dist (n i , c j )2 (81)
j=1 i=1

In literature, several distance measures were implemented for the calculation of

the SSE metric. Hence, this sub-section surveys the utilized distance functions for
evaluating the validity of evolutionary clustering.

3.1.1 Euclidean Distance

The Euclidean distance between two points (a, b) represents the length of the line
that links them together. The distance between a and b with m dimensions is the
square root of the sum of the squared difference between (a) and (b). Given that (a)
and (b) are located in the Euclidean space, then the Euclidean distance is illustrated
by Eq. 82 in regard to the Cartesian coordinate system.

dist (a, b) = (a1 − b1 )2 + (a2 − b2 )2 + · · · + (am − bm )2 (82)
54 I. Aljarah et al.

The Euclidean distance is a metric that obeys certain mathematical properties. First,
the distance is a non-negative value. Second, the distance from the point to itself is
zero. Third, the distance satisfies the symmetry property, where dist(a, b) = dist(b,
a). Lastly, the distance follows the triangle inequality property. Thus, dist (a, b) ≤
dist (a, c) + dist (c, b) [72].
Several research studies have implemented the SSE metric for validating the good-
ness of the clustering process. For instance, [113, 124] proposed a GA algorithm with
k-means for clustering. In which, the sum of squared Euclidean distances was the
fitness function. Also, [94] implemented a two-stage clustering approach for order
clustering. In which, the proposed approach is a combination of neural networks, GA
and PSO. Where the sum of Euclidean distances was the fitness function. Further-
more, [96] designed a clustering approach that is a hybrid of differential evolution
and k-means. Where the sum of squared errors was the fitness criterion. [50] devel-
oped an improved adaptive GA into fuzzy c-means. In order to optimize the initial
centroids by optimizing the SSE based on the Euclidean distance. While [186] intro-
duced the parallel k-bat algorithm for clustering big high-dimensional datasets. In
which, the sum of mean squared Euclidean distance was the objective function. In
addition, in [134], the authors utilized a spiral cuckoo search algorithm for cluster-
ing. The proposed approach was used for spam detection, where SSE metric was a
fitness function. [174] used the SSE measure for the validation of k-means based
on the GA clustering method. Nonetheless, [31] used a weighted within group sum
of squared error as a fitness function. Where the proposed method was a hybrid of
fuzzy c-means and PSO algorithm for finding the optimal number of clusters.

3.1.2 Mahalanobis Distance

Mahalanobis distance is a measurement for calculating the distance between two

vectors in a multivariate space. One of the advantages of Mahalanobis distance is the
adoption of a covariance matrix which leads for quantifying the similarity between
the vectors. However, it has been known to have a high computational cost [59].
−
→
Given a data point (−
→a ) and a centroid ( b ), then Mahalanobis distance is defined by
−1
Eq. 83. In which, C is the inverse covariance matrix [64].

−
→ −
→
dab = (−
→
a − b )C −1 (−
→
a − b )T (83)

Authors in [174] proposed an evolutionary clustering approach for distributed

datasets. The proposed approach is a hybrid of GA algorithm and k-means, where
Mahalanobis distance is integrated as a fitness function. While in [65], the authors
proposed a data filling algorithm using a combination of k-means and information
entropy. In which, Mahalanobis distance is used for clustering instead of Euclidean
distance. Moreover, [29] developed a clustering approach for topic categorization.
Where the squared Mahalanobis distance was the clustering criterion.
A Comprehensive Review of Evaluation and Fitness Measures … 55

3.1.3 Minkowski Distance

Minkowski distance is a generalization form of Euclidean distance. Often known

as L p norm metric. It is formulated in Eq. 84. When the value of p = 1, then it is
Manhattan distance (City-Block), while p = 2 is the Euclidean distance and p = ∞
is the Chebyshev distance [72].

 n

dist (a, b) = 
p
|ai − bi | p , wher e p ≥ 1 (84)
i=1

In [84], the authors proposed an improved clustering approach for multimedia

databases. In which, Minkowski distance is used for clustering validation. In addi-
tion, [76] implemented a hierarchical clustering approach for the detection of drug
toxicity using biomarker genes. The proposed algorithm examined the integration of
Minkowski distance for assessing the goodness of clusters. Whereas, [88] proposed
an incremental clustering approach for time-series datasets. The proposed algorithm
is a combination of two clustering algorithms; k-means and fuzzy c-means. As well
as, four distance measures including the Minkowski algorithm.

3.1.4 Cosine Distance

The cosine distance is the angular distance between two vectors of the same length.
Roughly, it is defined by (1-cosine similarity) and given by Eq. 85. The cosine
similarity for two vectors is calculated by finding the cosine of the enclosed angle
between them. Smaller angles mean higher similarity score, while θ ≥ 90 means no
similarity. The cosine distance is a value ∈ [0, 1], where value (1) means highly distant
vectors and (0) means much closer vectors. Figure 3 presents the cosine distance of
two vectors.
a ·b
cosine distance(a, b) = 1 − (85)
||a|| · ||b||
m
where (a.b) corresponds to the dot product and equals to i=0 (ai × bi ), while ||a||
 2
is the Euclidean norm of vector (a) and equals to ai .
Several studies have implemented the cosine distance for the calculation of SSE
fitness function. [104] designed an evolutionary clustering approach using k-means
and GA algorithm. In which, the cosine distance is used as a fitness function. In
[175], authors proposed a PSO approach for clustering. The proposed approach
implemented the average of cosine distance as a fitness score for optimizing the
clustering. Further, [199] implemented a clustering approach for power curve mod-
eling of wind turbines. The clustering approach utilized different clustering algo-
rithms including k-means. As well as, they deployed several fitness functions such
as the Euclidean, Cosine, and City-Block distances. Also, [15] authors formulated
56 I. Aljarah et al.

Fig. 3 A representation of
cosine distance between
point (a1 ) and centroid (c1 )

a1

c1

θ
cosine distance
= 1 - cosine similarity

0.0

a clustering approach for time analysis of air pollution using k-means. In which,
various distances were used involving the Euclidean, City-Block, Cosine, and corre-
lation distances. Nonetheless, [128] utilized k-means for color-based segmentation
of images. In their approach, authors implemented different distance measures for
SSE metric including the cosine distance.

3.2 (Intra and Inter) Cluster Distances

The Intra-Cluster distance is concerned with quantifying the similarity of points

within a single cluster. As the objective of a clustering problem is maximizing the
similarity among data points; the closer the points the more similar they are. Hence,
the Intra-Cluster distance is considered as a minimization problem. Authors in [25]
suggested three criteria for calculating the Intra-Cluster distance. Mainly, they are the
complete diameter, the average diameter, and the centroid diameter. In the case of the
complete diameter, the Intra-Cluster distance is characterized by the length of the link
between the farthest two points in the cluster. For the average diameter, the distance
is represented by the average of distances among all data points in the cluster. While
the centroid diameter is characterized by the average distance between all data points
and the centroid. Even that, there are other strategies different than the diameter such
as the radius, the variance, and the variance multiplied by the SSE. Variance is a
statistical measure describes the dispersion of data. A high variance value means the
data points are very close to their mean value. The variance is presented by Eq. 86,
where N is the number of data points, X is the mean value [72].

1 
N
σ =
2
(xi − x)2 (86)
N i=1
A Comprehensive Review of Evaluation and Fitness Measures … 57

x1
Single_Linkage Complete_Linkage

Average_Linkage Centroid_Linkage

x2

Fig. 4 A description of the linkage measures between two clusters represented by the red and
black colors in 2-dimensional space. Depicting the single, the complete, the average, and centroid
Inter-Cluster measures

Zang et al. [201] implemented a new clustering method based on the GA algorithm
for improving spectral clustering. Authors used clustering variance ratio as a fitness
function which finds the ratio of the within-cluster variance and the between-cluster
variance. Further, [136] utilized GA, PSO, and DE for partitional clustering, in which,
several fitness criteria have been adopted based heavily on the variance measure.
They include the Marriotte criterion, the trace-within criterion, and the variance
ratio criterion.
Inter-Cluster distance reflects how much the clusters are dissimilar. The more
they are separated from each other the more dissimilar the clusters are. Finding
the Inter-Cluster distances can be performed using the linkage criteria. There are
several Linkage measures such as the single-linkage, complete-linkage, average-
linkage, centroid or median linkage and Ward’s method [72]. Figure 4 illustrates the
difference between the single, complete, average, and centroid measures. The single-
linkage measure finds the pair of data with the minimum distance as determined by
(Min (|x − x  |), x ∈ Ci , x  ∈ C j ) given two clusters (Ci and C j ) and two data
points (x and x  ).
Regarding the complete-linkage criterion, the distance is a maximization metric
represented by the length of the line segment between the two farthest neighbor
points (Max (|x − x  |)). While the average-linkage distance is the average distance
between all pairs of data represented by Eq. 87, where n i and n j are the number of
points in clusters Ci and C j , respectively. Whereas, the centroid-linkage computes
58 I. Aljarah et al.

the center of each cluster then finds the distance between centroids.
1 
d(Ci , C j ) = |x − x  | (87)
ni × n j x∈Ci , x  ∈C j

Nguyen and Kuo [129] proposed a fuzzy-GA clustering for categorical data. In
the proposed approach, single-objective and multi-objective functions were deployed
for assessing the clustering fitness. Where the used objectives are the inter-cluster
distance and the intra-cluster distance. Also, [81] suggested a hybrid strategy for
clustering relying on Grey Wolf and Whale optimizers. The new hybrid approach
used a combination of the inter-cluster distance, the intra-cluster distance, and the
cluster density for fitness evaluation. As a result, the hybrid method outperformed
other evolutionary algorithms in terms of F-measure, Jaccard, and Rand indices.
Furthermore, [108] implemented a clustering method using Ant colony and Ant Lion
optimizations, in which, the combination of inter-cluster and intra-cluster distances
were used for the fitness function structure.

3.3 Jaccard Distance

Jaccard Distance (JD) is a metric that satisfies the properties of the conventional
metrics including the triangle inequality [72]. JD metric is used to measure the
dissimilarity between two samples of data that constitutes subtracting the Jaccard
Similarity (JS) from 1, as represented in Eq. 88 [72]. In which, JS is the Jaccard
similarity between two given data sets (x and y).

J D = 1 − J S(x, y) (88)

Evidently, as the name implies, the Jaccard similarity resembles how much the two
sets (x,y) have common elements. It is defined by the ratio of the similar elements
between the two sets to all similar and not similar elements. In other words, is the
ratio of the intersection over the union of the two sets as in Eq. 89.

|x ∩ y|
J S(x, y) = (89)
|x ∪ y|

Both Jaccard distance and Jaccard similarity have been utilized for the fitness
evaluation within evolutionary clustering. Andrade et al. [14] designed a biased
random-key GA algorithm for overlapping clustering. Where the Jaccard similarity
is used for assessing the fitness during the local search of clustering. While [104]
proposed an evolutionary k-means based on the GA algorithm which adopted the
Jaccard distance as a quality measure for the objective function. Whereas, in [103]
authors suggested a new clustering and validity clustering technique for soft sub-
space clustering. The new algorithm depends on the adaptive evolutionary algorithm
A Comprehensive Review of Evaluation and Fitness Measures … 59

that implemented the Jaccard distance for clustering evaluation. Remarkably, incor-
porating the Jaccard distance exhibited a better performance in comparison to the
results of Euclidean distance. Further, [1] designed a DE-based approach for clus-
tering texts that used Jaccard similarity as a fitness function. Where it outperformed
the Normalized Google Distance (NGD) similarity measure.

3.4 Compactness and Separation Measures of Clusters

Essentially cluster validation can be designed to maximize two objectives; the com-
pactness and the separation. Compactness is a terminology that reflects how much
the points are dense or close. In other words, it quantifies and minimizes the variance
within a cluster. The lower the variance the higher the compactness and the better
the clustering performance. On the other hand, the separation measures how much
the created clusters are disconnected from each other. Separation can be computed
based on different measures such as the distance or the variance. Evidently, having
the clusters being well-separated from each other is an indicator of higher clustering
quality.
Compactness and separation have been utilized for evaluating clustering fitness,
but they also widely used within internal cluster validation. Most internal clustering
validation measures depend on their implementation on the compactness, separation,
or both.
Essentially, [187] implemented a clustering approach based on Principal Com-
ponent Analysis (PCA) for identifying markers of a genotypic data. In which, the
author utilized C-Index, Point Biserial, and Silhouette Indices as evaluation measures
for the best number of clusters. Moreover, [37] used several evolutionary algorithms
such as the ABC and Cuckoo Search for image clustering. In the proposed method,
the Gap statistics has been utilized for evaluating the fitness of the used evolu-
tionary algorithms, while Davies-Bouldin, Silhouette, Dunn, and R-Squared indices
were used for evaluating the clustering results. Further, [112] created an adaptive
k-determination clustering algorithm which adopted the connectivity and deviation
as a fitness evaluation measures. Where the connectivity depends on a similarity
measure to increase the compactness and, similarly, is the deviation that depends on
a distance measure.
Interestingly, various research studies have implemented compactness and sep-
aration as a multi-objective fitness functions that targets maximizing the com-
pactness and the separation. Authors in [27] have formulated the fitness func-
tion in a way that comprises both compactness and separation in the form of
(α × (separation/compactness)), where it utilized for Fractional Lion Optimiza-
tion for clustering. Generally speaking, clustering can be formulated as multi-
objective optimization problem as given by Eq. 90. Where C is the cluster and
(i, j) refer for clusters’ numbers.
60 I. Aljarah et al.

f 1 = compactness(Ci )
f itness = Maximi ze (90)
f 2 = separation(Ci , C j )

For example, [161] demonstrated the usage of multi-objective DE algorithm for

clustering, where both PBM and Silhouette indices were handled as objective func-
tions. While in [198] two objective functions utilized for clustering in the context
of image segmentation. The clustering approach was based on multi-objective arti-
ficial immune learning that optimizes two measures; the cluster connectedness and
the cluster scattering. Whereas [102] integrated three objectives for optimization
including the intra-cluster dispersion, the inter-cluster separation, and the negative
Shannon entropy alongside the Non-dominated Sorting GA algorithm (NSGA-III)
for clustering.

3.5 Clusters Density

Generally, density-based methods are efficient for finding randomly-shaped clusters.

Where the dense regions represent the potential clusters that are separated by the
scattered regions. Primarily, the density of a data item i is characterized by the
number of neighbors around it that satisfies a threshold value. Hence, density-based
methods collect points with the high-density neighborhood to form the cluster. Often,
the density of the neighborhood is quantified by a radius value of r . Figure 5 shows
how the density measure affect the coherence of the candidate clusters.
On the contrary, density-based neighborhood methods are highly sensitive to
the radius value. Therefore, one of the alternative approaches is the utilization of
the kernel density estimation criterion. For instance, [13] proposed an evolutionary
approach for clustering using the PSO algorithm and the density-based estimator
for clusters validation. While [101] suggested a combination of density clustering
algorithm and ACO algorithm for tackling the traveling salesman problem.
The previous sub-sections demonstrated different fitness evaluation functions for
clustering approaches optimized by evolutionary search algorithms. However, there
are other quality validation functions used to optimize the objective function of a
clustering problem without the integration of evolutionary algorithms. Such objective
functions are the Manhattan distance, Negated Average Association, Hamming sim-
ilarity, Gaussian similarity, Spearman correlation, Pearson correlation, and Kendall-
tau ranking. Table 3 shows the definition of those functions. In essence, [178] used
Manhattan distance instead of Euclidean distance for optimizing Ward’s cluster-
ing algorithm. Whereas, [34] implemented the negated average association measure
and the association average as optimization objectives for spectral clustering. [182]
implemented a clustering approach for detecting bias in machine learning algorithms.
In which, k-means was deployed with the hamming distance function. Also, [144]
utilized Gaussian similarity for clustering software modules and textual documents.
While [171] developed a clustering approach based on the Spearman correlation
coefficient for indoor-outdoor environmental detection. Further, [41] presented sev-
A Comprehensive Review of Evaluation and Fitness Measures … 61

Fig. 5 An explanation of
selecting a proper radius of a x1
cluster might result in better
cluster density. Setting r2 as
the radius of density
maximizes the compactness
of the cluster, while both r1 r1
r2
and r3 are improper choices
r3

x2

Table 3 A list of objective functions not used with evolutionary clustering

Function References Formula
n
Manhattan distance [91] dist(a, b) = i=1 |ai − bi |
k assoc(Vi ,Vi )
Negated average [34] NAA = T r (W ) − i=1 |V |
association
n−1
Hamming distance [71] d(a, b) = k=0 (ya,k ∈ / yb,k )
Gaussian similarity [190] s(a ,b) = ex p( ||xa2σ
−x b ||
2 )

6 di2
Spearman correlation [125] Cs = 1 − n(n 2 −1)
i (rui −r u )(rvi −r v )
Pearson correlation [22] C p (u, v) = √ √ 
i (rui −r u ) i (rv i−r v )
2 2
n c −n d
Kendall-Tau ranking [87] τ= n(n−1)/2

eral correlation metrics including Pearson for clustering electricity markets. Yet, in
[44], Kendall-Tau has been adopted in a clustering criterion for preference rankings.
To this end, validating clustering is ubiquitous. Evidently, different problems can
be tackled in different strategies depending on the nature of the problem and its
context. Therefore, clustering validation under the realm of evolutionary algorithms
or even generally is open-ended, where interested researchers can inspect and explore
more.

4 Conclusion

In conclusion, this chapter introduced thoroughly the clustering validation indices

generally and particularly for evolutionary clustering. It is included the internal and
external clustering quality measures, as well as the objective (fitness) functions.
Additionally, it presented the implementation areas of each respective index. As
62 I. Aljarah et al.

clustering is a fundamental unsupervised learning method with immense applica-

tions, quantifying the quality of clustering analysis is not trivial and has been studied
extensively. This chapter covers in-depth clustering validation indices and cluster-
ing objective functions found in the literature. Hence, it is devoted as a reference
point that facilitates the work of researchers who are interested in clustering and
evolutionary clustering.

References

1. Abuobieda, Albaraa, Naomie Salim, Yogan Jaya Kumar, and Ahmed Hamza Osman. 2013.
An improved evolutionary algorithm for extractive text summarization. In Asian conference
on intelligent information and database systems, 78–89. Springer.
2. Al-Madi, Nailah, Ibrahim Aljarah, and Simone A. Ludwig. 2014. Parallel glowworm swarm
optimization clustering algorithm based on mapreduce. In 2014 IEEE symposium on swarm
intelligence (SIS), 1–8. IEEE.
3. Al Shorman, Amaal, Hossam Faris, and Ibrahim Aljarah. 2020. Unsupervised intelligent
system based on one class support vector machine and grey wolf optimization for iot botnet
detection. Journal of Ambient Intelligence and Humanized Computing 11 (7): 2809–2825.
4. Aldouri, Yamur K., Hassan Al-Chalabi, and Liangwei Zhang. 2018. Data clustering and imput-
ing using a two-level multi-objective genetic algorithm (GA): A case study of maintenance
cost data for tunnel fans. Cogent Engineering 5 (1): 1513304.
5. Aliniya, Zahra, and Seyed Abolghasem Mirroshandel. 2019. A novel combinatorial merge-
split approach for automatic clustering using imperialist competitive algorithm. Expert Sys-
tems with Applications 117: 243–266.
6. Aljarah, Ibrahim, and Simone A. Ludwig. 2012. Parallel particle swarm optimization cluster-
ing algorithm based on mapreduce methodology. In 2012 fourth world congress on Nature
and biologically inspired computing (NaBIC), 104–111. IEEE.
7. Aljarah, Ibrahim, and Simone A. Ludwig. 2013. Mapreduce intrusion detection system based
on a particle swarm optimization clustering algorithm. In 2013 IEEE congress on evolutionary
computation (CEC), 955–962. IEEE.
8. Aljarah, Ibrahim, and Simone A. Ludwig. 2013. A new clustering approach based on glow-
worm swarm optimization. In 2013 IEEE congress on evolutionary computation, 2642–2649.
IEEE.
9. Aljarah, Ibrahim, and Simone A. Ludwig. 2013. Towards a scalable intrusion detection sys-
tem based on parallel pso clustering using mapreduce. In Proceedings of the 15th annual
conference companion on genetic and evolutionary computation, 169–170. ACM.
10. Aljarah, Ibrahim, Majdi Mafarja, Ali Asghar Heidari, Hossam Faris, and Seyedali Mir-
jalili. 2019. Clustering analysis using a novel locality-informed grey wolf-inspired clustering
approach. Knowledge and Information Systems 1–33.
11. Aljarah, Ibrahim, Majdi Mafarja, Ali Asghar Heidari, Hossam Faris, and Seyedali Mir-
jalili. 2020. Clustering analysis using a novel locality-informed grey wolf-inspired clustering
approach. Knowledge and Information Systems 62 (2): 507–539.
12. Aljarah, Ibrahim, Majdi Mafarja, Ali Asghar Heidari, Hossam Faris, and Seyedali Mirjalili.
2020. Multi-verse optimizer: Theory, literature review, and application in data clustering. In
Nature-inspired optimizers, 123–141. Springer.
13. Alswaitti, Mohammed, Mohanad Albughdadi, and Nor Ashidi Mat Isa. 2018. Density-based
particle swarm optimization algorithm for data clustering. Expert Systems with Applications
91: 170–186.
14. Andrade, Carlos E., Mauricio G.C. Resende, Howard J. Karloff, and Flávio K. Miyazawa.
2014. Evolutionary algorithms for overlapping correlation clustering. In Proceedings of the
2014 annual conference on genetic and evolutionary computation, 405–412. ACM.
A Comprehensive Review of Evaluation and Fitness Measures … 63

15. Arroyo, Ángel, Verónica Tricio, Álvaro Herrero, and Emilio Corchado. 2016. Time analysis
of air pollution in a spanish region through k-means. In International joint conference SOCO-
16-CISIS-16-ICEUTE-16, 63–72. Springer.
16. Auffarth, Benjamin. 2010. Clustering by a genetic algorithm with biased mutation operator.
In IEEE congress on evolutionary computation, 1–8. IEEE.
17. Azarakhsh, Javad, and Zobeir Raisi. 2019. Automatic clustering using metaheuristic algo-
rithms for content based image retrieval. In Fundamental research in electrical engineering,
83–99. Springer.
18. Baker, Frank B., and Lawrence J. Hubert. 1975. Measuring the power of hierarchical cluster
analysis. Journal of the American Statistical Association 70 (349): 31–38.
19. Ball, Geoffrey H., and David J. Hall. 1965. Isodata, a novel method of data analysis and
pattern classification. Technical report, Stanford Research Institute, Menlo Park, CA.
20. Bandyopadhyay, Sanghamitra, and Ujjwal Maulik. 2002. Genetic clustering for automatic
evolution of clusters and application to image classification. Pattern Recognition 35 (6):
1197–1208.
21. Beale, E.M.L. 1969. Cluster analysis scientific control system.
22. Benesty, Jacob, Jingdong Chen, Yiteng Huang, and Israel Cohen. 2009. Pearson correlation
coefficient. In Noise reduction in speech processing, 1–4. Springer.
23. Bezdek, James C. 2013. Pattern recognition with fuzzy objective function algorithms. Springer
Science & Business Media.
24. Bezdek, James C., and Nikhil R. Pal. 1998. Some new indices of cluster validity.
25. Bolshakova, Nadia, and Francisco Azuaje. 2003. Cluster validation techniques for genome
expression data. Signal Processing 83 (4): 825–833.
26. Caliński, Tadeusz, and Jerzy Harabasz. 1974. A dendrite method for cluster analysis. Com-
munications in Statistics-Theory and Methods 3 (1): 1–27.
27. Chander, Satish, P. Vijaya, and Praveen Dhyani. 2018. Multi kernel and dynamic fractional
lion optimization algorithm for data clustering. Alexandria Engineering Journal 57 (1): 267–
276.
28. Charrad, Malika, Mohamed Ben Ahmed, Yves Lechevallier, and Gilbert Saporta. 2009. Deter-
mining the number of clusters in CROKI2 algorithm.
29. Chen, Junyang, Zhiguo Gong, and Weiwen Liu. 2019. A nonparametric model for online topic
discovery with word embeddings. Information Sciences 504: 32–47.
30. Chen, Min, and Simone A. Ludwig. 2014. Fuzzy clustering using automatic particle swarm
optimization. In 2014 IEEE international conference on fuzzy systems (FUZZ-IEEE), 1545–
1552. IEEE.
31. Chen, Min, and Simone A. Ludwig. 2014. Particle swarm optimization based fuzzy clustering
approach to identify optimal number of clusters. Journal of Artificial Intelligence and Soft
Computing Research 4 (1): 43–56.
32. Chen, Scott Shaobing, and Ponani S. Gopalakrishnan. 1998. Clustering via the Bayesian
information criterion with applications in speech recognition. In Proceedings of the 1998
IEEE international conference on acoustics, speech and signal processing, ICASSP’98 (Cat.
No. 98CH36181), vol. 2, 645–648. IEEE.
33. Cheng, Fan, Su Tingting Cui, Yunyun Niu Yansen, and Xingyi Zhang. 2018. A local infor-
mation based multi-objective evolutionary algorithm for community detection in complex
networks. Applied Soft Computing 69: 357–367.
34. Chi, Yun, Xiaodan Song, Dengyong Zhou, Koji Hino, and Belle L. Tseng. 2007. Evolutionary
spectral clustering by incorporating temporal smoothness. In Proceedings of the 13th ACM
SIGKDD international conference on Knowledge discovery and data mining, 153–162. ACM.
35. Chou, Chien-Hsing, Su Mu-Chun, and Eugene Lai. 2003. A new cluster validity measure for
clusters with different densities. In IASTED international conference on intelligent systems
and control, 276–281.
36. Chovancova, Olga, Lucia Piatrikova, and Adam Dudas. 2019. Improving fuzzy c-means algo-
rithm using particle swarm optimization and fuzzy c-means++. In 2019 international confer-
ence on information and digital technologies (IDT), 173–179. IEEE.
64 I. Aljarah et al.

37. Civicioglu, P., U.H. Atasever, C. Ozkan, E. Besdok, A.E. Karkinli, and A. Kesikoglu. 2014.
Performance comparison of evolutionary algorithms for image clustering. The International
Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences 40 (7): 71.
38. Cobos, Carlos, Leydy Muñoz, Martha Mendoza, Elizabeth León, and Enrique Herrera-
Viedma. 2012. Fitness function obtained from a genetic programming approach for web
document clustering using evolutionary algorithms. In Ibero-American conference on artifi-
cial intelligence, 179–188. Springer.
39. Cobos, Carlos, Henry Muñoz-Collazos, Richar Urbano-Muñoz, Martha Mendoza, Elizabeth
León, and Enrique Herrera-Viedma. 2014. Clustering of web search results based on the
cuckoo search algorithm and balanced Bayesian information criterion. Information Sciences
281: 248–264.
40. Cowgill, Marcus Charles. 1993. Monte Carlo validation of two genetic clustering algorithms.
41. Cui, Tianyu, Francesco Caravelli, and Cozmin Ududec. 2018. Correlations and clustering
in wholesale electricity markets. Physica A: Statistical Mechanics and its Applications 492:
1507–1522.
42. Czekanowski, Jan. 1932. Coefficient of racial likeness" und durchschnittliche differenz.
Anthropologischer Anzeiger (H. 3/4): 227–249.
43. Dalrymple-Alford, E.C. 1970. Measurement of clustering in free recall. Psychological Bulletin
74 (1): 32.
44. D’Ambrosio, Antonio, and Willem J. Heiser. 2019. A distribution-free soft-clustering method
for preference rankings. Behaviormetrika 46 (2): 333–351.
45. Das, Swagatam, and Amit Konar. 2009. Automatic image pixel clustering with an improved
differential evolution. Applied Soft Computing 9 (1): 226–236.
46. Das, Swagatam, and Sudeshna Sil. 2010. Kernel-induced fuzzy clustering of image pixels
with an improved differential evolution algorithm. Information Sciences 180 (8): 1237–1256.
47. Davies, David L., and Donald W. Bouldin. 1979. A cluster separation measure. IEEE Trans-
actions on Pattern Analysis and Machine Intelligence 2: 224–227.
48. Desgraupes, Bernard. 2013. Clustering indices. University of Paris Ouest-Lab Modal-X 1:
34.
49. Dice, Lee R. 1945. Measures of the amount of ecologic association between species. Ecology
26 (3): 297–302.
50. Ding, Yi, and Fu Xian. 2016. Kernel-based fuzzy c-means clustering algorithm based on
genetic algorithm. Neurocomputing 188: 233–238.
51. Duan, Ganglong, Hu Wenxiu, and Zhiguang Zhang. 2016. A novel data clustering algorithm
based on modified adaptive particle swarm optimization. International Journal of Signal
Processing, Image Processing and Pattern Recognition 9 (3): 179–188.
52. Duda, Richard O., and Peter E Hart. 1973. Pattern classification and scene analysis. Wiley.
53. Dunn, Joseph C. 1973. A fuzzy relative of the isodata process and its use in detecting compact
well-separated clusters.
54. Faris, Hossam, Ibrahim Aljarah, and Ja’far Alqatawna. 2015. Optimizing feedforward neural
networks using krill herd algorithm for e-mail spam detection. In 2015 IEEE Jordan confer-
ence on applied electrical engineering and computing technologies (AEECT), 1–5. IEEE.
55. Faris, Hossam, Ibrahim Aljarah, Seyedali Mirjalili, Pedro A. Castillo, and Juan J. Merelo.
2016. EvoloPy: An open-source nature-inspired optimization framework in Python. In IJCCI
(ECTA), 171–177.
56. Folino, Francesco, and Clara Pizzuti. 2010. A multiobjective and evolutionary clustering
method for dynamic networks. In 2010 international conference on advances in social net-
works analysis and mining, 256–263. IEEE.
57. Forestier, Germain, Cédric Wemmert, and Pierre Gançarski. 2010. Background knowledge
integration in clustering using purity indexes. In International conference on knowledge sci-
ence, engineering and management, 28–38. Springer.
58. Fowlkes, Edward B., and Colin L. Mallows. 1983. A method for comparing two hierarchical
clusterings. Journal of the American Statistical Association 78 (383): 553–569.
A Comprehensive Review of Evaluation and Fitness Measures … 65

59. Franco, Manuel, and Juana-María Vivo. 2019. Cluster analysis of microarray data. In Microar-
ray bioinformatics, 153–183. Springer.
60. Frey, T., and H. Van Groenewoud. 1972. A cluster analysis of the D2 matrix of white spruce
stands in Saskatchewan based on the maximum-minimum principle. The Journal of Ecology
873–886.
61. Friedman, Herman P., and Jerrold Rubin. 1967. On some invariant criteria for grouping data.
Journal of the American Statistical Association 62 (320): 1159–1178.
62. Fukuyama, Yoshiki. 1989. A new method of choosing the number of clusters for the fuzzy
c-mean method. In Proceedings of the 5th fuzzy systems symposium, 1989, 247–250.
63. Georgieva, Kristina S., and Andries P. Engelbrecht. 2013. Dynamic differential evolution
algorithm for clustering temporal data. In International conference on large-scale scientific
computing, 240–247. Springer.
64. Gnanadesikan, R., J.W. Harvey, and J.R. Kettenring. 1993. Mahalanobis metrics for cluster
analysis. Sankhyā: The Indian Journal of Statistics, Series A 494–505.
65. Gong, Xiaofei, Jie Zhang, and Yijie Shi. 2018. Research on data filling algorithm based on
improved k-means and information entropy. In 2018 IEEE 4th international conference on
computer and communications (ICCC), 1774–1778. IEEE.
66. Guan, Chun, Kevin Kam Fung Yuen, and Frans Coenen. 2019. Particle swarm optimized
density-based clustering and classification: Supervised and unsupervised learning approaches.
Swarm and Evolutionary Computation 44: 876–896.
67. Gurrutxaga, Ibai, Iñaki Albisua, Olatz Arbelaitz, José I. Martín, Javier Muguerza, Jesús M.
Pérez, and Iñigo Perona. 2010. SEP/COP: An efficient method to find the best partition in
hierarchical clustering based on a new cluster validity index. Pattern Recognition 43 (10):
3364–3373.
68. Gutoski, Matheus, Manassés Ribeiro, Nelson Marcelo Romero Aquino, Leandro Takeshi
Hattori, André Eugênio Lazzaretti, and Heitor Silvério Lopes. 2018. Feature selection using
differential evolution for unsupervised image clustering. In International conference on arti-
ficial intelligence and soft computing, 376–385. Springer.
69. Halkidi, Maria, and Michalis Vazirgiannis. 2001. Clustering validity assessment: Finding the
optimal partitioning of a data set. In Proceedings 2001 IEEE international conference on data
mining, 187–194. IEEE.
70. Halkidi, Maria, Michalis Vazirgiannis, and Yannis Batistakis. 2000. Quality scheme assess-
ment in the clustering process. In European conference on principles of data mining and
knowledge discovery, 265–276. Springer.
71. Hamming, Richard W. 1950. Error detecting and error correcting codes. The Bell System
Technical Journal 29 (2): 147–160.
72. Han, Jiawei, Micheline Kamber, and Jian Pei. 2011. Getting to know your data. Data Mining:
Concepts and Techniques 3 (744): 39–81.
73. Hancer, Emrah. 2020. A new multi-objective differential evolution approach for simultane-
ous clustering and feature selection. Engineering Applications of Artificial Intelligence 87:
103307.
74. Handl, Julia, and Joshua Knowles. 2007. An evolutionary approach to multiobjective cluster-
ing. IEEE Transactions on Evolutionary Computation 11 (1): 56–76.
75. Hartigan, John A. 1975. Clustering algorithms. New York, NY: Wiley Inc.
76. Hasan, Mohammad Nazmol, Masuma Binte Malek, Anjuman Ara Begum,Moizur Rahman,
Md. Mollah, and Nurul Haque. 2019. Assessment of drugs toxicity and associated biomarker
genes using hierarchical clustering. Medicina 55 (8): 451.
77. He, Zhenfeng, and Chunyan Yu. 2019. Clustering stability-based evolutionary k-means. Soft
Computing 23 (1): 305–321.
78. Hubert, Lawrence, and Phipps Arabie. 1985. Comparing partitions. Journal of Classification
2 (1): 193–218.
79. Hubert, Lawrence, and James Schultz. 1976. Quadratic assignment as a general data analysis
strategy. British Journal of Mathematical and Statistical Psychology 29 (2): 190–241.
66 I. Aljarah et al.

80. Jaccard, Paul. 1901. Étude comparative de la distribution florale dans une portion des alpes
et des jura. Bulletin de la Societe Vaudoise des Sciences Naturelles 37: 547–579.
81. Jadhav, Amolkumar Narayan, and N. Gomathi. 2018. WGC: Hybridization of exponential
grey wolf optimizer with whale optimization for data clustering. Alexandria Engineering
Journal 57 (3): 1569–1584.
82. Jauhiainen, Susanne, and Tommi Kärkkäinen. 2017. A simple cluster validation index with
maximal coverage. In European symposium on artificial neural networks, computational
intelligence and machine learning. ESANN.
83. Jiang, Lei, and Datong Xie. 2018. An efficient differential memetic algorithm for clustering
problem. IAENG International Journal of Computer Science 45 (1).
84. Jiang, Xiaoping, Chenghua Li, and Jing Sun. 2018. A modified k-means clustering for mining
of multimedia databases based on dimensionality reduction and similarity measures. Cluster
Computing 21 (1): 797–804.
85. Joopudi, Sreeram, Suraj S. Rathi, Shankar Narasimhan, and Raghunathan Rengaswamy. 2013.
A new cluster validity index for fuzzy clustering. IFAC Proceedings Volumes 46 (32): 325–
330.
86. Kapp, Amy V., and Robert Tibshirani. 2006. Are clusters found in one dataset present in
another dataset? Biostatistics 8 (1): 9–31.
87. Kendall, Maurice G. 1938. A new measure of rank correlation. Biometrika 30 (1/2): 81–93.
88. Khobragade, Sneha, and Preeti Mulay. 2018. Enhance incremental clustering for time series
datasets using distance measures. In International conference on intelligent computing and
applications, 543–556. Springer.
89. Kim, Minho, and R.S. Ramakrishna. 2005. Some new indexes of cluster validity. Pattern
Recognition Letters 26 (15): 2353–2363.
90. Kirkland, Oliver, Victor J. Rayward-Smith, and Beatriz de la Iglesia. 2011. A novel multi-
objective genetic algorithm for clustering. In International conference on intelligent data
engineering and automated learning, 317–326. Springer.
91. Klamroth, Kathrin. 2002. Measuring distances. In Single-facility location problems with bar-
riers, 3–14. Springer.
92. Krzanowski, Wojtek J., and Y.T. Lai. 1988. A criterion for determining the number of groups
in a data set using sum-of-squares clustering. Biometrics 23–34.
93. Kumar, Lalit, and Kusum Kumari Bharti. 2019. An improved BPSO algorithm for feature
selection. In Recent trends in communication, computing, and electronics, 505–513. Springer.
94. Kuo, R.J., and L.M. Lin. 2010. Application of a hybrid of genetic algorithm and particle
swarm optimization algorithm for order clustering. Decision Support Systems 49: 451–462.
95. Kurada, Ramachandra Rao, and Karteeka Pavan Kanadam. 2019. A novel evolutionary
automatic clustering technique by unifying initial seed selection algorithms into teaching–
learning-based optimization. In Soft computing and medical bioinformatics, 1–9. Springer.
96. Kwedlo, Wojciech. 2011. A clustering method combining differential evolution with the k-
means algorithm. Pattern Recognition Letters 32 (12): 1613–1621.
97. Kwon, Soon H. 1998. Cluster validity index for fuzzy clustering. Electronics Letters 34 (22):
2176–2177.
98. Lago-Fernández, Luis F., and Fernando Corbacho. 2010. Normality-based validation for crisp
clustering. Pattern Recognition 43 (3): 782–795.
99. Li, Chaoshun, Jianzhong Zhou, Pangao Kou, and Jian Xiao. 2012. A novel chaotic particle
swarm optimization based fuzzy clustering algorithm. Neurocomputing 83: 98–109.
100. Li, Yanyan, Qing Wang, Jianping Chen, Liming Xu, and Shengyuan Song. 2015. K-means
algorithm based on particle swarm optimization for the identification of rock discontinuity
sets. Rock Mechanics and Rock Engineering 48 (1): 375–385.
101. Liao, Erchong, and Changan Liu. 2018. A hierarchical algorithm based on density peaks
clustering and ant colony optimization for traveling salesman problem. IEEE Access 6: 38921–
38933.
102. Liu, Chao, Yuanrui Li, Qi Zhao, and Chenqi Liu. 2019. Reference vector-based multi-objective
clustering for high-dimensional data. Applied Soft Computing 78: 614–629.
A Comprehensive Review of Evaluation and Fitness Measures … 67

103. Liu, Chao, Jing Xie, Qi Zhao, Qiwei Xie, and Chenqi Liu. 2019. Novel evolutionary multi-
objective soft subspace clustering algorithm for credit risk assessment. Expert Systems with
Applications 138: 112827.
104. Liu, Chuanren, Tianming Hu, Yong Ge, and Hui Xiong. 2012. Which distance metric is right:
An evolutionary k-means view. In Proceedings of the 2012 SIAM international conference on
data mining, 907–918. SIAM.
105. Liu, Yimin, Tansel Özyer, Reda Alhajj, and Ken Barker. 2005. Integrating multi-objective
genetic algorithm and validity analysis for locating and ranking alternative clustering. Infor-
matica 29 (1).
106. Łukasik, Szymon, Piotr A. Kowalski, Małgorzata Charytanowicz, and Piotr Kulczycki. 2016.
Clustering using flower pollination algorithm and Calinski-Harabasz index. In 2016 IEEE
congress on evolutionary computation (CEC), 2724–2728. IEEE.
107. Luna-Romera, José María, Jorge García-Gutiérrez, María Martínez-Ballesteros, and José C.
Riquelme Santos. 2018. An approach to validity indices for clustering techniques in big data.
Progress in Artificial Intelligence 7 (2): 81–94.
108. Mageshkumar, C., S. Karthik, and V.P. Arunachalam. 2019. Hybrid metaheuristic algorithm
for improving the efficiency of data clustering. Cluster Computing 22 (1): 435–442.
109. Marriott, F.H.C. 1971. Practical problems in a method of cluster analysis. Biometrics 501–514.
110. Martarelli, Nádia Junqueira, and Marcelo Seido Nagano. 2018. A constructive evolutionary
approach for feature selection in unsupervised learning. Swarm and Evolutionary Computa-
tion 42: 125–137.
111. Martínez-Peñaloza, María-Guadalupe, Efrén Mezura-Montes, Nicandro Cruz-Ramírez,
Héctor-Gabriel Acosta-Mesa, and Homero-Vladimir Ríos-Figueroa. 2017. Improved multi-
objective clustering with automatic determination of the number of clusters. Neural Comput-
ing and Applications 28 (8): 2255–2275.
112. Matake, Nobukazu, Tomoyuki Hiroyasu, Mitsunori Miki, and Tomoharu Senda. 2007. Mul-
tiobjective clustering with automatic k-determination for large-scale data. In Proceedings of
the 9th annual conference on genetic and evolutionary computation, 861–868. ACM.
113. Maulik, Ujjwal, and Sanghamitra Bandyopadhyay. 2000. Genetic algorithm-based clustering
technique. Pattern Recognition 33 (9): 1455–1465.
114. Maulik, Ujjwal, and Sanghamitra Bandyopadhyay. 2002. Performance evaluation of some
clustering algorithms and validity indices. IEEE Transactions on Pattern Analysis and
Machine Intelligence 24 (12): 1650–1654.
115. Maulik, Ujjwal, and Indrajit Saha. 2009. Modified differential evolution based fuzzy clustering
for pixel classification in remote sensing imagery. Pattern Recognition 42 (9): 2135–2149.
116. McClain, John O., and Vithala R. Rao. 1975. CLUSTISZ: A program to test for the quality of
clustering of a set of objects. JMR, Journal of Marketing Research (pre-1986) 12 (000004):
456.
117. Milligan, Glenn W. 1980. An examination of the effect of six types of error perturbation on
fifteen clustering algorithms. Psychometrika 45 (3): 325–342.
118. Milligan, Glenn W., and Martha C. Cooper. 1985. An examination of procedures for deter-
mining the number of clusters in a data set. Psychometrika 50 (2): 159–179.
119. Milligan, Glenn W., and Martha C. Cooper. 1986. A study of the comparability of external
criteria for hierarchical cluster analysis. Multivariate Behavioral Research 21 (4): 441–458.
120. Mirkin, Boris. 1998. Mathematical classification and clustering: From how to what and why.
In Classification, data analysis, and data highways, 172–181. Springer.
121. Mohapatra, Smruti Sourava, Prasanta Kumar Bhuyan, and K.V. Rao. 2012. Genetic algorithm
fuzzy clustering using GPS data for defining level of service criteria of urban streets.
122. Mukhopadhyay, Anirban, Ujjwal Maulik, and Sanghamitra Bandyopadhyay. 2015. A survey
of multiobjective evolutionary clustering. ACM Computing Surveys (CSUR) 47 (4): 61.
123. Muravyov, Sergey, Denis Antipov, Arina Buzdalova, and Andrey Filchenkov. 2019. Efficient
computation of fitness function for evolutionary clustering. MENDEL 25: 87–94.
124. Murthy, Chivukula A., and Nirmalya Chowdhury. 1996. In search of optimal clusters using
genetic algorithms. Pattern Recognition Letters 17: 825–832.
68 I. Aljarah et al.

125. Myers, Leann, and Maria J. Sirois. 2004. Spearman correlation coefficients, differences
between. In Encyclopedia of statistical sciences, 12.
126. Naik, Nitin, Ren Diao, and Qiang Shen. 2015. Choice of effective fitness functions for genetic
algorithm-aided dynamic fuzzy rule interpolation. In 2015 IEEE international conference on
fuzzy systems (FUZZ-IEEE), 1–8. IEEE.
127. Nerurkar, Pranav, Aruna Pavate, Mansi Shah, and Samuel Jacob. 2019. Performance of internal
cluster validations measures for evolutionary clustering. In Computing, communication and
signal processing, 305–312. Springer.
128. Neumann, Aneta, and Frank Neumann. 2018. On the use of colour-based segmentation in
evolutionary image composition. In 2018 IEEE congress on evolutionary computation (CEC),
1–8. IEEE.
129. Nguyen, Thi Phuong Quyen, and R.J. Kuo. 2019. Partition-and-merge based fuzzy genetic
clustering algorithm for categorical data. Applied Soft Computing 75: 254–264.
130. Oujezsky, Vaclav, and Tomas Horvath. 2018. Traffic similarity observation using a genetic
algorithm and clustering. Technologies 6 (4): 103.
131. Ozyer, Tansel, and Reda Alhajj. 2006. Achieving natural clustering by validating results
of iterative evolutionary clustering approach. In 2006 3rd international IEEE conference
intelligent systems, 488–493. IEEE.
132. Özyer, Tansel, Ming Zhang, and Reda Alhajj. 2011. Integrating multi-objective genetic algo-
rithm based clustering and data partitioning for skyline computation. Applied Intelligence 35
(1): 110–122.
133. Pakhira, Malay K., Sanghamitra Bandyopadhyay, and Ujjwal Maulik. 2004. Validity index
for crisp and fuzzy clusters. Pattern Recognition 37 (3): 487–501.
134. Pandey, Avinash Chandra, and Dharmveer Singh Rajpoot. 2019. Spam review detection using
spiral cuckoo search clustering method. Evolutionary Intelligence 12 (2): 147–164.
135. Pantula, Priyanka D., Srinivas S. Miriyala, and Kishalay Mitra. 2019. A novel ANN-fuzzy
formulation towards evolution of efficient clustering algorithm. In 2019 Fifth Indian control
conference (ICC), 254–259. IEEE.
136. Paterlini, Sandra, and Thiemo Krink. 2006. Differential evolution and particle swarm optimi-
sation in partitional clustering. Computational Statistics & Data Analysis 50 (5): 1220–1247.
137. Patnaik, Ashish Kumar, and Prasanta Kumar Bhuyan. 2016. Application of genetic program-
ming clustering in defining LOS criteria of urban street in Indian context. Travel Behaviour
and Society 3: 38–50.
138. Peng, Hong, Peng Shi, Jun Wang, Agustín Riscos-Núñez, and Mario J. Pérez-Jiménez. 2017.
Multiobjective fuzzy clustering approach based on tissue-like membrane systems. Knowledge-
Based Systems 125: 74–82.
139. Pomponi, Eraldo, and Alexei Vinogradov. 2013. A real-time approach to acoustic emission
clustering. Mechanical Systems and Signal Processing 40 (2): 791–804.
140. Qaddoura, Raneem, Hossam Faris, and Ibrahim Aljarah. 2020. An efficient clustering algo-
rithm based on the k-nearest neighbors with an indexing ratio. International Journal of
Machine Learning and Cybernetics 11 (3): 675–714.
141. Qaddoura, Raneem, Hossam Faris, Ibrahim Aljarah, and Pedro A. Castillo. 2020. EvoClus-
ter: An open-source nature-inspired optimization clustering framework in Python. In Interna-
tional conference on the applications of evolutionary computation (part of EvoStar), 20–36.
Springer.
142. Qaddoura, R., H. Faris, and I. Aljarah. 2020. An efficient evolutionary algorithm with a
nearest neighbor search technique for clustering analysis. Journal of Ambient Intelligence
and Humanized Computing 1–26.
143. Qaddoura, R., H. Faris, I. Aljarah, J. Merelo, and P. Castillo. 2020. Empirical evaluation of
distance measures for nearest point with indexing ratio clustering algorithm. In Proceedings
of the 12th international joint conference on computational intelligence, vol. 1, 430–438.
NCTA. ISBN 978-989-758-475-6, https://doi.org/10.5220/0010121504300438.
144. Radhakrishna, Vangipuram, Chintakindi Srinivas, and C.V. GuruRao. 2014. A modified gaus-
sian similarity measure for clustering software components and documents. In Proceedings of
A Comprehensive Review of Evaluation and Fitness Measures … 69

the international conference on information systems and design of communication, 99–104.

ACM.
145. Raitoharju, Jenni, Kaveh Samiee, Serkan Kiranyaz, and Moncef Gabbouj. 2017. Particle
swarm clustering fitness evaluation with computational centroids. Swarm and Evolutionary
Computation 34: 103–118.
146. Rand, William M. 1971. Objective criteria for the evaluation of clustering methods. Journal
of the American Statistical Association 66 (336): 846–850.
147. Ratkowsky, D.A., and G.N. Lance. 1978. Criterion for determining the number of groups in
a classification.
148. Ray, Siddheswar, and Rose H. Turi. 1999. Determination of number of clusters in k-means
clustering and application in colour image segmentation. In Proceedings of the 4th inter-
national conference on advances in pattern recognition and digital techniques, 137–143.
Calcutta, India.
149. Rezaee, Babak. 2010. A cluster validity index for fuzzy clustering. Fuzzy Sets and Systems
161 (23): 3014–3025.
150. Ripon, Kazi Shah Nawaz, and Mia Nazmul Haque Siddique. 2009. Evolutionary multi-
objective clustering for overlapping clusters detection. In 2009 IEEE congress on evolutionary
computation, 976–982. IEEE.
151. Ripon, K.S. Nawaz, Chi-Ho Tsang, Sam Kwong, and Man-Ki Ip. 2006. Multi-objective evo-
lutionary clustering using variable-length real jumping genes genetic algorithm. In 18th inter-
national conference on pattern recognition (ICPR’06), vol. 1, 1200–1203. IEEE.
152. Ritz, Christian, and Ib Skovgaard. 2005. Module 2: Cluster analysis. Master of Applied Statis-
tics, 1–20. http://www2.imm.dtu.dk/~pbb/MAS/ST116/module02/module.pdf.
153. Rohlf, F.James. 1974. Methods of comparing classifications. Annual Review of Ecology and
Systematics 5 (1): 101–113.
154. Rosenberg, Andrew, and Julia Hirschberg. 2007. V-measure: A conditional entropy-based
external cluster evaluation measure. In Proceedings of the 2007 joint conference on empir-
ical methods in natural language processing and computational natural language learning
(EMNLP-CoNLL), 410–420.
155. Ross, Edith M., and Florian Markowetz. 2016. OncoNEM: Inferring tumor evolution from
single-cell sequencing data. Genome Biology 17 (1): 69.
156. Rousseeuw, Peter J. 1987. Silhouettes: A graphical aid to the interpretation and validation of
cluster analysis. Journal of Computational and Applied Mathematics 20: 53–65.
157. Roy, Parthajit, and J.K. Mandal. 2016. An SVD based real coded genetic algorithm for graph
clustering.
158. Runkler, Thomas A., and James C. Bezdek. 2019. Optimizing the c index using a canonical
genetic algorithm. In International conference on the applications of evolutionary computa-
tion (part of EvoStar), 287–298. Springer.
159. Saha, Sriparna, and Sanghamitra Bandyopadhyay. 2008. Application of a new symmetry-
based cluster validity index for satellite image segmentation. IEEE Geoscience and Remote
Sensing Letters 5 (2): 166–170.
160. Saha, Sriparna, and Ranjita Das. 2018. Exploring differential evolution and particle swarm
optimization to develop some symmetry-based automatic clustering techniques: Application
to gene clustering. Neural Computing and Applications 30 (3): 735–757.
161. Saini, Naveen, Sriparna Saha, and Pushpak Bhattacharyya. 2019. Automatic scientific doc-
ument clustering using self-organized multi-objective differential evolution. Cognitive Com-
putation 11 (2): 271–293.
162. Saitta, Sandro, Benny Raphael, and Ian F.C. Smith. 2007. A bounded index for cluster validity.
In International workshop on machine learning and data mining in pattern recognition, 174–
187. Springer.
163. Saitta, Sandro, Benny Raphael, and Ian F.C. Smith. 2008. A comprehensive validity index for
clustering. Intelligent Data Analysis 12 (6): 529–548.
164. Santos, Adam, Eloi Figueiredo, Moisés Silva, Reginaldo Santos, Claudomiro Sales, and João
C.W.A. Costa. 2017. Genetic-based EM algorithm to improve the robustness of gaussian
70 I. Aljarah et al.

mixture models for damage detection in bridges. Structural Control and Health Monitoring
24 (3): e1886.
165. Sarle, W.S. 1983. SAS technical report A-108, cubic clustering criterion, 56. Cary, NC: SAS
Institute Inc.
166. Sayed, Gehad Ismail, Ashraf Darwish, and Aboul Ella Hassanien. 2019. Binary whale opti-
mization algorithm and binary moth flame optimization with clustering algorithms for clinical
breast cancer diagnoses. Journal of Classification 1–31.
167. Schwarz, Gideon, et al. 1978. Estimating the dimension of a model. The Annals of Statistics
6 (2): 461–464.
168. Scott, Allen J., and Symons, Michael J. 1971. Clustering methods based on likelihood ratio
criteria. Biometrics 387–397.
169. sheng Li, Chun. 2011. The improved partition coefficient. Procedia Engineering 24: 534–538.
170. Shirakawa, Shinichi, and Tomoharu Nagao. 2009. Evolutionary image segmentation based on
multiobjective clustering. In 2009 IEEE congress on evolutionary computation, 2466–2473.
IEEE.
171. Shtar, Guy, Bracha Shapira, and Lior Rokach. 2019. Clustering Wi-Fi fingerprints for indoor-
outdoor detection. Wireless Networks 25 (3): 1341–1359.
172. Shukri, Sarah, Hossam Faris, Ibrahim Aljarah, Seyedali Mirjalili, and Ajith Abraham. 2018.
Evolutionary static and dynamic clustering algorithms based on multi-verse optimizer. Engi-
neering Applications of Artificial Intelligence 72: 54–66.
173. Silva, Jesús, Omar Bonerge Pineda Lezama, Noel Varela, Jesús García Guiliany, Ernesto Stef-
fens Sanabria, Madelin Sánchez Otero, and Vladimir Álvarez Rojas. 2019. U-control chart
based differential evolution clustering for determining the number of cluster in k-means. In
International conference on green, pervasive, and cloud computing, 31–41. Springer.
174. Sinha, Ankita, and Prasanta K. Jana. 2018. A hybrid mapreduce-based k-means clustering
using genetic algorithm for distributed datasets. The Journal of Supercomputing 74 (4): 1562–
1579.
175. Song, Wei, Wei Ma, and Yingying Qiao. 2017. Particle swarm optimization algorithm with
environmental factors for clustering analysis. Soft Computing 21 (2): 283–293.
176. Sørensen, Thorvald Julius. 1948. A method of establishing groups of equal amplitude in
plant sociology based on similarity of species content and its application to analyses of the
vegetation on Danish commons. I kommission hos E. Munksgaard.
177. Starczewski, Artur. 2017. A new validity index for crisp clusters. Pattern Analysis and Appli-
cations 20 (3): 687–700.
178. Strauss, Trudie, and Michael Johan von Maltitz. 2017. Generalising ward’s method for use
with Manhattan distances. PloS One 12 (1): e0168288.
179. Strehl, Alexander. 2002. Relationship-based clustering and cluster ensembles for high-
dimensional data mining. PhD thesis.
180. Strehl, Alexander, and Joydeep Ghosh. 2002. Cluster ensembles—a knowledge reuse frame-
work for combining multiple partitions. Journal of Machine Learning Research 3: 583–617.
181. Tan, Teck Yan, Li Zhang, Chee Peng Lim, Ben Fielding, Yu. Yonghong, and Emma Anderson.
2019. Evolving ensemble models for image segmentation using enhanced particle swarm
optimization. IEEE Access 7: 34004–34019.
182. Thomson, Robert, Elie Alhajjar, Joshua Irwin, and Travis Russell. 2018. Predicting bias in
machine learned classifiers using clustering. In Annual social computing, behavior prediction,
and modeling-behavioral representation in modeling simulation conference.
183. Tibshirani, Robert, Guenther Walther, and Trevor Hastie. 2001. Estimating the number of
clusters in a data set via the gap statistic. Journal of the Royal Statistical Society: Series B
(Statistical Methodology) 63 (2): 411–423.
184. Tinós, Renato, Zhao Liang, Francisco Chicano, and Darrell Whitley. 2016. A new evaluation
function for clustering: The NK internal validation criterion. In Proceedings of the genetic
and evolutionary computation conference 2016, 509–516. ACM.
185. Trauwaert, E., P. Rousseeuw, and L. Kaufman. 1993. Fuzzy clustering by minimizing the total
hypervolume. In Information and classification, 61–71. Springer.
A Comprehensive Review of Evaluation and Fitness Measures … 71

186. Tripathi, Ashish Kumar, Kapil Sharma, and Manju Bala. 2018. Dynamic frequency based
parallel k-bat algorithm for massive data clustering (DFBPKBA). International Journal of
System Assurance Engineering and Management 9 (4): 866–874.
187. van Heerwaarden, Joost, T.L. Odong, and F.A. van Eeuwijk. 2013. Maximizing genetic differ-
entiation in core collections by PCA-based clustering of molecular marker data. Theoretical
and Applied Genetics 126 (3): 763–772.
188. Vazirgiannis, Michalis. 2009. Clustering validity, 388–393. Boston, MA: Springer, US.
189. Vinh, Nguyen Xuan, Julien Epps, and James Bailey. 2010. Information theoretic measures
for clusterings comparison: Variants, properties, normalization and correction for chance.
Journal of Machine Learning Research 11: 2837–2854.
190. Von Luxburg, Ulrike. 2007. A tutorial on spectral clustering. Statistics and Computing 17 (4):
395–416.
191. Wallace, David L. 1983. A method for comparing two hierarchical clusterings: Comment.
Journal of the American Statistical Association 78 (383): 569–576.
192. Wu, Kuo-Lung, and Miin-Shen Yang. 2005. A cluster validity index for fuzzy clustering.
Pattern Recognition Letters 26 (9): 1275–1291.
193. Wu, Yi-Leh, Cheng-Yuan Tang, Maw-Kae Hor, and Wu Pei-Fen. 2011. Feature selection using
genetic algorithm and cluster validation. Expert Systems with Applications 38 (3): 2727–2732.
194. Xie, Xuanli Lisa, and Gerardo Beni. 1991. A validity measure for fuzzy clustering. IEEE
Transactions on Pattern Analysis & Machine Intelligence 8: 841–847.
195. Xu, Kevin S., Mark Kliger, and Alfred O. Hero. 2010. Evolutionary spectral clustering with
adaptive forgetting factor. In 2010 IEEE international conference on acoustics, speech and
signal processing, 2174–2177. IEEE.
196. Xu, Kevin S., Mark Kliger, and Alfred O. Hero III. 2014. Adaptive evolutionary clustering.
Data Mining and Knowledge Discovery 28 (2): 304–336.
197. Xu, Rui, Jie Xu, and Donald C. Wunsch. 2012. A comparison study of validity indices on
swarm-intelligence-based clustering. IEEE Transactions on Systems, Man, and Cybernetics,
Part B (Cybernetics) 42 (4): 1243–1256.
198. Yang, Dongdong, Xiaowei Zhang, Lintao Lv, and Wenzhun Huang. 2018. An automatic SAR
image segmentation framework by multi-objective clustering and artificial immune learn-
ing. In 2018 international conference on mathematics, modelling, simulation and algorithms
(MMSA 2018). Atlantis Press.
199. Yesilbudak, Mehmet. 2018. Implementation of novel hybrid approaches for power curve
modeling of wind turbines. Energy Conversion and Management 171: 156–169.
200. Žalik, Krista Rizman, and Borut Žalik. 2011. Validity index for clusters of different sizes and
densities. Pattern Recognition Letters 32 (2): 221–234.
201. Zang, Wenke, Zhenni Jiang, and Liyan Ren. 2017. Improved spectral clustering based on
density combining DNA genetic algorithm. International Journal of Pattern Recognition and
Artificial Intelligence 31 (04): 1750010.
202. Zhang, Yunjie, Weina Wang, Xiaona Zhang, and Yi Li. 2008. A cluster validity index for
fuzzy clustering. Information Sciences 178 (4): 1205–1218.
203. Zhao, Qinpei, and Pasi Fränti. 2014. WB-index: A sum-of-squares based index for cluster
validity. Data & Knowledge Engineering 92: 77–89.
204. Zhao, Ying, and George Karypis. 2002. Evaluation of hierarchical clustering algorithms for
document datasets. In Proceedings of the eleventh international conference on Information
and knowledge management, 515–524. ACM.
205. Zhou, Shibing, and Xu Zhenyuan. 2018. A novel internal validity index based on the cluster
centre and the nearest neighbour cluster. Applied Soft Computing 71: 78–88.
206. Zhu, Shuwei, Lihong Xu, and Erik D. Goodman. 2019.Evolutionary multi-objective automatic
clustering enhanced with quality metrics and ensemble strategy. Knowledge-Based Systems
105018.
A Grey Wolf-Based Clustering Algorithm
for Medical Diagnosis Problems

Raneem Qaddoura, Ibrahim Aljarah, Hossam Faris, and Seyedali Mirjalili

Abstract Evolutionary and swarm intelligence algorithms are used as optimization

algorithms for solving the clustering problem. One of the most popular optimization
algorithms is the Grey Wolf Optimizer (GWO). In this chapter, we use GWO on
seven medical data sets to optimize the initial clustering centroids represented by
the individuals of each population at each iteration. The aim is to minimize the
distances between instances of the same cluster to predict certain diseases and medical
problems. The results show that solving the clustering task using GWO outperforms
the other well-regarded evolutionary and swarm intelligence clustering algorithms,
by converging toward enhanced solutions having low dispersion from the average
values, for all the selected data sets.

Keywords Data clustering · Nature-inspired algorithms · Metaheuristics ·

Medical applications · Grey wolf optimizer · Evolutionary computation ·
EvoCluster · EvoloPy

R. Qaddoura
Information Technology, Philadlphia University, Amman, Jordan
e-mail: [email protected]
I. Aljarah (B) · H. Faris
King Abdullah II School for Information Technology, The University of Jordan,
Amman, Jordan
e-mail: [email protected]
H. Faris
e-mail: [email protected]
S. Mirjalili
Center for Artificial Intelligence Research and Optimization, Torrens University Australia,
Fortitude Valley, Brisbane, Australia
e-mail: [email protected]

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer 73
Nature Singapore Pte Ltd. 2021
I. Aljarah et al. (eds.), Evolutionary Data Clustering: Algorithms and Applications, Algo-
rithms for Intelligent Systems,
https://doi.org/10.1007/978-981-33-4191-3_3
74 R. Qaddoura et al.

1 Introduction

Clustering is an unsupervised learning problem, which gathers similar instances to

the same cluster by analyzing the patterns of the features of these instances [10–12,
26]. The clustering task can be solved using partitional, density-based, hierarchical
clustering [27, 56, 60], and nature-inspired clustering algorithms [4, 59, 63].
Grey Wolf Optimizer (GWO) [47] as one of the nature-inspired clustering algo-
rithms, is a popular algorithm that is used for optimization problems and can be used
for solving the clustering task [37, 67, 70].
Clustering applications include financial risk analysis [34], document catego-
rization [14, 43], dental radiography segmentation [55], information retrieval [18,
58], image processing [33, 36, 66], search engines [40], academics [52], and pat-
tern recognition [39, 62]. Some other applications related to biological and medical
applications include bioinformatics [24], cancerous data [30, 65], and drug activity
prediction [5], and many others [3, 9, 22, 23].
In this chapter, we use GWO, which is customized for clustering as part of the
EvoCluster framework [57], to optimize the centroids of the clustering solution for
medical data sets, by minimizing the sum of squared error (SSE) of the instances
of each cluster, to predict certain diseases and medical problems. The clustering
solutions are represented by the individuals of the population at each iteration. The
experiments are conducted on seven popular data sets, and the results are compared
with other well-known evolutionary and swarm intelligence clustering algorithms.
The remainder of the chapter is organized as follows: Section 2 summarizes the
related work in the field of evolutionary and swarm intelligence clustering. Section 3
describes GWO in detail and how it is customized to solve the clustering task. Section
4 presents the results of conducting the experiments on well-known medical data sets.
Finally, Sect. 5 concludes the work.

2 Related Work

Nature-inspired algorithms try to solve the optimization problems by applying natural

operations on candidate solutions [53, 54]. Fuad [25] categorized the nature-inspired
algorithms into evolutionary algorithms (EA) and swarm intelligence (SI). The most
popular EAs are the genetic algorithm (GA) [16], genetic programming (GP) [35],
multi-verse optimizer (MVO) [46], evolutionary strategies (ES) [28], and differential
evolution (DE) [63]. In contrast, many SIs exist including particle swarm optimiza-
tion (PSO) [20], firefly algorithm (FFA) [68], ant colony optimization (ACO) [19],
cuckoo search (CS) [69], artificial bee colony (ABC) [32], moth-flame optimization
(MFO) [44], whale optimization algorithm (WOA) [45], and grey wolf optimizer
(GWO) [47].
A Grey Wolf-Based Clustering Algorithm for Medical Diagnosis Problems 75

These algorithms are used to solve different optimization problems including the
clustering problem. Many studies can be found in the literature concerning clustering
with evolutionary and swarm intelligence algorithms which are reviewed in several
surveys. The authors of [27] reviewed the studies of determining the number of
clusters using evolutionary algorithms, while the authors of [51] reviewed nature-
inspired algorithms of partitional clustering. In addition, authors of [49, 50] reviewed
the evolutionary clustering algorithms in a two-parts survey. Faris et al. [21] reviewed
the recent variants and applications of GWO and presented the clustering applications
of using GWO as part of their review.
In 2014, Mirjalili et al. [47] implemented GWO which becomes a very popular
swarm intelligence algorithm for solving the optimization problems. Many studies
were observed since then which are classified into updating mechanisms, new oper-
ators, encoding scheme of the individuals, and population structure and hierarchy as
per the review study by Faris et al. [21].
Some studies can be found on GWO for performing the clustering task by optimiz-
ing the centroids of the clustering solution. Grey Wolf Algorithm Clustering (GWAC)
[37], GWO with Powell local optimization (PGWO) [70], Grey Wolf Optimizer and
K-means (GWO-KM) [67], and K-means with GWO (K-GWO) are clustering algo-
rithms which work on enhancing the centroids decision for the individuals in the
population of the GWO at each iteration. In addition, the authors in [48] have devel-
oped mGWO which balances between the exploratory and exploitative behavior of
the algorithm to generate enhanced results. An enhanced GWO (EGWO) was pro-
posed by [64], which is a hybrid combination of the hunting process of GWO and
the binomial crossover and lévy flight steps. Another hybrid algorithm named WGC
was proposed by [29] as a combination of the GWO and the WOA algorithms with
a newly formulated fitness function.
GWO are used in some clustering applications including satellite image segmen-
tation [31] and wireless sensor networks [1, 48]. However, many other clustering
applications can be experimented using the GWO as a clustering algorithm which is
the main goal of this chapter. Thus, we experiment GWO as a clustering algorithm
on medical data sets.

3 Grey Wolf Optimizer for Clustering

This section presents the GWO algorithm details which are used to optimize general
problems. This section also includes the details of optimizing the centroids selection
of the clustering solution using GWO.
76 R. Qaddoura et al.

3.1 GWO Algorithm

GWO is inspired by the social behavior of grey wolf packs which includes the
following:

• Social hierarchy of leadership: The leader of the pack is referred to as the alpha
wolf. Beta wolves support the alpha wolf in decision-making and substitute them
in case of death or illness [21, 47]. The remaining hierarchy of leadership includes
delta and omega wolves.
• Group hunting: It includes tracking, chasing, pursuing, encircling, harassing, and
attacking the prey [47].

The social behavior of wolves is translated into an optimization algorithm by

mathematically modeling the behavior.
The hierarchy of leadership is obtained by considering the best solution in a
population as the alpha wolf (α), the following best solutions are considered as beta
wolf (β) and delta wolf (δ), respectively. Consequently, the remaining solutions in
the population are considered as omega wolves (ω).
The hunting behavior of the wolves is obtained by creating a set of solutions
of grey wolves for the first iteration of the algorithm. For the following iterations,
the position of the prey is predicted by the alpha, beta, and delta wolves. Candidate
solutions converge toward the prey, which results in exploiting the search space,
and diverge from the prey, which results in exploring fitter solutions in the search
space when the values of A are considered between −1 and 1, or the values of A
are considered larger than 1 or smaller than −1, respectively. This can be further
explained by applying the following processes:
• Encircling: Equations 1, 2, 3, 4 simulate this process. Equation 1 represents a
vector calculated according to the position vector of the prey and the current
position vector of the wolf. Equation 2 represents the next position vector of the
wolf calculated according to the current position vector of the wolf and the vector
calculated in Eq. 1. Equations 3 and 4 represents the coefficient vectors A and C,
respectively.
D = |C.X p (t) − X (t)| (1)

X (t + 1) = X (t) − A.D (2)

C = 2.r2 (3)

A = 2a.r1 − a (4)

Where t is the current iteration, X and X p are the positions of the grey wolf and
the prey, respectively, r1 and r2 are random vectors of values between 0 and 1, and
a is a vector which linearly decreases from 2 to 0.
A Grey Wolf-Based Clustering Algorithm for Medical Diagnosis Problems 77

• Hunting: The next position vector of the omega wolves are calculated by Eq. 5
according to the position vectors of the alpha, beta, and delta wolves. The values
of the X 1 , X 2 , X 3 are calculated by Eqs. 6, 7, and 8, respectively.

X (t + 1) = (X 1 + X 2 + X 3 )/3 (5)

X 1 = X α − A1 .(|C1 .X α − X |) (6)

X 2 = X β − A2 .(|C2 .X β − X |) (7)

X 3 = X δ − A3 .(|C3 .X δ − X |) (8)

Where X α , X β , and X δ are the position vectors of the alpha, beta, and delta wolves,
X and X (t + 1) are the current and next position vectors of the wolves, respectively.
A1 & C1 , A2 & C2 , and A3 & C3 are coefficient vectors for alpha, beta, and delta
wolves, respectively.
• Attacking (exploitation): Decreasing the value of a from 2 to 0 causes a consequent
decrease of the value of A from 1 to −1 which results in moving toward the prey
and finally attacking the prey when the prey stops moving. This technique results
in the exploitative behavior of searching for the fitter preys.
• Searching (exploration): Grey wolves diverge from the prey by considering values
of A less than −1 or larger than 1. In addition, the random values of weights
of C, which are between 0 and 2, provide different effects of the position of the
prey. Both techniques show the exploratory behavior of the search space and local
optima avoidance of the solutions.

3.2 Clustering with GWO

The main idea of clustering with GWO is to optimize the centroids selection of
the clustering solutions. This can be achieved by selecting an initial population of
solutions, which represents the individuals of the initial centroids. Then, the instances
are assigned to the closest centroids and the evaluation of each individual is achieved
by calculating the sum of squared error (SSE) of the distances between the centroids
and the corresponding instances at each cluster. For the following iterations, GWO
optimizes the candidate solutions, according to the aforementioned processes, which
are discussed in the previous section. The fittest solution of the last iteration is
returned to reflect the solution achieved by applying the algorithm.
The pseudocode of performing the clustering task using GWO can be observed by
Algorithm 1. The algorithm accepts the data set instances, the number of clusters (k),
the number of iterations (#iterations), and the number of individuals (#individuals)
as parameter values. The initial population is created in line 2 which represents the
initial centroids of the candidate individuals. Then, the algorithm repeats until the
value of iterations is reached which is presented in lines 3–10. For each iteration,
78 R. Qaddoura et al.

instances of each solution are assigned to the closest cluster and the SSE value
of each individual is calculated. These operations are presented in lines 4–7. The
algorithm recognizes the alpha, beta, delta, and omega individuals in line 8. Then,
the Encircling, Hunting, Attacking, and Searching processes are performed in line
9. The algorithm terminates in line 11 by returning the alpha individual of the last
iteration.

Algorithm 1 Clustering with GWO

1: procedure CGWO(#iterations, #individuals, k, instances)
2: Create initial population of individuals
3: repeat
4: for all individuals ∈ population do
5: Assign instances to the closest centroids
6: Calculate SSE
7: end for
8: Point out alpha, beta, delta, and omega individuals
9: Perform Encircling, Hunting, Attacking, and Searching processes
10: until #iterations is reached
11: Return alpha individual
12: end procedure

4 Experimental Results

This section discusses the data sets used in the experiments, the parameter selec-
tion for each algorithm, the fitness function used in the experiments, and a detailed
analysis of the obtained results.

4.1 Data Sets

Seven medical data sets are used for the conducted experiments, which are gathered
from the UCI machine learning repository1 [17] and KEEL.2 Table 1 shows the
number of classes, instances, and features for each data set.
The description of these data sets are given as follows:
• Appendicitis2 : Represents the presence or absence of appendicitis for 106 patients
depending on 7 measures.
• Blood1 : This data set is gathered from the Blood Transfusion Service Center in
Hsin-Chu City in Taiwan. It represents the status of denoting blood in March 2007,

1 https://archive.ics.uci.edu/ml/.
2 https://sci2s.ugr.es/keel/.
A Grey Wolf-Based Clustering Algorithm for Medical Diagnosis Problems 79

Table 1 Characteristics of Data sets

ID Name #Classes #Instances #Features Type
1 Appendicitis 2 106 7 Real
2 Blood 2 748 5 Real
3 Diagnosis II 2 120 6 Real
4 Heart 2 270 13 Real
5 Liver 2 345 7 Real
6 Vertebral II 2 310 6 Real
7 Vertebral III 3 310 6 Real

based on 4 attributes of 748 donors. The attributes include the months since last
donation, total number of donations, total blood donated in c.c., and months since
first donation.
• Diagnosis II1 : Represents the presence or absence of nephritis of renal pelvis
origin for 120 patients depending on 6 symptoms.
• Heart1 : Represents the presence or absence of a heart disease for 270 patients
depending on 13 attributes.
• Liver1 : Represents the presence or absence of liver disorders that might be caused
by alcohol consumption for 345 patients depending on 7 attributes.
• Vertebral II1 : The data set is gathered by Dr. Henrique da Mota inCentre
MAmedico-Chirurgical de RAadaptation des Massues, Lyon, France. It represents
the presence or absence of vertebral problems which are classified into normal or
abnormal vertebral for 310 patients depending on 6 attributes
• Vertebral III1 : The same as the Vertebral II data set but abnormal vertebral are
split into Disk Hernia and Spondylolisthesis forming 3 classes in addition to the
normal vertebral.

4.2 Algorithms and Parameters Selection

Recent and well-known algorithms are used to compare the results achieved by
running the CGWO algorithm. These algorithms include CSSA, CGA, CPSO, and
CWOA which represent the clustering variation of the original algorithms which are
SSA, GA, PSO, and WOA, respectively. They use a population of individuals which
represent the clustering solutions of centroids that are optimized across iterations.
The number of classes for each data set is passed as a parameter to each algorithm.
The population size of 50 and the iteration value of 100 are also considered for all
algorithms. In addition, Roulette Wheel selection mechanism and the values of 0.8
and 0.001 for the crossover and mutation probabilities, respectively, are considered
for CGA which are extensively found in the literature [13, 15, 41, 42, 61]. Table 2
80 R. Qaddoura et al.

Table 2 Parameters setting

Algorithm Parameter Value
CGA Crossover probability 0.8
Mutation probability 0.001
CPSO Vmax 6
wMax 0.9
wMin 0.2
c1 2
c2 2
CGWO a From 2 to 0
CSSA No. iterations 100
No. population size 50
No. of leaders 0.5
CWOA a From 2 to 0

shows the selection of parameters for these algorithms. The population/swarm size
is unified for all algorithms and set to 50, while the number of iterations is set to 100.

4.3 Fitness Function

We use the Sum of Squared Error (SSE) [38] to evaluate the convergence of the
algorithms toward the optimal solution, which indicates compact clusters. The SSE
is the sum of the squares of the euclidean distance between an instance and the
corresponding centroid of its cluster. The objective is to minimize the value of SSE
to obtain optimized clustering results. SSE is calculated by the following equation
[38]:


k 
J
SS E = || p j − ci ||2 (9)
i=1 j=1

where k is the number of clusters J is the number of points for a cluster i, p j is the
j th point in cluster i, ci is the centroid of cluster i.

4.4 Results and Discussion

The experiments are conducted on the selected data sets for 30 independent runs, then
the averages of the results are reported. Table 3 shows the performance of CGWO
against other recent and well-known algorithms which are CSSA, CGA, CPSO, and
A Grey Wolf-Based Clustering Algorithm for Medical Diagnosis Problems 81

Table 3 Average values of SSE of different algorithms for 30 independent runs

Dataset CGWO CSSA CGA CPSO CWOA
Appendicitis 17.88 19.63 21.11 21.13 22.18
Blood 36.8 46.67 43.27 41.55 39.39
Diagnosis II 106.11 106.43 109.21 106.76 108.7
Heart 259.18 259.39 284.64 295.33 301.15
Liver 24.12 27.82 30.02 29.83 31.71
Vertebral2 21.33 25.18 26.8 26.4 29.59
Vertebral3 18.85 20.95 26.58 21.07 27.29

CWOA. It is observed from the table that CGWO outperforms the other algorithms,
having the lowest SSE value compared to the other algorithms for all the selected
data sets. It has a recognizable lower value than CWOA and CGA for most of the
data sets. It also has close values for Diagnosis II and heart data sets compared to
CSSA.
The convergence curves of the algorithms for the selected data sets are also shown
in Fig. 1. The convergence curve represents the average SSE value for 30 runs for each
of the algorithms. It shows the tendency of each algorithm for finding an enhanced
solution by minimizing the SSE value across 100 iterations. It is observed from
the figure that CGWO has recognizable optimized solutions during the course of
iterations for Liver, Blood, Vertebral 2, and Appendicitis data sets. It slightly improves
the solution for Diagnosis II data set compared to CSSA and CPSO algorithms.
It competes CSSA in optimizing the solution during the course of iterations for
Vertebral II, Vertebral III, Diagnosis II, and Heart data sets, but finally reaches a
better solution than CSSA for these data sets.
Furthermore, the box plots of the SSE values for each data set are shown in
Fig. 2 to access the stability of CGWO. The box plot shows the interquartile range,
average SSE value, best SSE values, and worst SSE value [6] for the 30 runs of every
algorithm in comparison.
It is observed from the figure that CGWO has the most compacted box compared
to the other algorithms for all the data sets, which indicates low standard deviation
and stability of the algorithm. Specifically, it has almost the same values for SSE
for different runs of CGWO for Liver and Blood data sets having very low values
of standard deviation. It also has a recognizable minimal SSE value compared to the
other algorithms for most of the data sets.
82 R. Qaddoura et al.

(a) Appendicitis (b) Blood

(c) Diagnosis II (d) Heart

(e) Liver (f) Vertebral II

(g) Vertebral III

Fig. 1 Convergence curve for CSSA, CWOA, CPSO, CGA, and CGWO using SSE objective
function for a Appendicitis; b Blood; c Diagnosis II; d Heart; e Liver; f Vertebral II; and g Vertebral
III
A Grey Wolf-Based Clustering Algorithm for Medical Diagnosis Problems 83

Fig. 2 Box plot for CSSA, CWOA, CPSO, CGA, and CGWO using SSE objective function for a
Appendicitis; b Blood; c Diagnosis II; d Heart; e Liver; f Vertebral II; and g Vertebral III
84 R. Qaddoura et al.

5 Conclusion

In this chapter, we have applied GWO, which is customized for clustering as part of
the EvoCluster framework, as an optimizer for solving the clustering task for seven
medical data sets. The individuals at each iteration represent the initial centroids that
are used to cluster the data set instances into certain diseases and medical problems.
The results show that applying GWO on the selected medical data sets causes a
recognizable convergence of the algorithm toward enhanced solutions compared to
other well-regarding evolutionary and swarm intelligence algorithms. In addition, the
box plots of thirty independent runs for each algorithm show that lower dispersion
from the average values is achieved using GWO compared to the other algorithms.
For future work, we plan to investigate other clustering applications using the
same customized clustering algorithm of GWO.

References

1. Al-Aboody, N.A., and H.S. Al-Raweshidy. 2016. Grey wolf optimization-based energy-
efficient routing protocol for heterogeneous wireless sensor networks. In 2016 4th International
Symposium on Computational and Business Intelligence (ISCBI), pp. 101–107. IEEE.
2. Al-Madi, Nailah, Ibrahim, Aljarah, and Simone A. Ludwig. 2014. Parallel glowworm swarm
optimization clustering algorithm based on mapreduce. In 2014 IEEE Symposium on Swarm
Intelligence, pp. 1–8. IEEE.
3. Al Shorman, Amaal, Hossam, Faris, and Ibrahim, Aljarah. 2020. Unsupervised intelligent
system based on one class support vector machine and grey wolf optimization for iot botnet
detection. Journal of Ambient Intelligence and Humanized Computing 11(7):2809–2825.
4. Alam, Shafiq, Gillian, Dobbie, Yun Sing, Koh, Patricia, Riddle, and Saeed Ur, Rehman. 2014.
Research on particle swarm optimization based clustering: a systematic review of literature
and techniques. Swarm and Evolutionary Computation 17:1–13.
5. Alhalaweh, Amjad, Ahmad Alzghoul, and Waseem Kaialy. 2014. Data mining of solubility
parameters for computational prediction of drug-excipient miscibility. Drug Development and
Industrial Pharmacy 40 (7): 904–909.
6. Aljarah, Ibrahim, Al-Zoubi, AlaM, Hossam, Faris, Mohammad A. Hassonah, Seyedali, Mir-
jalili, and Heba, Saadeh. 2018. Simultaneous feature selection and support vector machine
optimization using the grasshopper optimization algorithm. Cognitive Computation, pp. 1–18.
7. Aljarah, Ibrahim, and Simone A. Ludwig. 2012. Parallel particle swarm optimization clustering
algorithm based on mapreduce methodology. In 2012 Fourth World Congress on Nature and
Biologically Inspired Computing (NaBIC), pp. 104–111. IEEE.
8. Aljarah, Ibrahim, and Simone A. Ludwig. 2013. Mapreduce intrusion detection system based
on a particle swarm optimization clustering algorithm. In 2013 IEEE congress on evolutionary
computation, pp. 955–962. IEEE.
9. Aljarah, Ibrahim, and Simone A. Ludwig. 2013. A new clustering approach based on glowworm
swarm optimization. In 2013 IEEE congress on evolutionary computation, pp. 2642–2649.
IEEE.
10. Aljarah, Ibrahim, and Simone A. Ludwig. 2013. Towards a scalable intrusion detection system
based on parallel pso clustering using mapreduce. In Proceedings of the 15th annual conference
companion on genetic and evolutionary computation, pp. 169–170.
11. Aljarah, Ibrahim, Majdi, Mafarja, Ali Asghar, Heidari, Hossam, Faris, and Seyedali, Mir-
jalili. 2020. Clustering analysis using a novel locality-informed grey wolf-inspired clustering
approach. Knowledge and Information Systems, 62(2):507–539.
A Grey Wolf-Based Clustering Algorithm for Medical Diagnosis Problems 85

12. Aljarah, Ibrahim, Majdi, Mafarja, Ali Asghar, Heidari, Hossam, Faris, and Seyedali, Mirjalili.
2020. Multi-verse optimizer: theory, literature review, and application in data clustering. In
Nature-Inspired Optimizers, pp. 123–141. Berlin: Springer.
13. Beg, A.H. and M.d. Zahidul, Islam. 2015. Clustering by genetic algorithm-high quality chro-
mosome selection for initial population. In 2015 IEEE 10th conference on industrial electronics
and applications (ICIEA), pp. 129–134. IEEE.
14. Brodić, Darko, Alessia, Amelio, and Zoran N. Milivojević. 2017. Clustering documents in
evolving languages by image texture analysis. Applied Intelligence, 46(4):916–933.
15. Chang, Dong-Xia, Xian-Da Zhang, and Chang-Wen Zheng. 2009. A genetic algorithm with
gene rearrangement for k-means clustering. Pattern Recognition 42 (7): 1210–1222.
16. Davis, Lawrence. 1991. Handbook of genetic algorithms.
17. Dheeru, Dua, and Efi Karra, Taniskidou. 2017. UCI machine learning repository, 2017.
18. Djenouri, Youcef, Asma, Belhadi, Philippe, Fournier-Viger, and Jerry Chun-Wei, Lin. 2018.
Fast and effective cluster-based information retrieval using frequent closed itemsets. Informa-
tion Sciences, 453:154–167.
19. Marco Dorigo and Luca Maria Gambardella. 1997. Ant colony system: a cooperative learning
approach to the traveling salesman problem. IEEE Transactions on Evolutionary Computation
1 (1): 53–66.
20. Eberhart, Russell, and Kennedy, James. 1995. A new optimizer using particle swarm theory.
In MHS’95. Proceedings of the sixth international symposium on micro machine and human
science, pp. 39–43. IEEE.
21. Faris, Hossam, Ibrahim, Aljarah, Mohammed Azmi, Al-Betar, and Seyedali, Mirjalili. 2018.
Grey wolf optimizer: a review of recent variants and applications. Neural Computing and
Applications, 30(2):413–435.
22. Faris, Hossam, Ibrahim, Aljarah, and Ja’far, Alqatawna. Optimizing feedforward neural net-
works using krill herd algorithm for e-mail spam detection. In 2015 IEEE Jordan conference
on applied electrical engineering and computing technologies (AEECT), pp. 1–5. IEEE.
23. Faris, Hossam, Ibrahim, Aljarah, Seyedali, Mirjalili, Pedro A. Castillo, and Juan Julián Merelo,
Guervós. 2016. Evolopy: An open-source nature-inspired optimization framework in python.
In IJCCI (ECTA), pp. 171–177.
24. Frank, Eibe, Mark, Hall, Len, Trigg, Geoffrey, Holmes, and Ian H. Witten. 2004. Data mining
in bioinformatics using weka. Bioinformatics, 20(15):2479–2481.
25. Fuad, Muhammad Marwan Muhammad. 2019. Applying nature-inspired optimization algo-
rithms for selecting important timestamps to reduce time series dimensionality. Evolving Sys-
tems 10 (1): 13–28.
26. Han, Jiawei, Jian Pei, and Micheline, Kamber. 2011. Data mining: concepts and techniques.
Elsevier.
27. Hancer, Emrah, and Dervis Karaboga. 2017. A comprehensive survey of traditional, merge-
split and evolutionary approaches proposed for determination of cluster number. Swarm and
Evolutionary Computation 32: 49–67.
28. Hansen, Nikolaus, and Stefan, Kern. 2004. Evaluating the cma evolution strategy on multimodal
test functions. In International conference on parallel problem solving from nature, pp. 282–
291. Berlin: Springer.
29. Jadhav, Amolkumar Narayan, and N. Gomathi. 2018. Wgc: hybridization of exponential grey
wolf optimizer with whale optimization for data clustering. Alexandria engineering journal,
57(3):1569–1584.
30. Jang, Ho, Youngmi, Hur, and Hyunju, Lee. 2016. Identification of cancer-driver genes in focal
genomic alterations from whole genome sequencing data. Scientific Reports 6.
31. Kapoor, Shubham, Irshad, Zeya, Chirag, Singhal, and Satyasai Jagannath, Nanda. 2017. A grey
wolf optimizer based automatic clustering algorithm for satellite image segmentation. Procedia
Computer Science, 115:415–422.
32. Karaboga, Dervis. 2005. An idea based on honey bee swarm for numerical optimization.
Technical report, Technical report-tr06, Erciyes University, Engineering Faculty, Computer,
2005.
86 R. Qaddoura et al.

33. Khan, Zubair, Jianjun Ni, Xinnan Fan, and Pengfei Shi. 2017. An improved k-means clustering
algorithm based on an adaptive initial parameter estimation procedure for image segmentation.
International Journal Of Innovative Computing Information and Control 13(5):1509–1525.
34. Kou, Gang, Yi Peng, and Guoxun Wang. 2014. Evaluation of clustering algorithms for financial
risk analysis using mcdm methods. Information Sciences 275: 1–12.
35. Koza, John R., and John R. Koza. 1992. Genetic programming: on the programming of com-
puters by means of natural selection, vol. 1. MIT Press.
36. Kumar, Sushil, Millie Pant, Manoj Kumar, and Aditya Dutt. 2018. Colour image segmen-
tation with histogram and homogeneity histogram difference using evolutionary algorithms.
International Journal of Machine Learning and Cybernetics 9 (1): 163–183.
37. Kumar, Vijay, Jitender Kumar, Chhabra, and Dinesh, Kumar. 2017. Grey wolf algorithm-based
clustering technique. Journal of Intelligent Systems, 26(1):153–168.
38. Lee, C.-Y., and E.K. Antonsson. 2000. Dynamic partitional clustering using evolution strategies.
In Industrial Electronics Society, 2000. IECON 2000. 26th Annual Confjerence of the IEEE,
vol. 4, pp. 2716–2721. IEEE.
39. Liu, Anan, Yuting, Su, Weizhi, Nie, and Mohan S. Kankanhalli. 2017. Hierarchical clustering
multi-task learning for joint human action grouping and recognition. IEEE Trans. Pattern Anal.
Mach. Intell. 39(1):102–114.
40. Liu, Ting, Charles, Rosenberg, and Henry A. Rowley. 2007. Clustering billions of images with
large scale nearest neighbor search. In Applications of Computer Vision, 2007. WACV’07. IEEE
Workshop on, pp. 28–28. IEEE.
41. Liu, Yongguo, Wu Xindong, and Yidong Shen. 2011. Automatic clustering using genetic algo-
rithms. Applied Mathematics and Computation 218 (4): 1267–1279.
42. Maulik, Ujjwal, and Sanghamitra Bandyopadhyay. 2000. Genetic algorithm-based clustering
technique. Pattern Recognition 33 (9): 1455–1465.
43. Mei, Jian-Ping, Yangtao Wang, Lihui Chen, and Chunyan Miao. 2017. Large scale document
categorization with fuzzy clustering. IEEE Transactions on Fuzzy Systems 25 (5): 1239–1251.
44. Mirjalili, Seyedali. 2015. Moth-flame optimization algorithm: A novel nature-inspired heuristic
paradigm. Knowledge-Based Systems 89: 228–249.
45. Mirjalili, Seyedali, and Andrew Lewis. 2016. The whale optimization algorithm. Advances in
Engineering Software 95: 51–67.
46. Mirjalili, Seyedali, Seyed Mohammad, Mirjalili, and Abdolreza, Hatamlou. 2016. Multi-verse
optimizer: a nature-inspired algorithm for global optimization. Neural Computing and Appli-
cations, 27(2):495–513.
47. Mirjalili, Seyedali, Seyed Mohammad, Mirjalili, and Andrew, Lewis. 2014. Grey wolf opti-
mizer. Advances in Engineering Software, 69:46–61.
48. Mittal, Nitin, Urvinder, Singh, and Balwinder Singh, Sohi. 2016. Modified grey wolf optimizer
for global engineering optimization. Applied Computational Intelligence and Soft Computing,
2016:8.
49. Mukhopadhyay, Anirban, Ujjwal, Maulik, Sanghamitra, Bandyopadhyay, and Carlos A. Coello.
2014. Survey of multiobjective evolutionary algorithms for data mining: Part ii. IEEE Trans-
actions on Evolutionary Computation, 18(1):20–35.
50. Mukhopadhyay, Anirban, Ujjwal Maulik, Sanghamitra Bandyopadhyay, and Carlos Artemio
Coello Coello. 2013. A survey of multiobjective evolutionary algorithms for data mining: Part
i. IEEE Transactions on Evolutionary Computation 18 (1): 4–19.
51. Satyasai Jagannath Nanda and Ganapati Panda. 2014. A survey on nature inspired metaheuristic
algorithms for partitional clustering. Swarm and Evolutionary computation 16: 1–18.
52. Oyelade, O.J. O.O. Oladipupo, and I.C. Obagbuwa. 2010. Application of k means clustering
algorithm for prediction of students academic performance. arXiv:1002.2425.
53. Qaddoura, R., H. Faris, and I. Aljarah. 2020. An efficient evolutionary algorithm with a near-
est neighbor search technique for clustering analysis. Journal of Ambient Intelligence and
Humanized Computing 1–26.
54. Qaddoura, R., H. Faris, I. Aljarah, J. Merelo, and P. Castillo. 2020. Empirical evaluation of
distance measures for nearest point with indexing ratio clustering algorithm. In Proceedings
A Grey Wolf-Based Clustering Algorithm for Medical Diagnosis Problems 87

of the 12th International Joint Conference on Computational Intelligence - Volume 1: NCTA,

ISBN 978-989-758-475-vol. 6, pp. 430–438. https://doi.org/10.5220/0010121504300438.
55. Qaddoura, Raneem, Waref, Al Manaseer, Mohammad A.M. Abushariah, and Mohammad Aref,
Alshraideh. 2020. Dental radiography segmentation using expectation-maximization clustering
and grasshopper optimizer. multimedia Tools and Applications.
56. Qaddoura, Raneem, Hossam Faris, and Ibrahim Aljarah. 2020. An efficient clustering algorithm
based on the k-nearest neighbors with an indexing ratio. International Journal of Machine
Learning and Cybernetics 11 (3): 675–714.
57. Qaddoura, Raneem, Hossam, Faris, Ibrahim, Aljarah, and Pedro A. Castillo. 2020. Evocluster:
An open-source nature-inspired optimization clustering framework in python. In International
Conference on the Applications of Evolutionary Computation (Part of EvoStar), pp. 20–36.
Berlin: Springer.
58. Sharma, Manorama, G.N. Purohit, and Saurabh, Mukherjee. 2018. Information retrieves from
brain mri images for tumor detection using hybrid technique k-means and artificial neural
network (kmann). In Networking Communication and Data Knowledge Engineering, pp. 145–
157. Berlin: Springer.
59. Sheikh, Rahila H., Mukesh M. Raghuwanshi, and Anil N. Jaiswal. 2008. Genetic algorithm
based clustering: a survey. In First International Conference on Emerging Trends in Engineer-
ing and Technology, pp. 314–319. IEEE.
60. Shukri, Sarah, Hossam Faris, Ibrahim Aljarah, Seyedali Mirjalili, and Ajith Abraham. 2018.
Evolutionary static and dynamic clustering algorithms based on multi-verse optimizer. Engi-
neering Applications of Artificial Intelligence 72: 54–66.
61. Siddiqi, Umair F., and Sadiq M. Sait. 2017. A new heuristic for the data clustering problem.
IEEE Access, 5:6801–6812.
62. Silva, Samuel, Rengan, Suresh, Feng, Tao, Johnathan, Votion, and Yongcan, Cao. 2017. A multi-
layer k-means approach for multi-sensor data pattern recognition in multi-target localization.
arXiv:1705.10757.
63. Storn, Rainer, and Kenneth Price. 1997. Differential evolution-a simple and efficient heuristic
for global optimization over continuous spaces. Journal of Global Optimization 11 (4): 341–
359.
64. Tripathi, Ashish Kumar, Kapil, Sharma, and Manju, Bala. 2018. A novel clustering method
using enhanced grey wolf optimizer and mapreduce. Big Data Research, 14:93–100.
65. Wang, X.Y., and Jon M. Garibaldi. 2005. A comparison of fuzzy and non-fuzzy clustering tech-
niques in cancer diagnosis. In Proceedings of the 2nd International Conference in Computa-
tional Intelligence in Medicine and Healthcare, BIOPATTERN Conference, Costa da Caparica,
Lisbon, Portugal, vol. 28.
66. Yadav, Ms Chandni, Ms Shrutika, Zele, Ms Tejashree, Patil, Ms Vishakha, Bombadi, and
Mr Tushar, Chaudhari. 2018. Automatic blood cancer detection using image processing. Cell,
4(03).
67. Yang, Hongguang, and Jiansheng Liu. 2015. A hybrid clustering algorithm based on grey wolf
optimizer and k-means algorithm. J Jiangxi Univ Sci Technol 5: 015.
68. Yang, Xin-She. 2009. Firefly algorithms for multimodal optimization. In International sympo-
sium on stochastic algorithms, pp. 169–178. Berlin: Springer.
69. Yang, Xin-She, and Suash, Deb. 2009. Cuckoo search via lévy flights. In 2009 World Congress
on Nature & Biologically Inspired Computing (NaBIC), pp. 210–214. IEEE.
70. Zhang, Sen, and Yongquan, Zhou. 2015. Grey wolf optimizer based on powell local optimiza-
tion method for clustering analysis. Discrete Dynamics in Nature and Society 2015.
EEG-Based Person Identification Using
Multi-Verse Optimizer as Unsupervised
Clustering Techniques

Zaid Abdi Alkareem Alyasseri, Ammar Kamal Abasi,

Mohammed Azmi Al-Betar, Sharif Naser Makhadmeh, João P. Papa,
Salwani Abdullah, and Ahamad Tajudin Khader

Abstract Recently, electroencephalogram (EEG) signal provides great potential for

identification systems. Many studies have shown that the EEG introduces unique,
universal features and natural robustness for spoofing attacks. The EEG represents
a graphic recording of the electrical activity of the brain that can be measured by
placing sensors (electrodes) at different locations on the scalp. This chapter proposes
a new technique using unsupervised clustering and optimization techniques for user
identification-based EEG signals. The proposed method employs four algorithms
which are Genetic Algorithm (GA), Multi-verse Optimizer (MVO), Particle Swarm
Optimization (PSO) and the k-means algorithm. A standard EEG motor imagery

Z. A. A. Alyasseri (B)
ECE Department, Faculty of Engineering, University of Kufa, Najaf, Iraq
e-mail: [email protected]
Z. A. A. Alyasseri · S. Abdullah
Faculty of Information Science and Technology, Center for Artificial Intelligence, Universiti
Kebangsaan Malaysia, 43600 Bangi, Selangor, Malaysia
e-mail: [email protected]
A. K. Abasi · S. N. Makhadmeh · A. T. Khader
School of Computer Sciences, Universiti Sains Malaysia, George Town, Pulau Pinang, Malaysia
e-mail: [email protected]
S. N. Makhadmeh
e-mail: [email protected]
A. T. Khader
e-mail: [email protected]
M. Al-Betar
Department of Information Technology - MSAI, College of Engineering and Information
Technology, Ajman University, Ajman, United Arab Emirates
e-mail: [email protected]
IT Department, Al-Huson University College, Al-Balqa Applied University, Irbid, Jordan
J. P. Papa
Department of Computing, São Paulo State University - UNESP, Bauru, Brazil
e-mail: [email protected]

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer 89
Nature Singapore Pte Ltd. 2021
I. Aljarah et al. (eds.), Evolutionary Data Clustering: Algorithms and Applications, Algo-
rithms for Intelligent Systems,
https://doi.org/10.1007/978-981-33-4191-3_4
90 Z. A. A. Alyasseri et al.

dataset is used to evaluate the proposed method’s performance, and its results are
evaluated using four criteria: (i) Precision, (ii) Recall, (iii) F-Score and (v) Purity. It
is worth mentioning that this work is one of the first to employ optimization methods
with unsupervised clustering methods for person identification using EEG. As a
conclusion, the MVO algorithm achieved the best results compared with GA, PSO
and k-means. Finally, the proposed method can draw future directions to apply to
different research areas.

Keywords EEG · Biometric · Multi-verse optimizer · Evolutionary computation ·

Meta-heuristics algorithms · Clustering · Auto-regressive

1 Introduction

Electroencephalogram (EEG) is a visualization of brain neural activity measured

from scalp that reflects the variations of the voltage that originate from ionic currents
in the brain’s neurons [14, 18, 39, 53]. EEG signals could therefore produce most
of the brain activity information needed and be collected through non-invasive or
invasive methods [10, 19, 48]. The key distinction between these approaches is that
the intrusive technique entails the deployment of electrode sets embedded in the brain,
for example, an eastern group-brain oncology computer interface for coordination
of arm movement [49]. Berger [24] introduced the first non-invasive method for
recording human brain activities using brain signals (EEG).
Researchers have expanded the technique from Han’s for flexible applications in
the last few decades. For instance, brain signals are successfully applied in medical
for many purposes such as to treat, rehabilitate and re-establish patients. The EEG was
also utilized in non-medical purposes, for example, self-regulation and education,
the neuromarketing and advertising sector as well as for the neuroergonomics and
the smart environment [6, 36]. EEG signals have effectively been utilized as a new
identification and authentication technique for security and authorization purposes [6,
17, 19, 38, 39]. The researchers concluded that EEG signals are distinctive and
powerful that are also hard to fudge, and also, recognize as a latest biometric technique
[48, 50].
Asanza et al. [21] introduced a new EEG pre-processing method for noise removal.
The signals were collected from two electrodes which are Left Occipital (LO)
and Right Occipital (RO) using the EMOTIV headset with 14 channels. The pro-
posed method used the k-means algorithm for clustering purposes right after fea-
ture extraction using frequency and temporal characteristics. Lit later on, the same
authors applied five different clustering algorithms, say that k-means, Hierarchical
k-medoids, BDSCAN, and Spectral to detect motor and imaginary motor tasks [20].
They used a standard EEG dataset with 25 healthy persons using 64 channels. In
the feature extraction phase, they extracted spectral density, being the best results
for detecting motor activity obtained by k-means and k-medoids, while hierarchical
clustering achieved the best results for imaginary tasks of hands.
EEG-Based Person Identification Using Multi-Verse Optimizer … 91

Recently, there have been major development relative with traditional techniques
for the EEG-based user identification with the supervised classification and opti-
mization methods [11, 15]. Zaid Alyaseri has developed a novel technique for the
detection of users based on the EEG signal [17]. This method used a novel technique
called MOFPA-WT. The MOFPA-WT method used a multi-objective FPA and the
Wavelet to derive EEG features. Several variations of EEG energy features from the
EEG sub-bands have been extracted. Later, this work has been extended with more
features extracted from the EEG using WT decomposition process to 10 levels [18]. A
number of measures such as accuracy, true acceptance rate, false acceptance rate and
F-score were used to test the proposed method. The MOFPA-WT method extracted
several time-main features such as mean, entropy, standard deviation, energy the log-
arithm of the energy, absolute energy and Resting Energy Expenditure (REE) [16].
The performance results were evaluated using accuracy, sensitivity, specificity, false
acceptance rate and F-Ore. The MOFPA-WT method was comparable to some edge-
cutting techniques with prospective outcomes. Moreover, the optimization technique
has shown successfully significant in the field of EEG-based user identification.
The literature refers that there are a few works which are applying unsupervised
clustering techniques with for EEG-based biometric person identification. The main
reason for that is because most researchers prefer to use the classification methods to
obtain good results compared to the methods of clustering, which does not contain a
training phase, but the new data is determined to its cluster by calculating the distance
between the centre of clusters and then the new data will join the closest one. Recently,
the metaheuristic-based algorithms are utilized for EEG, they are normally classify
into three main categories: Trajectory-based Algorithms (TAs), Evolutionary-based
Algorithms (EAs) and Swarm Intelligence (SI) [15]. In general, all these types of
metaheuristic algorithms are nature-inspired with common characteristics such as
stochastic behaviours, set more than one control parameter to deal with the problem
[32, 40].
Recently, the metaheuristic-based algorithms are utilized for EEG, they are nor-
mally classified into three main categories: Trajectory-based Algorithms (TAs),
Evolutionary-based Algorithms (EAs) and Swarm Intelligence (SI) [13, 42]. In gen-
eral, all these types of metaheuristic algorithms are nature-inspired with common
characteristics such as stochastic behaviours, set more than one control parameter to
deal with the problem.
The first type of metaheuristic algorithms is TAs; in the initial phase, the TAs are
started with a single solution (one solution), and based on the neighbouring-moves
operators, this initial solution is then evolved, or moved over the generations until a
local optimal solution, which is the best region in the search space is reached from
these algorithms. An example of TAs is Self-Organizing Maps (SOMs) and β-hill
climbing [7, 41]. In the same context, there are techniques which are related with
the problem nature (i.e. clustering), these techniques utilized the same mechanism of
evolving the solution during the search spaces like K-means, K-medoids. Although
TAs can deeply search the search space region of the initial solution and reach local
optima, they cannot navigate several search space regions simultaneously.
92 Z. A. A. Alyasseri et al.

The second type of metaheuristic algorithms is EA, which is initiated with a group
of provisional individuals called population. Generation after generation, the pop-
ulation is evolved on the basis of three main operators: recombination for mixing
the individual features, mutation for diversifying the search and selection for util-
ising the survival-of-the-fittest principle [12]. The EA is stopped when no further
evolution can be achieved. The main shortcoming of EAs is that although they can
simultaneously navigate several areas in the search space, they cannot perform deep
searching in each area to which they navigate. Consequently, EAs mostly suffer from
premature convergence. EAs that have been successfully utilized for TDC include
Genetic Algorithm (GA) [34], harmony search [30] and cuckoo search [55].
The last type of metaheuristic algorithms is SI; an SI algorithm is also initiated
with a set of random solutions called a swarm. Iteration after iteration, the solutions
in the swarm are reconstructed by means of attracting them by the best solutions that
are so far found [44]. SI-based algorithms can easily converge prematurely. Several
SI-based TDC are utilized, such as Particle Swarm Optimization (PSO) [25, 43],
Grey Wolf Optimizer (GWO) [3] and artificial bee colony [35].
The Multi-Verse Optimizer (MVO) algorithm was recently proposed as a stochas-
tic population-based algorithm [45] inspired by multi-verse theory [23]. The big bang
theory [37] explains the origin of the universe to have been a massive explosion.
According to this theory, the origin of everything in our universe requires one big
bang. Multi-verse theory believes that more than one explosion (big bang) occurred,
with each big bang creating a new and independent universe. This theory is modelled
as an optimization algorithm with three concepts: white hole, black hole and worm-
hole, for performing exploration, exploitation and local search, respectively. The
MVO algorithm presents many benefits over other algorithms. Just a few parameters
are needed to be defined at the initial stage; no complex data mathematical derivation
is possible. During the search, it can easily balance exploration with exploitation. It
is also sound-and-complete, flexible, scalable, adaptable and simple. MVO has also
been used for a number of optimization problems, such as optimising SVM parame-
ters [8, 27], oil recovery [33], feature selection [4, 26] and text document clustering
[1, 2, 5].
However, the literature refers that there are a few works that applied clustering
techniques for EEG-based biometric person identification. Therefore, the main con-
tributions of this paper are two-fold: (i) to introduce a hybrid approach composed
of both unsupervised classification and optimization techniques for EEG-based bio-
metric person identification; (ii) to evaluate the performance of some metaheuristic
algorithms such as MVO, Genetic Algorithm (GA) and PSO together with k-means
for EEG-based biometric person identification. The performance of the proposed is
evaluated regarding four measurement factors: precision, recall, F-Score and purity.
Our work includes 109 EEG signal respondents recorded from 64 channels on var-
ious cognitive tasks. In the extraction process, auto-regressive are derived from the
original EEG signals with three separate coefficients (i.e. 5, 10 and 20 orders).
The organizer of this chapter is organized as follows. Section 2 provides a Prelim-
inaries of the techniques considered in this work. Section 3 describes the proposed
EEG-Based Person Identification Using Multi-Verse Optimizer … 93

approach, and the results are discussed in Sect. 4. Finally, the conclusion and future
works are stated in Sect. 5.

2 Preliminaries

In this section, we provided a brief background about the k-means, particle swarm
optimization, genetic algorithm and multi-verse optimizer algorithms which will use
in this work.

2.1 k-Means

The well-known k-means is a widely and simple utilized technique in clustering

domain [29, 54]. The idea behind k-means clustering is to group the samples into fixed
k clusters using their similarities. The Euclidean distance is one of the commonly used
measures to determine such a degree of similarity. Also, the clusters are represented
by their centroids. In each round of iteration, these centroids are recalculated and
used as reference points to group samples around their closest k clusters [22]. The
main idea of k-means is implemented in the following Algorithm 1

Algorithm 1 clustering method using k-means algorithm

1: Inputs: A reading t number of samples, and a k of clusters.
2: Outputs: Assign t to k .
3: determine an initial trials group with k clusters.
4: while achieved the final criterion or reach the stop condition do
5: for each K clusters compute new cluster centroid.
6: Generate a new groups by assigning each trial to its closest cluster centroid based on distance(trial; cluster centroid).
7: end while

2.2 Principles for Multi-Verse Optimizer (MVO)

The general information on MVO is provided in this subsection. Multi-version theory

and mathematical modelling as well as inspiration.

2.2.1 Inspiration

MVO is a population-based algorithm [45] inspired by nature phenomena (multi-

verse theory) [23]. According to multi-verse theory, universes connect and might even
collide with each other. MVO engages and reformulates using three main concepts:
white holes, black holes and wormholes. The probability is used for determining the
inflation rate (corresponds to the objective function in the optimization context) for
94 Z. A. A. Alyasseri et al.

the universe, thereby allowing the universe to assign one of these holes. Given that the
universe has a high inflation rate, the probability of a white hole existing increases.
Meanwhile, a low inflation rate leads to increased probability of a black hole existing
[45]. Regardless of the universe’s inflation rate, wormholes move objects towards the
best universe randomly [28, 33]. In the optimization of terms, the white and black
hole concepts play a central role in managing the exploration phase. On another side,
the wormhole concept ensures the phase of exploitation. In multi-verse theory, each
universe corresponds to a solution in optimization theory. In that solution, each object
in the universe corresponds to one of the decision variables. The multi-verse theory
is the main inspiration of MVO, which states that more than one universe exists and
the universes can interact with each other (i.e. exchange the objects) by three main
concepts: (i) the white hole, (ii) the black hole and (iii) the wormhole. Each one of
these concepts plays a different role in the MVO working mechanism [47].
Like any population-based optimization algorithms, MVO starts with a population
of feasible solutions, and it has been designed to improve solutions by mapping them
to different values based on their fitness function. In MVO, the decision variables
may face some exchange between the solutions randomly regardless of the fitness
function to keep population diversity and to skip the local optimum [28].

2.2.2 Mathematical Modelling of MVO

The black and white hole concepts in MVO are formulated for exploring search
spaces, and the wormhole concept is formulated for exploiting search spaces. In
other EAs, MVO is initiated by a population of individuals (universes). Thereafter,
MVO improves these solutions until a stopping criterion. Figure 1 illustrates the
conceptual model of the MVO and shows the movements of the objects between the
universes via white/black hole tunnels. These hole tunnels are created between two
universes on the basis of the inflation rate of each universe (i.e. one universe has a
higher inflation rate than the other universes.). Objects move from universes with
high inflation rates using white holes. These objects are received by universes with
low inflation rates using black holes.
After a population of solutions is initiated, all solutions in MVO are sorted from
high inflation rates to low ones. Thereafter, it visits the solutions one by one to attract
these solutions to the best one. This is done under the assumption that the solution
that has been visited has the black hole. As for the white holes, the roulette wheel
mechanism is used for selecting one solution. The formulation of the population U
is provided in Eq. (1).

⎡ ⎤
Sol11 Sol12 · · · Sol1d
⎢ Sol21 Sol22 · · · Sol2d ⎥
⎢ ⎥
U=⎢ . .. .. .. ⎥ . (1)
⎣ .. . . . ⎦
Soln Soln · · · Solnd
1 2
EEG-Based Person Identification Using Multi-Verse Optimizer … 95

Fig. 1 Conceptual model of the MVO algorithm I (U1 ) > I (U2 > · · · > I (Un )

where n is the number of solutions (i.e. candidate universes), d is the number of deci-
sion variables (i.e. objects) and U is the population matrix of size n × d containing
the set of universes. In solution i, the object j is generated randomly as follows:

j
Soli = lb j + rand()%((ub j − lb j ) + 1) ∀i ∈ (1, 2, . . . , n) ∧ ∀ j ∈ (1, 2, . . . , d),
(2)
where rand() is a function generating a discrete random number of the (1, 2, . . . ,
M AX _I N T ) distribution, M AX _I N T , where M AX _I N T is the maximum inte-
ger number which the machine can produce, and [lb j , ub j ] signifies the j object’s
discrete lower and upper limits.
In each iteration, in solution j, which has the black hole, the decision vari-
j
able i (i.e. Soli j ) can exchange the value from better solutions such that Soli ∈
j−1
(Soli1 , Soli2 , . . . , Soli ) or the value remains unchanged. This is formulated as
shown in Eq. (3).

j
j Solk z 1 < N oI (Ui ),
Soli = j (3)
Soli z 1 ≥ N oI (Ui ),
j
where z 1 is function that generates a uniform random number between (0, 1), Solk
represent the jth decision variable of kth solution selected by a roulette wheel selec-
j
tion, Soli represent the jth decision variable of the i th solution and N oI (Ui ) is
normalized objective function of i th solution. At the same time, regardless of the
fitness function, the decision variables of the solution j perform random movements
in respect of the optimal value for solutions diversity improvement in the MVO
algorithm. This method is modelled using the following formula:
96 Z. A. A. Alyasseri et al.

Fig. 2 Flowchart of MVO algorithm

⎧
⎪
⎨ Sol j + T D R × ((ub j − lb j ) × z 4 + lb j ) z 3 < 0.5,
j z2 < P W E
Soli = Sol j − T D R × ((ub j − lb j ) × z 4 + lb j ) z 3 ≥ 0.5, (4)
⎪
⎩ j
Soli r2 ≥ P W E

where Sol j represents the jth location of the best universe in the population; (T D R)
refers to the travelling distance rate and (P W E) refers to wormhole existence proba-
bility which are coefficient parameters; ubj and lbj are the upper and lower bounds,
respectively; z2, z3 and z4 are random value (0, 1). The formulas for T D R and
P W E are as follows:

max − min
P W E = min + l × , (5)
L

l 1/ p
T DR = 1 − , (6)
L 1/ p
where min and max are constant pre-defined values, l represents the current iteration,
L is the maximum number of iterations and p is a constant that denotes the accuracy of
exploitation over the iterations. The P W E coefficient values are smoothly increased
during the iterations for increasing the attendance chance of wormholes in universes.
Thus, the exploitation phase is stressed in every iteration. At the same time, T D R
coefficient values decrease the distance of decision variables around the best solution.
Therefore, the accuracy of local search is improved. A flowchart of the optimization
procedure of MVO is shown in Fig. 2.
Algorithm 2 provides the pseudocodes of the MVO algorithm.
EEG-Based Person Identification Using Multi-Verse Optimizer … 97

Algorithm 2 Pseudocodes and general steps of MVO algorithm

1: Initialize MVO parameters (Min, Max, Best solution = 0, LB, UB, Number of solutions, Number of dimensions
(Dim), Number of iterations).
2: Create random solutions (Sol) based on LB, UB, Number of solutions, Dim.
3: while achieved the final criterion or reach the stop condition do
4: Evaluate the objective function for every solutions.
5: Sort_s=Ranking Solutions using fitness value.
6: Normalize=Normalize the objective function for all solutions.
7: Update the Best solution vector.
8: for each solutions indexed by i except the best solution do
9: Calculate PWE and TDR using Equations (5 and 6).
10: Black_hole_index=i.
11: for each object indexed by j do
12: z1 = random([0, 1]).
13: if z1 < N or mali ze(Si ) then
14: White_hole_index = Roulette_Wheel_Selection .
15: S(Black_hole_index,j)= Sort_s(White_hole_index,j).
16: end if
17: z2 = random([0, 1]).
18: if z2 < W or mhole_existance_ pr obabilit y then
19: z3 = random([0, 1]).
20: z4 = random([0, 1]).
21: if z3 < 0.5 then
22: Update the Position of Solution using Equation 4 case one.
23: else
24: Update the Position of Solution using Equation 4 case two.
25: end if
26: end if
27:
28: end for
29: end for
30: end while
31: Produce the optimal Solution.

3 Hybridizing EEG with MVO for Unsupervised Person

Identification: Proposed Method

In this section, independent EEG-based user recognition method is proposed. The

method suggested inclusive of five steps in which the output of each step is used to
instruct the next step. The proposed method, which is described , is shown in Fig. 3.

3.1 Signal Acquisition

A standard EEG signal data set is used to acquire the EEG signal [31]. The EEG
signals were obtained from 109 healthy people using a BCl2000 application pro-
gram [52]. EEG signals from 64 electrodes (i.e. sensors) are then captured. Each
subject runs 14 motor/imagery tasks, including neurological rehabilitation and brain-
computer interface applications. Such activities are usually to visualize or imitate an
action, such as eye opening and eye closing. The EEG signal input for each user, with
98 Z. A. A. Alyasseri et al.

Fig. 3 EEG-based user identification system proposed in this work

EEG-Based Person Identification Using Multi-Verse Optimizer … 99

Fig. 4 Electrodes
distribution for motor
movement/imagery dataset

a length of one minute, hereby collected three times. Figure 4 shows the distribution
of the EEG channels which are used in this work.

3.2 Signal Pre-processing

The input EEG signal is split into 6 segments of 10s each. We have used a bandpass
and notch filter to de-noise, as the EEG signal can compromise throughout recording
[53].

3.3 EEG Feature Extraction

Effective features are important to extract in every authentication system [51,

53]. This phase consequently focuses on extracting unique information from auto-
regressive models with three different coefficients, i.e. 5, 10 and 20 orders that allow
the clustering phase to achieve good results.1 In this chapter, we applied the Yule-
Walker approach to find the coefficients of the Auto-Regressive (AR) model by using
the least square method criterion.

1 Such numbers are suggested in the work of Rodrigues et al. [9, 50].
100 Z. A. A. Alyasseri et al.

3.4 EEG Signal Clustering

The main role of any clustering method is gathering similar samples in the same
cluster, and dissimilar ones in a different cluster. In this paper, different optimization
techniques are considered to optimize k-means algorithm in the context of EEG-
based person identification, as explained further.

3.4.1 Representation of Solutions

In the proposed approach, each possible solution is represented as vector s ∈ n ,

where n is number of dataset samples. The value of each element si falls within the
range [1, k], where k is any cluster number. Figure 5 shows an illustrative example of
the proposed solution representation with the steps to obtain the optimal/near-optimal
solution.

3.4.2 Fitness Function

As stated by Papa et al. [46], the k-means working mechanism is essentially an

optimization problem, where the idea is to minimize the distance from each dataset
sample to its closest centroid. Therefore, the fitness function used to evaluate each
solution can be computed by the following steps:

1. calculate the cluster centroid for each cluster;

2. calculate the distance between each sample (task) and the cluster centroid asso-
ciated with it; and
3. calculate average distance of samples to the cluster centroid (ADDC).

To calculate the clusters’ centroids, we can use the following formula:

n
si
C j = n i=1 , (7)
i=1 N ji

where C j denotes the centroid of cluster j, j ∈ {1, 2, . . . , k}, n is the number of

dataset samples and Nn×k is a binary-valued matrix defined as follows:

1 sample (task) si belongs to the cluster j
Nki = (8)
0 otherwise.

We used the average distance of samples to the cluster centroid measure, which
is defined as follows:
k  1  
i=1 n i ∀s j ∈Ci D(C i , s j )
AD DC = , (9)
k
EEG-Based Person Identification Using Multi-Verse Optimizer … 101

Fig. 5 Proposed approach for unsupervised EEG-based person identification using metaheuristic
optimization

where D(C j , s j ) denotes the distance between the centroid of cluster j and sample
s j . Therefore, the main idea of this work is to minimize the value of ADDC for a
given possible solution, i.e. to associate the “best cluster” to each dataset sample.

4 Results and Discussions

In this section, we present the evaluation measures, results and the discussion about
the findings of our work.
102 Z. A. A. Alyasseri et al.

4.1 Evaluation Measures

In order to evaluate the proposed approach, four measures are used Purity, Recall,
Precision and F-measure, which can be defined as follows:
• Purity: This measure computes the maximum number of correct class for each
cluster over all tasks in the cluster [54], and it can be computed as follows:

1
k
Purit y = max(i, j), (10)
n i=1

where max(i, j) stands for the maximum number of correct label assignment (i.e.
correct classification) from class i in cluster j.
• Precision: Such measure computes the ratio of correct label assignment from class
i over the total number of tasks in cluster j [54]:
L i, j
P(i, j) =   , (11)
C j 

where
 L i j is the number of tasks from class i correctly identified in cluster j, and
C j  is the numbers of tasks (samples) in cluster j.
• Recall: Such measure computes the ratio of tasks from cluster j correctly identified
(i.e. L̂ i, j ) against the total numbers of tasks (samples) from the class i:

L̂ i, j
R(i, j) = , (12)
Ti

where Ti is total number of tasks (samples) from class i.

• F-score: It is combing the precision and recall defines as the F-score, which tells
us the clustering quality. The larger its values, the better is the solution:
2 × P(i, j) × R(i, j)
F(i, j) = . (13)
P(i, j) + R(i, j)

The overall F-score is computed by taking the weighted F-score for each class.
Table 1 shows the experiment results of 30 runs concerning standard k-means and
its versions optimized by MVO, GA and PSO. The results are presented in terms of
the best, average and worst fitness values. Besides, the MVO algorithm achieved the
best results according to all measures, i.e. Purity, Recall, Precision and F-measure. In
the best case, the MVO obtained 0.9643 with A R5 (5-order auto-regressive model),
0.9740 with A R10 (10-order auto-regressive model) and 0.9804 with A R20 (20-order
auto-regressive model). For the GA obtained 0.8098 with A R5 , 0.8145 with A R10
and 0.8179 with A R20 . For the PSO obtained 0.8128 with A R5 , 0.8158 with A R10
and 0.8158 with A R20 where the k-means obtained 0.8074 with A R5 , 0.8180 with
A R10 and 0.6884 with A R20 .
EEG-Based Person Identification Using Multi-Verse Optimizer … 103

Fig. 6 Precision measure of the auto-regression model with 5 coefficients

Fig. 7 F-measure of the auto-regression model with 5 coefficients

Such results are pretty very interesting, and it refers that the MVO algorithm is
suitable for EEG biometric applications.
Figures 6, 7, 8 and 9 show the performance of the proposed methods. The results
show the MOV obtained a great significant advantage using all criteria, while GA
and PSO have closer results. Notice that k-means has achieved a fluctuation change
in the results.
104 Z. A. A. Alyasseri et al.

Table 1 Experimental results

Measures Method/ AR5 MVO k -means PSO GA
Precision Best 0.9643 0.8074 0.8128 0.8098
Recall 0.9665 0.8487 0.8324 0.8283
F-Score 0.9654 0.8275 0.8225 0.8189
Purity 0.9643 0.8074 0.8128 0.8098
Measures Method/ AR5 MVO k -means PSO GA
Precision Average 0.9471 0.5697 0.7836 0.7816
Recall 0.9499 0.6007 0.8047 0.8021
F-Score 0.9485 0.5845 0.7940 0.7917
Purity 0.9471 0.7088 0.7836 0.7816
Measures Method/ AR5 MVO k -means PSO GA
Precision Worst 0.9199 0.2595 0.7255 0.7276
Recall 0.9228 0.2520 0.7504 0.7517
F-Score 0.9214 0.2557 0.7377 0.7394
Purity 0.9199 0.6334 0.7255 0.7276
Measures Method/ AR10 MVO k -means PSO GA
Precision Best 0.9740 0.8180 0.8158 0.8145
Recall 0.9751 0.8661 0.8335 0.8328
F-Score 0.9746 0.8414 0.8246 0.8236
Purity 0.9740 0.8503 0.8158 0.8145
Measures Method/ AR10 MVO k -means PSO GA
Precision Average 0.9562 0.5582 0.7850 0.7878
Recall 0.9578 0.5872 0.8050 0.8081
F-Score 0.9570 0.5719 0.79492 0.7978
Purity 0.9562 0.7186 0.7850 0.7878
Measures Method/ AR10 MVO k -means PSO GA
Precision Worst 0.9170 0.2561 0.7570 0.7507
Recall 0.9197 0.2622 0.7791 0.7744
F-Score 0.9184 0.2591 0.7679 0.7624
Purity 0.9170 0.6123 0.7570 0.7507
Measures Method/ AR20 MVO k -means PSO GA
Precision Best 0.9804 0.6884 0.8158 0.8179
Recall 0.9811 0.7096 0.8335 0.8345
F-Score 0.9807 0.6988 0.8246 0.8261
Purity 0.9804 0.8101 0.8158 0.8179
Measures Method/ AR20 MVO k -means PSO GA
Precision Average 0.9575 0.4578 0.7836 0.7919
Recall 0.9588 0.4608 0.8037 0.8113
F-Score 0.9581 0.4587 0.7935 0.8015
Purity 0.9575 0.6964 0.7836 0.7919
Measures Method/ AR20 MVO k -means PSO GA
Precision Worst 0.9309 0.1501 0.7570 0.7652
Recall 0.9324 0.1414 0.7791 0.7863
F-Score 0.9317 0.1534 0.7679 0.7756
Purity 0.9309 0.6126 0.7570 0.7652
Bold value indicates the best value
EEG-Based Person Identification Using Multi-Verse Optimizer … 105

Fig. 8 Recall measure of the auto-regression model with 5 coefficients

Fig. 9 Purity measure of the auto-regression model with 5 coefficients

Figure 10 summarizes the average performance over A R5 , A R10 and A R20 con-
figurations where the MVO achieved the best results according to Purity, Recall,
Precision and F-measure.
106 Z. A. A. Alyasseri et al.

Fig. 10 Average Performance over: a AR5, b AR10, and c AR20 configurations

EEG-Based Person Identification Using Multi-Verse Optimizer … 107

5 Conclusions and Future Work

In this chapter, a new method for unsupervised EEG-based user identification was
proposed. The main purpose of the proposed approach is to find the optimal solution
of the original brain EEG signal which can give unique features from auto-regressive
models with three different coefficients (AR5, Ar10 and AR20). In general, the idea
is to model the k-means working mechanism as an optimization process, where
the samples are associated with their closest centroids. In this context, GA, PSO
and MVO were considered to optimize k-means for EEG person identification. The
proposed method was tested using EEG motor imagery which is a standard EEG
dataset. The results of the proposed method are evaluated using four criteria, namely,
precision, recall, F-score and purity.
The best results obtained by the MVO for all the evaluation measures are the PSO,
GA and k-means, respectively.
For the future works, we intend to investigate another unsupervised techniques
for EEG-based user identification with more challenging and more complex EEG
dataset. Also, using multi-objective signal problem instances, such as user authenti-
cation or early detection of epilepsy based on EEG signals.

References

1. Abasi, Ammar Kamal, Ahamad Tajudin Khader, Mohammed Azmi Al-Betar, Syibrah Naim,
Zaid Abdi Alkareem Alyasseri, and Sharif Naser Makhadmeh. 2020. An ensemble topic extrac-
tion approach based on optimization clusters using hybrid multi-verse optimizer for scientific
publications. Journal of Ambient Intelligence and Humanized Computing 1–37.
2. Abasi, Ammar Kamal, Ahamad Tajudin Khader, Mohammed Azmi Al-Betar, Syibrah Naim,
Zaid Abdi Alkareem Alyasseri, and Sharif Naser Makhadmeh. 2020. A novel hybrid multi-verse
optimizer with k-means for text documents clustering. Neural Computing & Applications.
3. Abasi, Ammar Kamal, Ahamad Tajudin Khader, Mohammed Azmi Al-Betar, Syibrah Naim,
Sharif Naser Makhadmeh, and Zaid Abdi Alkareem Alyasseri. 2019. An improved text feature
selection for clustering using binary grey wolf optimizer. In Proceedings of the 11th national
technical seminar on unmanned system technology 2019, 503–516. Springer.
4. Abasi, Ammar Kamal, Ahamad Tajudin Khader, Mohammed Azmi Al-Betar, Syibrah Naim,
Sharif Naser Makhadmeh, and Zaid Abdi Alkareem Alyasseri. 2019. A text feature selection
technique based on binary multi-verse optimizer for text clustering. In 2019 IEEE Jordan
international joint conference on electrical engineering and information technology (JEEIT),
1–6. IEEE.
5. Abasi, Ammar Kamal, Ahamad Tajudin Khader, Mohammed Azmi Al-Betar, Syibrah Naim,
Sharif Naser Makhadmeh, and Zaid Abdi Alkareem Alyasseri. 2020. Link-based multi-verse
optimizer for text documents clustering. Applied Soft Computing 87: 106002.
6. Abdulkader, Sarah N., Ayman Atia, and Mostafa-Sami M. Mostafa. 2015. Brain computer
interfacing: Applications and challenges. Egyptian Informatics Journal 16 (2): 213–230.
7. Al-Betar, Mohammed Azmi. 2017. β-hill climbing: An exploratory local search. Neural Com-
puting and Applications 28 (1): 153–168.
8. Aljarah, Ibrahim, Al-Zoubi Ala’M, Hossam Faris, Mohammad A. Hassonah, Seyedali Mir-
jalili, and Heba Saadeh. 2018. Simultaneous feature selection and support vector machine
108 Z. A. A. Alyasseri et al.

optimization using the grasshopper optimization algorithm. Cognitive Computation 10 (3):

478–495.
9. Alyasseri, Zaid Abdi Alkareem, Ahamad Tajudin Khadeer, Mohammed Azmi Al-Betar, Ammar
Abasi, Sharif Makhadmeh, and Nabeel Salih Ali. 2019. The effects of EEG feature extraction
using multi-wavelet decomposition for mental tasks classification. In Proceedings of the inter-
national conference on information and communication technology, 139–146.
10. Alyasseri, Zaid Abdi Alkareem, Ahamad Tajudin Khader, and Mohammed Azmi Al-Betar.
2017. Electroencephalogram signals denoising using various mother wavelet functions: A com-
parative analysis. In Proceedings of the international conference on imaging, signal processing
and communication, 100–105. ACM.
11. Alyasseri, Zaid Abdi Alkareem, Ahamad Tajudin Khader, and Mohammed Azmi Al-Betar.
2017. Optimal electroencephalogram signals denoising using hybrid β-hill climbing algorithm
and wavelet transform. In Proceedings of the international conference on imaging, signal
processing and communication, 106–112. ACM.
12. Alyasseri, Zaid Abdi Alkareem, Ahamad Tajudin Khader, Mohammed Azmi Al-Betar,
Ammar Kamal Abasi, and Sharif Naser Makhadmeh. 2019. EEG signal denoising using
hybridizing method between wavelet transform with genetic algorithm. In Proceedings of the
11th national technical seminar on unmanned system technology 2019, 449–469. Springer.
13. Alyasseri, Zaid Abdi Alkareem, Ahamad Tajudin Khader, Mohammed Azmi Al-Betar, Ammar
Kamal Abasi, and Sharif Naser Makhadmeh. 2019. EEG signals denoising using optimal
wavelet transform hybridized with efficient metaheuristic methods. IEEE Access 8: 10584–
10605.
14. Alyasseri, Zaid Abdi Alkareem, Ahamad Tajudin Khader, Mohammed Azmi Al-Betar, and
Osama Ahmad Alomari. 2020. Person identification using EEG channel selection with hybrid
flower pollination algorithm. Pattern Recognition 107393.
15. Alyasseri, Zaid Abdi Alkareem, Ahamad Tajudin Khader, Mohammed Azmi Al-Betar, and
Mohammed A. Awadallah. 2018. Hybridizing β-hill climbing with wavelet transform for
denoising ECG signals. Information Sciences 429: 229–246.
16. Alyasseri, Zaid Abdi Alkareem, Ahamad Tajudin Khader, Mohammed Azmi Al-Betar, J.P.
Papa, Osama Ahmad Alomari, and Sharif Naser Makhadme. 2018. An efficient optimization
technique of EEG decomposition for user authentication system. In 2018 2nd international
conference on biosignal analysis, processing and systems (ICBAPS), 1–6. IEEE.
17. Alyasseri, Zaid Abdi Alkareem, Ahamad Tajudin Khader, Mohammed Azmi Al-Betar, João P.
Papa, and Osama ahmad Alomari. EEG-based person authentication using multi-objective
flower pollination algorithm. In 2018 IEEE congress on evolutionary computation (CEC),
pages 1–8. IEEE, 2018.
18. Alyasseri, Zaid Abdi Alkareem, Ahamad Tajudin Khader, Mohammed Azmi Al-Betar, João P.
Papa, and Osama Ahmad Alomari. 2018. EEG feature extraction for person identification using
wavelet decomposition and multi-objective flower pollination algorithm. IEEE Access.
19. Alyasseri, Zaid Abdi Alkareem, Ahmad Tajudin Khader, Mohammed Azmi Al-Betar, Joao P.
Papa, Osama Ahmad Alomari, and Sharif Naser Makhadmeh. 2018. Classification of EEG
mental tasks using multi-objective flower pollination algorithm for person identification. Inter-
national Journal of Integrated Engineering 10 (7).
20. Asanza, Víctor, Kerly Ochoa, Christian Sacarelo, Carlos Salazar, Francis Loayza, Carmen
Vaca, and Enrique Peláez. 2016. Clustering of EEG occipital signals using k-means. In Ecuador
technical chapters meeting (ETCM), IEEE, 1–5. IEEE.
21. Asanza, Víctor, Enrique Pelaez, and Francis Loayza. 2017. EEG signal clustering for motor
and imaginary motor tasks on hands and feet. In Ecuador technical chapters meeting (ETCM),
2017 IEEE, 1–5. IEEE.
22. Bai, Liang, Xueqi Cheng, Jiye Liang, Huawei Shen, and Yike Guo. 2017. Fast density clustering
strategies based on the k-means algorithm. Pattern Recognition 71: 375–386.
23. Barrow, John D., Paul C.W. Davies, and Charles L. Harper Jr. 2004. Science and ultimate
reality: Quantum theory, cosmology, and complexity. Cambridge University Press.
EEG-Based Person Identification Using Multi-Verse Optimizer … 109

24. Berger, Hans. 1929. Über das elektrenkephalogramm des menschen. European Archives of
Psychiatry and Clinical Neuroscience 87 (1): 527–570.
25. Cura, Tunchan. 2012. A particle swarm optimization approach to clustering. Expert Systems
with Applications 39 (1): 1582–1588.
26. Ewees, Ahmed A., Mohamed Abd El Aziz, and Aboul Ella Hassanien. 2017. Chaotic multi-
verse optimizer-based feature selection. Neural Computing and Applications 1–16.
27. Faris, Hossam, Mohammad A. Hassonah, Al-Zoubi Ala’M, Seyedali Mirjalili, and Ibrahim
Aljarah. 2017. A multi-verse optimizer approach for feature selection and optimizing SVM
parameters based on a robust system architecture. Neural Computing and Applications 1–15.
28. Fathy, Ahmed, and Hegazy Rezk. 2018. Multi-verse optimizer for identifying the optimal
parameters of PEMFC model. Energy 143: 634–644.
29. Fränti, Pasi, and Sami Sieranoja. 2017. K-means properties on six clustering benchmark
datasets. Applied Intelligence 1–17.
30. Geem, Zong Woo, Joong Hoon Kim, and Gobichettipalayam Vasudevan Loganathan. 2001. A
new heuristic optimization algorithm: Harmony search. Simulation 76 (2): 60–68.
31. Goldberger, Ary L., Luis A.N. Amaral, Leon Glass, Jeffrey M. Hausdorff, Plamen Ch. Ivanov,
Roger G. Mark, Joseph E. Mietus, George B. Moody, Chung-Kang Peng, and H. Eugene
Stanley. 2000. Physiobank, physiotoolkit, and physionet. Circulation 101 (23): e215–e220.
32. Hussain, Kashif, Mohd Najib Mohd Salleh, Shi Cheng, and Yuhui Shi. 2018. Metaheuristic
research: A comprehensive survey. Artificial Intelligence Review 1–43.
33. Janiga, Damian, Robert Czarnota, Jerzy Stopa, Paweł Wojnarowski, and Piotr Kosowski. 2017.
Performance of nature inspired optimization algorithms for polymer enhanced oil recovery
process. Journal of Petroleum Science and Engineering 154: 354–366.
34. Jiang, Jian-Hui, Ji-Hong Wang, Xia Chu, and Ru-Qin Yu. 1997. Clustering data using a modified
integer genetic algorithm (IGA). Analytica Chimica Acta 354 (1): 263–274.
35. Karaboga, Dervis, and Bahriye Basturk. 2008. On the performance of artificial bee colony
(ABC) algorithm. Applied Soft Computing 8 (1): 687–697.
36. Katrawi, Anwar H., Rosni Abdullah, Mohammed Anbar, and Ammar Kamal Abasi. 2020. Ear-
lier stage for straggler detection and handling using combined CPU test and late methodology.
International Journal of Electrical & Computer Engineering 10. ISSN: 2088-8708.
37. Khoury, Justin, Burt A. Ovrut, Nathan Seiberg, Paul J. Steinhardt, and Neil Turok. 2002. From
big crunch to big bang. Physical Review D 65 (8): 086007.
38. Kumari, Pinki, and Abhishek Vaish. 2014. Brainwave based authentication system: Research
issues and challenges. International Journal of Computer Engineering and Applications 4 (1):
2.
39. Kumari, Pinki, and Abhishek Vaish. 2015. Brainwave based user identification system: A pilot
study in robotics environment. Robotics and Autonomous Systems 65: 15–23.
40. Makhadmeh, Sharif Naser, Ahamad Tajudin Khader, Mohammed Azmi Al-Betar, and Syi-
brah Naim. 2018. Multi-objective power scheduling problem in smart homes using grey wolf
optimiser. Journal of Ambient Intelligence and Humanized Computing 1–25.
41. Makhadmeh, Sharif Naser, Ahamad Tajudin Khader, Mohammed Azmi Al-Betar, Syibrah
Naim, Ammar Kamal Abasi, and Zaid Abdi Alkareem Alyasseri. 2019. Optimization methods
for power scheduling problems in smart home: Survey. Renewable and Sustainable Energy
Reviews 115: 109362.
42. Makhadmeh, Sharif Naser, Ahamad Tajudin Khader, Mohammed Azmi Al-Betar, Syibrah
Naim, Zaid Abdi Alkareem Alyasseri, and Ammar Kamal Abasi. 2019. A min-conflict algo-
rithm for power scheduling problem in a smart home using battery. In Proceedings of the 11th
national technical seminar on unmanned system technology 2019, 489–501. Springer.
43. Makhadmeh, Sharif Naser, Ahamad Tajudin Khader, Mohammed Azmi Al-Betar, Syibrah
Naim, Zaid Abdi Alkareem Alyasseri, and Ammar Kamal Abasi. 2019. Particle swarm opti-
mization algorithm for power scheduling problem using smart battery. In 2019 IEEE jordan
international joint conference on electrical engineering and information technology (JEEIT),
672–677. IEEE.
110 Z. A. A. Alyasseri et al.

44. Mavrovouniotis, Michalis, Changhe Li, and Shengxiang Yang. 2017. A survey of swarm intelli-
gence for dynamic optimization: Algorithms and applications. Swarm and Evolutionary Com-
putation 33: 1–17.
45. Mirjalili, Seyedali, Shahrzad Saremi, Seyed Mohammad Mirjalili, and Leandro dos S. Coelho.
2016. Multi-objective grey wolf optimizer: A novel algorithm for multi-criterion optimization.
Expert Systems with Applications 47: 106–119.
46. Papa, J.P., L.P. Papa, R.R.J. Pisani, and D.R. Pereira. 2016. A hyper-heuristic approach for unsu-
pervised land-cover classification. IEEE Journal of Selected Topics in Applied Earth Observa-
tions and Remote Sensing 9: 2333–2346.
47. Patel, Monika Raghuvanshi Rahul. 2017. An improved document clustering with multiview
point similarity/dissimilarity measures. International Journal of Engineering and Computer
Science 6 (2).
48. Ramadan, Rabie A., and Athanasios V. Vasilakos. 2017. Brain computer interface: Control
signals review. Neurocomputing 223: 26–44.
49. Rao, Rajesh P.N. 2013. Brain-computer interfacing: An introduction. Cambridge University
Press.
50. Rodrigues, Douglas, Gabriel F.A. Silva, J.P. Papa, Aparecido N. Marana, and Xin-She Yang.
2016. EEG-based person identification through binary flower pollination algorithm. Expert
Systems with Applications 62: 81–90.
51. Sarier, Neyire Deniz. 2010. Improving the accuracy and storage cost in biometric remote
authentication schemes. Journal of Network and Computer Applications 33 (3): 268–274.
52. Schalk, Gerwin, Dennis J. McFarland, Thilo Hinterberger, Niels Birbaumer, and Jonathan R.
Wolpaw. 2004. BCI2000: A general-purpose brain-computer interface (BCI) system. IEEE
Transactions on Biomedical Engineering 51 (6): 1034–1043.
53. Sharma, Pinki Kumari, and Abhishek Vaish. 2016. Individual identification based on neuro-
signal using motor movement and imaginary cognitive process. Optik-International Journal
for Light and Electron Optics 127 (4): 2143–2148.
54. Xiong, Hui, Junjie Wu, and Jian Chen. 2009. K-means clustering versus validation measures:
A data-distribution perspective. IEEE Transactions on Systems, Man, and Cybernetics, Part B
(Cybernetics) 39 (2): 318–331.
55. Zaw, Moe Moe, and Ei Ei Mon. 2013. Web document clustering using cuckoo search clustering
algorithm based on Levy flight. International Journal of Innovation and Applied Studies 4 (1):
182–188.
Capacitated Vehicle Routing
Problem—A New Clustering Approach
Based on Hybridization of Adaptive
Particle Swarm Optimization and Grey
Wolf Optimization

Dao Vu Truong Son and Pham Nhat Tan

Abstract Capacitated vehicle routing problem (CVRP) is a challenging combinato-

rial optimization research and has drawn considerable interests from many scientists.
However, no standard method has been established yet to obtain optimal solutions
for all standard problems. In this research, we propose a two-phase approach: An
improved k-Means algorithm for the clustering phase and a hybrid meta-heuristic
based on Adaptive Particle Swarm—PSO and Grey Wolf Optimization—GWO
(APGWO) for the routing phase. Our approach gives results very close to the exact
solutions and better than the original k-Means algorithm. And for the routing phase,
our experimental results show highly competitive solutions compared with recent
approaches using PSO and GWO on many of the benchmark datasets.

Keywords Data clustering · Optimization · Evolutionary computation · Swarm

intelligence · Meta-heuristics · Vehicle routing problem · GWO · PSO

1 Introduction and Related Works

CVRP is a constraint satisfaction problem to find minimum traveling distances of

vehicles to serve all customers with varying demands. Previous researches proposed
a lot of exact algorithms which can find the optimal solutions. These methods are
branch and bound and cutting plane method by Lysgaard et al. [1], branch cut and
price by Fukasawa et al. [2], algorithm based on set partitioning with extra cuts
by Baldacci et al. [3] and so on. However, performance of these exact algorithms
highly depends on the problem, and the computing time for large-scale issues grows

D. V. T. Son · P. N. Tan (B)

International University, Vietnam National University - HCMC, Ho Chi Minh City, Vietnam
e-mail: [email protected]
D. V. T. Son
e-mail: [email protected]

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer 111
Nature Singapore Pte Ltd. 2021
I. Aljarah et al. (eds.), Evolutionary Data Clustering: Algorithms and Applications, Algo-
rithms for Intelligent Systems,
https://doi.org/10.1007/978-981-33-4191-3_5
112 D. V. T. Son and P. N. Tan

exponentially. The meta-heuristic approaches can obtain a “reasonably good” solu-

tion relatively quickly and can solve the large-scale problems, which make them
suitable for practical situations.
A solution for a CVRP problem is a set of customer sequences. Verdú et al.
[4] provided a comprehensive review of CVRP methods. One of the most popular
methods is a cluster-first, route-second heuristic. In these approaches, the original
problem is decomposed into smaller ones by grouping customers into clusters whose
total demand is less than or equal to the vehicle capacity. After that, the customers in
each group are routed using the traveling salesman problem (TSP). Clustering is the
operation of segmenting a heterogeneous population into a number of clusters [5].
Verdú et al. [4] summarized some important results in clustering analysis for elec-
trical customers. There are three key clustering approaches, namely, artificial neural
networks, fuzzy logic, and statistical techniques. Another comprehensive review of
clustering techniques from a data mining perspective could be found in Berkhin [6].
In statistical direction, multivariate statistics and inductive techniques are two
main approaches. In multivariate statistics, a group of variables is considered together
to classify a data set. MANOVA is a powerful tool for clustering customers by
assessing multiple metric-dependent variables. K-Means algorithm is another pop-
ular statistical method that repeats two key steps of center determination and object
allocation to clusters until total data variance to be minimized [7]. The K-Means
algorithm is easily understood and implemented. However, K-Means algorithm has
some shortcomings. Its results depend on initial value of centroids which cannot
be determined in advance. In addition, clustering data is very sensitive to outliers
that lead to unbalanced clusters. One popular inductive statistical method is deci-
sion tree. Decision tree classifies items in the form of a tree structure of attributes
and the target value with decision nodes and leaf nodes. In these nodes, a single
attribute-value is analyzed to choose which branch of the subtree will be selected to
move further. A decision node has two or more branches. The leaf node indicates
the value of the target attribute. Both categorical and numerical data can be solved
by decision trees [8].
In fuzzy classification direction, there are some interesting approaches such as
adaptive neuro-fuzzy inference system (ANFIS), fuzzy C-means (FCM), and fuzzy
subtractive clustering. ANFIS can handle fuzzy aspects of the object data by using
adaptive and learning control with fuzzy if-then rules. The FCM and related mountain
methods effectively estimate cluster centroids. However, these techniques rely on
“carefully” specified inputs such as number of clusters and initial partitions. Good
estimation of these values proves challenging. Furthermore, these techniques require
huge computational power. In this research, we choose the k-Means algorithm due
to its simplicity and ease of calculation and improved it using a meta-heuristic. By
combining the meta-heuristics in clustering, the disadvantages of K-means algorithm
are reduced. Korayem et al. [9] previously chose to improve K-means performance
for some of the CVRP dataset using GWO. Recently, swarm intelligence such as
particle swarm optimization (PSO), ant colony optimization (ACO), artificial bee
colony [10, 11] are proposed using swarm behaviors to solve complex problems.
Capacitated Vehicle Routing Problem—A New Clustering Approach … 113

In the PSO, solution particles try to move to better locations. PSO converges rela-
tively fast using two mechanisms: remembering personal best and sharing global best
experiences. Many have tried to use different versions of PSO to solve a CVRP, such
as Chen et al. [12], Kao and Chen [13], Ai and Kachitvichyanukul [14], Marinakis
et al. [15].
This study proposes a new hybrid algorithm for the CVRP, which is based on
PSO and is hybridized with GWO. We choose swarm intelligence because they are
population-based and do not rely on having a good initial solution. In this research,
our main contributions are combining PSO, adaptive PSO, and a hybridized PSO-
GWO with the traditional K-means clustering algorithm to generate the “K-PSO,
K-APSO, and K-APSOGWO” algorithms. The algorithm is tested on a number of
benchmark problems with promising results.

2 Mathematical Models of CVRP

Munari et al. [16] represented CVRP as a graph G(N , V ), where N = C ∪ {0, n + 1}

is the set of node associated to customers in C and to depot node 0 and n + 1. They
use two nodes to represent the the same single depot and imposed that all routes
start from 0 and return to n + 1. Set V contains the arcs (i, j) for each pair of node
(i, j) ∈ N . The cost associated with an arc (i, j) ∈ V is ci j . Each node has a demand
qi such that qi > 0 for each i ∈ C and q0 = qn+1 = 0. The objective of the problem
is to determine a set of minimal cost routes that satisfies all the requirements defined
above. They introduce binary decision variable xi j that equals 1 if arc (i, j) is chosen
and equals 0 otherwise; y j is a continuous decision variable denoting the cumulated
demand on the route that visits node j ∈ N up to this visit. The flow-formulation is
given as follows:


n+1 
n+1
Minimi ze ci j xi j
i=0 j=0

subject to


n+1
xi j = 1, ∀i = 1, . . . , n (1)
j=0, j=i


n 
n+1
xi h − x h j = 0, ∀h = 1, . . . , n (2)
i=0,i=h j=0, j=h
114 D. V. T. Son and P. N. Tan


n
x0 j ≤ K (3)
j=1

y j ≤ yi + q j xi j − Q(1 − xi j ), ∀i, j = 0, . . . , n + 1 (4)

di ≤ yi ≤ Q, ∀i = 0, . . . , n + 1 (5)

xi j ∈ {0, 1}, ∀i, j = 0, . . . , n + 1 (6)

Constraints (1) ensure that all customers are visited exactly once. Constraints (2)
guarantee the correct flow of vehicles through the arcs, by stating that if a vehicle
arrives to a node h ∈ N , then it must depart from this node. Constraint (3) limits
the maximum number of routes to K , the number of vehicles. Constraints (4) and
(5) ensure together that the vehicle capacity is not exceeded. The objective function
imposes that the total travel cost of the routes is minimized. Constraints (4) also
avoid subtours in the solution, i.e., cycling routes that do not pass through the depot.

3 Our Algorithm

3.1 Grey Wolf Optimization

GWO was first introduced by Mirjalili et al. [17] is an algorithm basing on the
behavior of the wolves in catching their prey. This algorithm also focuses on the
social dominant hierarchy of the wolf. The alphas are responsible for the leaders
which take a role in making decisions about hunting, sleeping place, and so on.
However, in some democratic behavior, the alpha follows the other wolves in the
pack. The alphas are the best one in managing the pack. The following rank of
wolf’s hierarchy is beta. These wolves are the subordinate, who help the alpha in
making decision. The main role of the beta is to support and give feedback to the
alpha. The lowest level is omega. They have to submit to all the other dominant
wolves. The omega also has the least right in the pack. Those wolves that do not
belong to those groups above will be called delta. These wolves have variety of roles
such as watching the boundaries of the territory and warning the pack and so on.
To develop the mathematical model, the alpha, beta, and delta are the best, second
best, and third best solutions, respectively. The next step of GWO is encircling prey,
which is calculated by proposed equations (Fig. 1).
Capacitated Vehicle Routing Problem—A New Clustering Approach … 115

Fig. 1 Grey wolf optimization

To develop the mathematically model, first, the fittest solution is considered as

alpha, the beta and delta are considered to the second and the third fittest solution,
respectively. The next step of GWO is encircling prey, which is calculated by pro-
posed equations:
→ −
− → − → → 
−
D =  C · X p (t) − X (t) (7)

−
→ −
→ −
→ − →
X (t + 1) = X p (t) − A · D (8)
−
→ −
→ −
→
where t indicates the current iteration, A and C are coefficient vectors, X p is the
−
→
position vector of the prey, and X is the position vector of a grey wolf. The coefficient
vectors are calculated by equations:
−
→
A = 2−
→ r1 − −
a ·−
→ →
a (9)

−
→
C = 2−
→
r1 (10)

where −→ r1 and −
a are linearly decreased from 2 to 0, −
→ →
r2 are random vector in [0, 1].
−→
The following step is hunting, which defines the final position of the wolf X (t + 1)
using these equations:
116 D. V. T. Son and P. N. Tan

−
→ − →
→ − → −
D α =  C 1 · Xα − X  (11)

−
→ − →
→ − → −
D β =  C 2 · Xβ − X  (12)

−
→ − →
→ − → −
D δ =  C 3 · Xδ − X  (13)

−
→ −
→ − → − →
X 1 = X α − A 1 · Dα (14)

−
→ −
→ − → − →
X 2 = X β − A 2 · Dβ (15)

−
→ −
→ − → − →
X 3 = X δ − A 3 · Dδ (16)

−
→ −
→ −
→
−
→ X1+ X2+ X3
X (t + 1) = (17)
3

3.2 Particle Swarm Optimization

PSO was a swarm intelligence algorithm, proposed by Kennedy and Eberhart [18]. In
PSO, each individual searches space with a velocity which is dynamically adjusted
according to its own and its herb’s experiences. The ith particle is represented as
X i = (xi1 , xi2 , . . . , xik ). The best previous position which giving the best fitness value
of the ith particle is recorded and represented as in pibest . The index of the best particle
k
in the population is noted as gbest . The velocity which defines the rate of the position
changing for particle ith is represented as Vi = (vi1 , vi2 , . . . , vik ). The mathematically
model of PSO is represented as the following equations:

   
vkt+1 = w ∗ vkt + c1 ∗ rand ∗ pbest tk − xkt + c2 ∗ rand ∗ g Best − xkt (18)

xkt+1 = xkt + vkt (19)

Capacitated Vehicle Routing Problem—A New Clustering Approach … 117

Fig. 2 Particle swarm

optimization

where c1 and c2 are two constants, and rand is any random number between [0, 1].
The term
 c1 ∗ rand  ∗ pbest t
k − x k is called “cognitive component” while c2 ∗
t

rand ∗ g Best − xkt is referred as “social component” (Fig. 2).

3.3 Adaptive Particle Swarm Optimization (APSO)

The value of c1 and c2 , which are usually called “acceleration coefficients”, are often
set as constants, most likely c1 = c2 = 1 or c1 = c2 = 2. These values are found by
empirical studies in order to balance the cognitive and social components, which also
balance the exploration and exploration phases. In this study, we propose a formula
to change the acceleration coefficients in each iteration. The new coefficients are
calculated as followed:
 
f xkt
c1 = 1.2 −
t
(20)
f (g Best)
 
f xkt
c2t = 0.5 + (21)
f (g Best)

where c1t and c2t stands for the coefficients at iteration t; f (xkt ) is the fitness of particle
k at iteration t, and f (g Best) is the swarm’s global best fitness. The values of 1.2
and 0.5 are also found by empirical studies. We also modify the formula for inertia
as followed:
wMax − wMin
wt = (max I ter − t) ∗ + wMin (22)
max I ter
118 D. V. T. Son and P. N. Tan

Finally, we update the velocity and position of particles by the following equations:
   
vkt+1 = w ∗ vkt + c1t ∗ rand ∗ pbest tk − xkt + c2t ∗ rand ∗ g Best − xkt (23)

xkt+1 = xkt + vkt (24)

The pseudocode of this algorithm is described as in Algorithm 1.

Algorithm 1 APSO algorithm

1: Initialize the particle population
2: Initialize parameters
3: while t < Max number o f iteration do
4: for each particle with position x p do calculate fitness value f (x p )
5: if f (x p ) is better than pbest p then
6: pbest p ← x p
7: end if
8: if f ( pbest p ) is better than gbest then
9: gbest ← pbest p
10: end if
11: end for
12: update w according to Eq. (22)
13: for each particle with position x p do
14: update c1 , c2 according to Eq. (20) and Eq. (21)
15: calculate velocity of each particle by Eq. (23)
16: update position of each particle by Eq. (24)
17: end for
18: t =t +1
19: end while
20: return gbest

3.4 Adaptive Particle Swarm Optimization–Grey Wolf

Optimization (APSOGWO)

Şenel et al. [19] provided a novel hybrid PSO-GWO by replacing a particle of the PSO
with a value being the mean of the three best wolves’ positions. In this hybrid variant,
we follow the same procedure of APSO and introduce a probability of mutation,
which will trigger a small number of iterations of GWO within the APSO main loop.
The probability of mutation is set at 0.1 in our case. The pseudocode for this is as in
Algorithm 2.
Capacitated Vehicle Routing Problem—A New Clustering Approach … 119

Algorithm 2 APSOGWO Algorithm

1: Initialize the particle population
2: Initialize parameters
3: while t < Max number o f iteration do
4: for each particle with position x p do calculate fitness value f (x p )
5: if f (x p ) is better than pbest p then
6: pbest p ← x p
7: end if
8: if f ( pbest p ) is better than gbest then
9: gbest ← pbest p
10: end if
11: end for
12: update w according to Eq. (22)
13: for each particle with position x p do
14: update c1 , c2 according to Eq. (20) and Eq. (21)
15: calculate velocity of each particle by Eq. (23)
16: update position of each particle by Eq. (24)
17: end for
18: if rand(0, 1) < pr ob then
19: run GWO
20: x p = position of the best wolf
21: end if
22: t =t +1
23: end while
24: return gbest

3.5 Clustering with Evolutionary Algorithms

This section presents the modified k-Means algorithm with evolutionary algorithms
such as PSO, GWO, APSO, and APSOGWO. In a standard K-means algorithm,
defining centroids is one important issue. Each centroid has a dimensional vector of
j j
location (cen x , cen y ). The matrix of centroids has a size of k × 5, where k is the
number of vehicle, which are [index of centroid, x coordinate, y coordinate, current
capacity of centroid C j , total distance in this cluster]. The distance matrix of each
set of centroids is number o f customer × (k + 2) matrix in which the first column
is the number of customer, the second to the (k + 1) column is di j , which is distance
from customer i to cluster j, the last column is the cluster that customer is assigned to.

3.5.1 Objective Function

Our objective is to minimize traveling distance from depot to all customers which is
based on the distance and capacity constraints of each cluster. Originally, we aim to
minimize the following function:
120 D. V. T. Son and P. N. Tan


k 
n
 
F0 = xi j − cen j 2 (25)
j=1 i=1

where k is the number of clusters, n is the number of customers, xi j is the i ( th)

customer that belongs to cluster j, cen j is the centroid of cluster j. As shown
earlier, each cluster must satisfy a capacity constraint. We introduce a penalty into
our objective function F0 as shown in Eq. (24). This function penalizes infeasible
solutions relative to the amount of constraint’s violation [11]. Accordingly, the new
fitness function can be defined as follows:
 
Δcapacit y c
F = F0 + p × (26)
Q

where Q is the capacity of each vehicle, p and c are penalty parameters, Δcapacit y
is the amount of capacity violation. In our study, these parameters are determined
experimentally.

3.5.2 k-Means with GWO

Korayem et al. [9] proposed a GWO-based clustering method to solve CVRP. In this
approach, the solutions are represented by wolves. There are k numbers of clusters.
Each wolf is a set of centroids corresponding to j cluster in k and is represented by
a k × 2 matrix. The population of wolves’ hunts for the best possible solution which
is defined as prey. The best solution is then the best position of the centroid. The
positions of all wolves is represented by (k × number o f population) × 2 matrix.
The alpha position is the matrix of centroid that has the smallest objective function
value, the beta position and delta position have the second and the third smallest
objective function values, respectively. The distance and new position of all wolves
is updated by Eqs. (11) to (17) by the parameter calculated by Eqs. (9) and (10) in
Sect. 3.1.

3.5.3 k-Means with PSO

In K-means-PSO, similar to K-means-GWO, each population is defined by set of

centroids. The best particle matrix of all centroids has the same size of the position
of all population. The global best position is the set of centroids that has the minimum
value of objective function. The velocity matrix is a (k × number o f population) ×
2 matrix. The velocity and new position of each population is calculated by Eqs. (18)
and (19) in Sect. 3.2.
Capacitated Vehicle Routing Problem—A New Clustering Approach … 121

3.5.4 k-Means with APSO and APSOGWO

In these two variants, the search procedures follow the principles of k-Means-PSO
as stated before, with the algorithms following the pseudocode in Algorithm 1 and
Algorithm 2, respectively.

3.5.5 Algorithm Approach

The clustering algorithm is developed as described below

• Initialize the position of population
• Calculate the fitness function for each population (a set of centroids)
Calculate the distance matrix from each customer to each centroids by d( j, i) =
– 
j j
(x xi − cen x )2 + (x yi − cen y )2
– Assign customer to the nearest centroid
– Calculate the current capacity and total distance of from each centroid to the
customers assign in this cluster
– Calculate the violated capacity, if violated, assign the customer to the next
nearest cluster.
– Until all nearest capacity is violate, assign the customer to the cluster that has
minimum violation
– Calculate fitness function as Eq. (24)
• Define the alpha, beta, delta wolves for GWO or swarm of particles for PSO.
• Update the new position of population
• Loop until the maximum iteration
• Calculate total traveling distance by TSP code and violation of capacity.

4 Computational Results

All of the algorithms in this paper are coded using Matlab, and all experiments were
run on a personal computer equipped with a Core i5 processor running at 2.2GHz.
Three different sets of benchmark problems were selected to evaluate the effec-
tiveness of our proposed algorithm for the CVRP. The small set consists of instances
having less than 100 customers. The medium set consists of instances having from
100 to 500 customers. The large set for those having more than 500 customers.
The benchmark set can be downloaded from the following website:
http://people.brunel.ac.uk/~mastjjb/jeb/orlib/vrpinfo.html. It has been widely
used in previous studies.
122 D. V. T. Son and P. N. Tan

Table 1 Computational results of best solutions

Instance k-PSO k-GWO k-APSO k-APSOGWO BKS
A-n60-k9 1392.7 1415.2 1439 1399 1354
B-n31-k5 687 687 687 684.7 672
B-n39-k5 573.2 567.7 573.2 564.7 549
B-n41-k6 859.3 854.3 855.2 851.4 829
B-n50-k7 767.3 765.5 767.7 767.6 741
P-n20-k2 210.5 210.5 210.5 210.5 216
P-n76-k5 663 652.4 663 648.9 627
E-n76-k8 782.4 776.1 774.9 759.4 735
E-n101-k8 885.7 871.5 894 873.1 815
E-n101-k14 1153.9 1156.5 1180.2 1154.7 1067
P-n101-k4 716.3 720.8 715.4 718.2 681
X-n101-k25 29685 29031.7 30084.9 28986.6 27591
X-n209-k16 32884.1 32619 32839.6 32597.7 30656
X-n524-k153 168891 168514.3 169890.7 167863.7 154593
X-n716-k3 50051 48123.4 52028.1 48034.4 43414
X-n801-k40 81069.4 81322 81460.1 80690 73331
X-n1001-k43 80468.7 81464.4 83819.2 81853.6 72402
Leuven2- 144107.2 158779.8 148483.7 139227 112998
n4000
Rank 2nd 3rd 4th 1st

The parameters set for different algorithms are as follows:

• GWO: 30 search agents, 500 iterations
• PSO: 30 search agents, 500 iterations, c1 = c2 = 1
• APSO: 30 search agents, 500 iterations
• APSOGWO: 30 search agents (for PSO main loop), 20 search agents (for nested
GWO loop), 500 iterations for PSO, 20 iterations for nested GWO, wMax =
0.9, wMin = 0.2.
The following tables list the computational results of our algorithm the test prob-
lems. The best solution, average solution, worst solution, and standard deviation
(Std.) computed over 5 independent runs on each problem are summarized. The best
results in each benchmark problems are in bold and underscored.
Table 1 reveals that APSOGWO is able to generate reasonable good solutions for
most of CVRPs in terms of solution quality. Twelve out of eighteen test problems
can be solved successfully by the proposed algorithm, and the result is much better
than those using GWO (three best), PSO (four best), and APSO (two best).
Table 2 shows that APSOGWO placed second after GWO for providing best worst
solutions and significantly better than PSO and APSO.
Capacitated Vehicle Routing Problem—A New Clustering Approach … 123

Table 2 Computational results of worst solutions

Instance k-PSO k-GWO k-APSO k-APSOGWO BKS
A-n60-k9 1459.3 1444.4 1466.5 1420.5 1354
B-n31-k5 696.1 687.2 694 687 672
B-n39-k5 575.8 567.9 589.8 572.5 549
B-n41-k6 933.3 912.2 925.8 864.8 829
B-n50-k7 781.8 770.8 783.2 773.2 741
P-n20-k2 210.5 210.5 210.5 210.5 216
P-n76-k5 689.9 660.6 677.3 661.9 627
E-n76-k8 804 792 826.8 794.2 735
E-n101-k8 932.1 905.8 923.5 882.1 815
E-n101-k14 1198.9 1180.1 1204.9 1182.2 1067
P-n101-k4 730.8 724.4 719.4 726.4 681
X-n101-k25 31247 30836.9 31358.5 31503 27591
X-n209-k16 33852.4 32894.9 33283.3 33398.1 30656
X-n524-k153 172386.6 173178 171572.6 169785.4 154593
X-n716-k3 51912.8 48724.5 53123.4 52107.5 43414
X-n801-k40 81296.6 82214 82613.9 84729.8 73331
X-n1001-k43 83290.3 82212.1 85794.1 84964.2 72402
Leuven2- 212129.2 235060.8 226410.7 156047.8 112998
n4000
Rank 3rd 1st 3rd 2nd

×104
3.8
PSO
3.6 GWO
Total distance - a.u.

APSO
3.4 APSOGWO

3.2

2.8
0 100 200 300 400 500
Iterations

Fig. 3 Convergence of dataset X-n101-k25

Table 3 shows that APSOGWO and GWO provide the same number of best
average results and significantly better than PSO and APSO.
Table 4 shows that APSOGWO placed second only after GWO and significantly
better than PSO and APSO.
124 D. V. T. Son and P. N. Tan

Table 3 Computational results of average solutions

Instance k-PSO k-GWO k-APSO k-APSOGWO BKS
A-n60-k9 1434.7 1434.1 1445.9 1410.3 1354
B-n31-k5 693 687.1 689.7 686.2 672
B-n39-k5 571.6 566 573.3 566 549
B-n41-k6 897.4 872.8 889.2 858.7 829
B-n50-k7 777.1 769.1 774.6 770.5 741
P-n20-k2 210.5 210.5 210.5 210.5 216
P-n76-k5 675.5 656.2 668.7 657.1 627
E-n76-k8 792.5 786.6 798.5 781.3 735
E-n101-k8 901.6 890.2 909.1 876.8 815
E-n101-k14 1178.2 1167.9 1192.9 1169.1 1067
P-n101-k4 720.4 722.6 718 721.8 681
X-n101-k25 30469.1 30069.9 30614.3 30562.3 27591
X-n209-k16 33260.2 32765.8 32980.3 32920.5 30656
X-n524-k153 169997.7 170107 170571.1 168958 154593
X-n716-k3 51061.5 48495.3 52494.1 49655.2 43414
X-n801-k40 81217.5 81813.8 82162.1 82287.7 73331
X-n1001-k43 82151.1 81960.9 84482.4 83179.4 72402
Leuven2- 144687.1 158221.9 148458.4 139724.3 112998
n4000
Rank 3rd 1st 4th 1st

×104
4.2
PSO
4 GWO
Total distance - a.u.

APSO
3.8 APSOGWO

3.6

3.4

3.2
0 100 200 300 400 500
Iterations

Fig. 4 Convergence of dataset X-n209-k16

Capacitated Vehicle Routing Problem—A New Clustering Approach … 125

Table 4 Computational results of standard deviations

Instance k-PSO k-GWO k-APSO k-APSOGWO
A-n60-k9 25.2 11.2 12.3 8.1
B-n31-k5 3.5 0.2 3.7 1
B-n39-k5 3.2 1.5 9.5 2.8
B-n41-k6 34.1 23.7 25.7 4.9
B-n50-k7 6 2.1 6.6 2.4
P-n20-k2 0 0 0 0
P-n76-k5 12.3 3.5 6.2 5.1
E-n76-k8 8.9 6.4 22.4 14.3
E-n101-k8 18.2 17.2 11 3.4
E-n101-k14 16.5 10.2 9.9 12.4
P-n101-k4 5.9 1.3 1.5 3.5
X-n101-k25 723.1 761.4 518.9 975.7
X-n209-k16 414.3 124.6 175.7 354.3
X-n524-k153 1433.2 1998.8 686.5 791.8
X-n716-k3 941 325.0 565.6 2160.2
X-n801-k40 128.3 453 616.2 2148.2
X-n1001-k43 1487.2 430 1136 1605.3
Leuven2-n4000 7116.4 6689.6 7513.5 4683
Rank 3rd 1st 4th 1st

×104
6.5
PSO
GWO
Total distance - a.u.

6
APSO
APSOGWO
5.5

4.5
0 100 200 300 400 500
Iterations

Fig. 5 Convergence of dataset X-n716-k3

126 D. V. T. Son and P. N. Tan

×104
10
PSO
Total distance - a.u.
GWO
9.5
APSO
APSOGWO
9

8.5

8
0 100 200 300 400 500
Iterations

Fig. 6 Convergence of dataset X-n801-k40

× 105
2.4
PSO
GWO
Total distance - a.u.

APSO
APSOGWO
1.9

1.3
0 100 200 300 400 500
Iterations

Fig. 7 Convergence of dataset Leuven4000

One thing worth noting is that with a very large problem size (with 4000 cus-
tomers), APSOGWO provided the best solutions for best, worst, average, and even
standard deviation results in 5 repetitions. Even though we did not compute for cases
that have more than 4,000 customers, we believe that APSOGWO will be better as
well.
We conducted a comparative study to compare our algorithm with a couple of
swarm intelligence methods available for the CVRP. The smaller the objective func-
tion value, the better the solution. Benchmark sets X-n101-k25, X-n209-k16, X-
n716-k3, X-n801-k40, and Leuven2-n4000-q150 are chosen to show the conver-
gence of different iteration-best solutions for the K-APSOGWO. The curves reveal
that K-PSO, K-APSO, and K-APSOGWO can converge faster than K-GWO. How-
ever, K-PSO and K-APSO cannot escape the local traps as well as the K-GWO. Since
K-APSOGWO inherits the trap-escaping capability of the GWO, it can avoid local
traps and provide the most accurate results (Figs. 3 and 4).
Capacitated Vehicle Routing Problem—A New Clustering Approach … 127

5 Conclusion

This chapter proposes a hybrid algorithm, APSOGWO, which takes advantage of

particle swarm optimization and grey wolf algorithm for capacitated vehicle prob-
lems. Our contributions are two-fold. First, the “acceleration coefficients”, c1 and c2
are modified in each iteration and dependent on the gBest value obtained so far. Sec-
ond, we also introduce a probability of mutation, which will trigger a small number of
iterations of GWO within the APSO main loop. The probability of mutation is set at
0.1 in our case. Computational results show that the performance of K-APSOGWO
is competitive in terms of solution quality when compared with existing GWO- and
PSO-based approaches. For future research, APSOGWO can be modified to extend
its application to vehicle routing problems with time windows or multiple depots,
among others (Figs. 5, 6, and 7).

References

1. Lysgaard, Jens, Adam Letchford, and R. Eglese. 2004. A new branch-and-cut algorithm for
the capacitated vehicle problem. Mathematical Programming.
2. Fukasawa, Ricardo, Jens Lysgaard, Marcus Poggi, Marcelo Reis, Eduardo Uchoa, and Renato
Werneck. 2004. Robust branch-and-cut-and-price for the capacitated vehicle routing problem.
In Integer programming and combinatorial optimization, 10th international IPCO conference,
New York, NY, USA, 7–11 June 2004.
3. Baldacci, R., E.A. Hadjiconstantinou, and A. Mingozzi. 2004. A new method for solving
capacitated location problems based on a set partitioning approach. Operation Research 52:
723–738.
4. Verdú, S.V., M.O. Garcia, C. Senabre, A.G. Marin, and F.J.G. Franco. 2006. Classification, fil-
tering, and identification of electrical customer load patterns through the use of self-organizing
maps. IEEE Transactions on Power Systems 21: 1672–1682.
5. Linoff, G.S., and M.J. Berry. 2011. Data mining techniques: For marketing, sales, and customer
relationship management. Wiley.
6. Berkhin, P. 2006. A survey of clustering data mining techniques. In Grouping multidimensional
data: Recent advances in clustering, ed. J. Kogan, C. Nicholas, and M. Teboulle, 25–71.
Springer.
7. Hammouda, K., and F. Karray. 2000. A comparative study of data clustering techniques. Uni-
versity of Waterloo, ON, Canada.
8. Amatriain, X., and J.M. Pujol. 2015. Data mining methods for recommender systems. In Rec-
ommender systems handbook, ed. F. Ricci, L. Rokach, and B. Shapira, 227–262. Springer
US.
9. Korayem, Lamiaa, M. Khorsid, and Sally Kassem. 2015. Using grey wolf algorithm to solve
the capacitated vehicle routing problem. IOP Conference Series: Materials Science and Engi-
neering 83.
10. Lin, S.W., Z.J. Lee, K.C. Ying, and C.Y. Lee. 2009. Applying hybrid meta-heuristics for
capacitated vehicle routing problem. Expert Systems with Applications 36 (2): 1505–1512.
11. Szeto, W.Y., Y. Wu, and S.C. Ho. 2011. An artificial bee colony algorithm for the capacitated
vehicle routing problem. European Journal of Operational Research 215 (1): 126–135.
12. Chen, A.L., G.K. Yang, and Z.M. Wu. 2006. Hybrid discrete particle swarm optimization
algorithm for capacitated vehicle routing problem. Journal of Zhejiang University: Science 7
(4): 607–614.
128 D. V. T. Son and P. N. Tan

13. Kao, Y., and M. Chen. 2011. A hybrid PSO algorithm for the CVRP problem. In Proceedings
of the international conference on evolutionary computation theory and applications (ECTA
’11), Paris, France, 539–543, Oct 2011.
14. Ai, T.J., and V. Kachitvichyanukul. 2009. Particle swarm optimization and two solution rep-
resentations for solving the capacitated vehicle routing problem. Computers and Industrial
Engineering 56 (1): 380–387.
15. Marinakis, Y., M. Marinaki, and G. Dounias. 2010. A hybrid particle swarm optimization
algorithm for the vehicle routing problem. Engineering Applications of Artificial Intelligence
23 (4): 463–472.
16. Munari, Pedro, Twan Dollevoet, and Remy Spliet. 2017. A generalized formulation for vehicle
routing problems.
17. Mirjalili, S., S.M. Mirjalili, and A. Lewis. 2014. Grey wolf optimizer. Journal of Advances in
Engineering Software 69: 46–61.
18. Kennedy, J., and R. Eberhart. 1995. Particle swarm optimization. In Proceedings IEEE inter-
national conference on neural networks (ICNN’95), Perth, WA, Australia, vol. 4, 1942–1948,
Nov–Dec 1995.
19. Şenel, F.A., F. Gökçe, A.S. Yüksel, et al. 2019. A novel hybrid PSO–GWO algorithm for
optimization problems. Engineering with Computers 35: 1359.
A Hybrid Salp Swarm Algorithm with
β-Hill Climbing Algorithm for Text
Documents Clustering

Ammar Kamal Abasi, Ahamad Tajudin Khader, Mohammed Azmi Al-Betar,

Zaid Abdi Alkareem Alyasseri, Sharif Naser Makhadmeh,
Mohamad Al-laham, and Syibrah Naim

Abstract Recently, the researchers’ attention became more interest in partitioning

particular sets of documents into various subsets, due to the massive number of doc-
uments that make pattern recognition, information retrieval, and text mining more
complicated. This problem is known as a text clustering problem (TCD). Several
metaheuristic optimization algorithms have been adapted to address TDC optimally.
A new efficient metaheuristic optimization algorithm mimics the behavior of salps

A. K. Abasi (B) · A. T. Khader · S. N. Makhadmeh

School of Computer Sciences, Universiti Sains Malaysia, George Town, Pulau Pinang, Malaysia
e-mail: [email protected]
A. T. Khader
e-mail: [email protected]
S. N. Makhadmeh
e-mail: [email protected]
M. A. Al-Betar
Department of Information Technology - MSAI, College of Engineering and Information
Technology, Ajman University, Ajman, United Arab Emirates
e-mail: [email protected]
IT Department, Al-Huson University College, Al-Balqa Applied University, Irbid, Jordan
Z. A. A. Alyasseri
Faculty of Information Science and Technology, Center for Artificial Intelligence, Universiti
Kebangsaan Malaysia, 43600 Bangi, Selangor, Malaysia
e-mail: [email protected]
ECE Department, Faculty of Engineering, University of Kufa, Najaf, Iraq
M. Al-laham
Department of Management Information Systems, Amman University College, Al-Balqa Applied
University (BAU), Amman, Jordan
e-mail: [email protected]
S. Naim
Technology Department, Endicott College of International Studies (ECIS), Woosong University,
Daejeon, South Korea
e-mail: [email protected]

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer 129
Nature Singapore Pte Ltd. 2021
I. Aljarah et al. (eds.), Evolutionary Data Clustering: Algorithms and Applications, Algo-
rithms for Intelligent Systems,
https://doi.org/10.1007/978-981-33-4191-3_6
130 A. K. Abasi et al.

in oceans known as the salp swarm algorithm (SSA) has been proposed and adapted
to address different optimization problems. However, hybridizing optimization algo-
rithms with another algorithm is becoming the focus of scholars to obtain a superior
solution for the optimization problems. In this paper, a new hybrid optimization
method of SSA algorithm and a well-known metaheuristic optimization algorithm
called β-hill climbing algorithm (BHC), namely H-SSA, is proposed. The main
aims of the proposed method to improve the quality of initial candidate solutions and
enhance the SSA in terms of local search ability and convergence speed in attempting
for optimal partitioning of the cluster. The proposed H-SSA performance is tested in
the data cluster field using five standard datasets. In addition, the proposed method is
tested using two scientific articles’ datasets, and six standard text datasets in the text
document clustering domain. The experiment results show that the proposed method
boosted the solutions in terms of convergence rate, recall, precision, F-measure, accu-
racy, entropy, and purity criteria. For comparative evaluation, the proposed H-SSA
compared with the pure SSA algorithm and well-known clustering techniques like
DBSCAN, agglomerative, spectral, k-means++ k-means clustering techniques and
the optimization algorithms like KHA, PSO, GA, HS, CMAES, COA, and MVO. The
comparative results prove the efficiency of the proposed method, where it exhibited
and yielded better performance than the compared algorithms and techniques.

Keywords Data clustering · Document clustering · Beta-hill climbing ·

Optimization · Evolutionary computation · Hill climbing · Swarm intelligence ·
Meta-heuristics

1 Introduction

Recently, one of the research topics became more interest due to its contributing
grouping a particular set of text documents into subsets of clusters [34]. This topic is
known as text document clustering (TDC). Several techniques have been structured
to efficiently clusterize documents that have high intra-similarity by allocating the
related documents at the same cluster [43]. This situation can be done according to
several attributes that characterize the data [46].
TDC is considered the most prominent unsupervised technique due to its undis-
covered area that increases the burden of finding the best cluster for documents [25].
The main aim of TDC algorithms is to cluster documents into related groups under
the same topics. However, these topics are not titled in advanced. In this study, the
procedures of clustering and partitioning the documents into specific subsets is the
primary objective. These documents will be clustered based on an objective function.
Several algorithms have been built for TDC, where the K-means algorithm is the
most prominent algorithm, due to its ability to deal with a massive number of data [4,
5, 34, 57]. However, the K-mean algorithm like other algorithms has disadvantages,
where it could easy to fall in local optimal and not obtain the optimal cluster in some
cases [62].
A Hybrid Salp Swarm Algorithm with β-Hill Climbing Algorithm … 131

The metaheuristic optimization algorithms have been adapted instead of TDC

algorithms, such as k-means, to efficiently address TDC, where this kind of opti-
mization algorithms showed its robust performance in different fields [19]. The
most prominent metaheuristic optimization algorithms are ray optimization algo-
rithm (ROA) [47], genetic algorithm (GA) [44], grey wolf optimizer (GWO) [49–
51], harmony search (HS) [13, 36], cuckoo search (CS) [68], binary artificial bee
colony (ABC) [58], multi-verse optimizer (MVO) [4], fruit fly optimization algo-
rithm (FFOA) [59], krill herd algorithm (KHA) [37], flower pollination algorithm
[23], ant colony optimization (ACO) [66], particle swarm optimization (PSO) [28,
52], ant lion optimizer (ALO) [53], and dragonfly algorithm (DA) [54].
Salp swarm algorithm (SSA) is one of the recent and famous metaheuristic opti-
mization algorithms due to its capability, flexibility, simplicity, and easy to be under-
stood and utilized. In addition, SSA has a good ability to balancing exploration and
exploitation through the searching mechanism [33]. SSA is one of the population-
based metaheuristic algorithms that mimic the behavior of salps in nature that pro-
posed by Mirjalili et al. [55]. In contrast, the SSA like other metaheuristic algorithms
has some disadvantages that reduce its searching performance, such as stuck in local
optima and slow convergence speed [38]. Therefore, an improvement in the SSA
mechanism to speed up the convergence rate toward optimal solution without any
stagnation in local optimal should be well prepared. This improvement can be done
by enhancing the best solution at each iteration to find a better solution. Hybridizing a
pure metaheuristic algorithm with another optimization algorithm is the most promi-
nent technique to address algorithms drawbacks and improve their performance and
results [19].
Normally, hybridizing a local search as a new operator in the SSA can improve
the exploitation capability. The recent local search algorithm called β-hill climbing
is proposed by [8]. It is utilized a stochastic operator with the neighborhood opera-
tor to enable the algorithm to escape the local optima. β-hill climbing optimizer is
simple, adaptable,easy-to-use, and efficient. It has been adapted for a wide range of
optimization problems such as ECG and EEG signal denoising [18, 20, 22], Gen-
erating Substitution-Boxes [24], gene selection [15], economic load dispatch prob-
lem [10, 12], mathematical optimization functions [6] multiple-reservoir scheduling
[16], numerical optimization, [9], Sudoku game [11], classification problems [17],
and cancer classification [14].
In this paper, a hybrid version of SSA and a local search optimization algorithm
known as β-hill climbing (H-SSA) is proposed. The proposed method is prepared to
enhance the SSA the best solution at each iteration and initial solutions. The proposed
H-SSA performance is tested in the data cluster field using five standard datasets. In
addition, the proposed method is tested using two scientific articles’ datasets, and six
standard text datasets in the text document clustering domain. The obtained results by
the proposed method are compared several comparative techniques and algorithms,
including optimization algorithms (e.g., PSO, GA, H-PSO, H-GA, MVO, KHA,
HS) and clustering techniques (e.g., Spectral, Agglomerative, K-mean++, K-mean,
DBSCAN).
132 A. K. Abasi et al.

The rest of the paper is divided into eight sections. The second section reviews pre-
viously conducted studies and related works. The third section presents and discusses
the TDC problem in detail. The fourth section introduces the salp swarm algorithm’s
basic principles. The fifth section discusses the β-hill climbing principle, whereas the
sixth section describes in detail the proposed method H-SSA. The seventh section
provides the data set used, the experimental findings, and the proposed method’s
significance. The conclusions, as well as recommendations for further studies, are
provided in the eighth section of the current paper.

2 Related Works

In the field of TDC, the researches use insistent and advanced techniques to address
TDC. TDC can be addressed by partitioning documents into distinct clusters accord-
ing to the similarity of the documents’ content [65]. Several clustering methods
were proposed to address the TDC problem, where the k-means method is one of
the most popular methods. k-means method is only searching for local solutions in
addressing the TDC problem. Therefore, it was hybridized with other algorithms
to improve the poor local searching [45]. Also, β-hill climbing is one of the most
prominent algorithms that only searching for local solutions. β-hill climbing is a
recent single-based metaheuristic algorithm tailored for various optimisation prob-
lems [1, 8]. β-hill climbing has been hybridized with other algorithms to enhance
the exploitation side of the searching process of particular problem. Using this algo-
rithm, the results demonstrate its feasibility, significant improvement in performance
while exploiting the search space [2, 15, 22].
As mentioned previously, several metaheuristic optimization algorithms have
been adapted to address TDC problem. SSA is one of the recent swarm-based meta-
heuristic optimization algorithms due to its capability, flexibility, simplicity, and easy
to be understood and utilized. SSA was adapted to address the choosing task for the
best conductor in a radial distribution system [41]. The authors of [31] adapt SSA to
tune a stabilizer in a power system. In the experiment results, SSA presented a high
performance compared with other algorithms.
Lately, a few numbers of research have been hybridized the SSA with other
algorithms to enhance the SSA searching mechanism and improve the solutions.
The authors of [32] hybridize the SSA with the differential evolution algorithm to
enhance the feature exploitation ability of the SSA due to the differential evolution
algorithm performance in searching locally. SSA was hybridized with PSO in [40]
to address the feature selection problem. The main purpose behind the hybridization
is to improve the exploration and exploitation processes of SSA. The results prove
the robust performance of the proposed hybrid version in addressing the problem.
A Hybrid Salp Swarm Algorithm with β-Hill Climbing Algorithm … 133

3 Text Document Clustering Problem

TDC can be characterized as the problem of the NP-complete, which involves finding
clusters in documents that are heterogeneous through the minimization of objective
in the present paper(min f (x) is equal to minimize the Euclidean distance). This
section describes the formulation of the TDC problem, the Preprocessing of TDC,
and the TDC algorithm, in addition to the similarity measures, as well as the objective
function.

3.1 Problem Formulation

The TDC problem can be characterized by considering a specific set (Docs) of the t
documents at a high level. It aims to partition the documents into a specific number
of predetermined clusters (k) subsets, whereby (Docs) represents a documents’ vec-
tor (Docs = (Doc1 , Doc2 , . . . , Doci , . . . , Docd )). Doci denotes the number of the
document i, and Docd stands for the entire documents’ number in Docs. Each clus-
ter holds a specific cluster’s centroid (K cent ), which represents a specific vector of
terms’ weights of length f (kcnt = (kcnt1 , kcnt2 , . . . , kcnt j , . . . , kcnt f )), whereby kcnt
signifies a centroid of kth cluster, kcnt1 stands for a position value 1 in the centroid
of the cluster k, and kcnt f stands for the number of the entire special features of the
centroid (terms) [39, 60].
To determine a partition kcnt = (kcnt1 , kcnt2 , . . . , kcnt j , . . . , kcnt f that satisfies cer-
tain conditions:

= ∅
• kcnt 
• kcnt kcnt  = ∅i f K = K 
K
• kcnt = 0
k=1
• Objects that are in the same cluster are very similar to one another. However,
objects that are in diverse clusters are different from each other.

3.2 Preprocessing of Text Document

Before the application of the clustering algorithm, standard preprocessing steps are
implemented to preprocess text documents, including tokenization and stop word
removal, as well as stemming, in addition to steps of term weighting. Text documents
can be converted to a format that is numerical or matrix via the preprocessing steps.
In the present work, the term frequency-inverse document frequency (TFIDF) can
be generally used as a scheme of weighting.
134 A. K. Abasi et al.

Fig. 1 Solution representation

3.3 Clustering Algorithm and Similarity Measures

The clustering algorithm implementation for documents, as well as the features of

the document to produce clusters based on the similarity measures, establishes a
substantial phase in the clustering of documents. The TDC algorithm represents
a method of unsupervised learning. This method aims to find the best solution to
partition a set of documents [36]. The TDC algorithm functions in accordance with
certain evaluation criteria, including objective function, in addition to fitness [27].
Often, similarity measures can be represented in relation to distance or dissimilarity
measures. The Euclidean distance measure [61] is a standard measure, and it is used as
an objective function for various algorithms of text clustering so that the resemblance
between the documents, as well as the cluster’s centroid, can be computed, whereby
the aim of the cost function is to minimize distance between the documents within
each cluster.

3.3.1 Solution Representation

Every solution within the population suggests one candidate solution to the problem;
each solution can be indicated as a specific vector x = (x1 , x2 , . . . , xd ), whereby d
signifies the number of documents, and each variable’s value xi can take the value
of k decision k ∈ {1, . . . , K }, whereby k signifies a random cluster’s number, and K
indicates the clusters’ number as shown in Fig. 1. Each dimension can be treated as
one document. As shown in Fig. 1, the solution X possesses dim (here dim = 20),
twenty documents can be distributed between five clusters, e.g., document 5 (namely,
x5 ) can be in the cluster one; cluster five possesses four documents {1, 2, 3, 19, 20}.

3.3.2 Objective Function

The measure of the Euclidean distance has been used in this work. It is a specific
standard measure and it has been widely used for similar purposes [27, 36]. This
measure has been widely applied with the aim of computing resemblance between
two documents (namely, the document, as well as the cluster’s centroid) [61], e.g.,
when considering two documents d1 = (t11 , t12 , . . . , t1n ) and d2 = (t21 , t22 , . . . , t2n ),
the Euclidean distance can be calculated according to:
A Hybrid Salp Swarm Algorithm with β-Hill Climbing Algorithm … 135
 n 1/2

D(d1 , d2 ) = |t1,i − t2,i |2 , (1)
i=1

where t1,i indicates the term’s weight i in document one, and t2,i indicates the term’s
weight i in document two (namely, the feature’s weight i in the cluster’s centroid).
The distance values range between (0, 1), whereby the 0 value denotes the most
appropriate value; the 1 value denotes the worst value. Euclidean distance signifies
a distinct case of the Minkowski distances [26] that can be expressed in Eq. (2),
except for the p value that equals 2 in the Euclidean distance, which represents the
key difference.
 n 1/ p

D(d1 , d2 ) = |t1,i − t2,i | p
(2)
i=1

For the text document’s solution x to be assessed, the resemblance between each
cluster of the documents, as well as the cluster’s centroid can be computed. In this
work, the measure of distance between documents in a similar cluster can be utilized
by means of computing the resemblance between each document, as well as each
cluster’s centroid for kcnt , where kcnt = (kcnt1 , kcnt2 , . . . , kcnt j , . . . , kcnt f ) is a specific
cluster centroid vector of the length f .
To distribute each of the documents in its appropriate cluster’s centroid, each
document cluster’s centroid is recomputed using Eq. (3). That is, a documents’ group
can be placed in a similar cluster with a closer clustering centroid.
d
i=1 (aki )di j
kcnt j =  d
, (3)
i=1 aki

whereby kcnt j represents the cluster’s centroid j, d signifies the overall number of
documents, whereas di j denotes to jth feature weight for document i. aki signifies a
binary matrix size d × k based on:

1 document i assigned to the cluster j,
aki = (4)
0 otherwise

Finally, each solution’s objective function x of the introduced algorithm should

be formulated. Thus, the distance of an average document to a cluster’s centroid
(ADDC) is used based on Eq. 5.
k 1 di
i=1 ( di j=1 D(kcnti , d j ))
min f (x) = , (5)
k

where f (x) signifies an objective function (namely, minimizing the distance), di

indicates the number of documents in cluster i, k signifies the number of clusters,
136 A. K. Abasi et al.

in addition to D(kcnti , d j ), which indicates distance between cluster’s centroid j, as

well as document i.

4 Salp Swarm Algorithm (SSA)

SSA is a swarm-intelligence algorithm which introduced by Mirjalili in [55]. It is

inspired from swarming behavior of salps when navigating and foraging in oceans.
The main procedure of the standard algorithm of SSA can be summarized by the
following phases:
Phase 1: Parameters Initialization. Both parameters of the SSA and the text docu-
ment problem should be initialized within possible parameters range value
s. In this case, the general formulation of the SSA initialization can be
introduced as follows:

max or min{ f (s) | s ∈ S},

where f (s) is the objective function; s = {si | i = 1, . . . , N } is the set of

decision variables. s = {si | i = 1, . . . , d} is the possible value range for
each decision variable, where si ∈ [L Bi , U Bi ], where L Bi and U Bi are
the lower and upper bounds for the decision variable si , respectively, and
N is the number of decision variables.
Also, for the rest SSA parameters must be initialized in this phase too, these
parameters can summarize as follows:
• Salps i ze: is a SSA population size.
∗
• Fbest : refers to the best current solution (best agent).
The next phases will provide a full explanation of these parameters.
Phase 2: Initialize SSA population memory. The SSA can be represented as a
two-dimensional matrix with size Salps i ze × N which have sets of slap
location vectors as many as Salps i ze (see Eq. (6)). Where these slaps
j
are randomly generated as follows: si = L Bi + (U Bi − L Bi ) × U (0, 1),
∀i = 1, 2, . . . , z and ∀ j = 1, 2, . . . , Salps i ze, and U (0, 1) generates a uni-
form random number between 0 and 1. The generated solutions are stored
in the SSAM in ascending order according to their objective function values
where f (s1 ) ≤ f (s2 ) ≤ . . . ≤ f (s SS As ).

⎡ ⎤
s11 s21 · · · s N1
⎢ s12 s22 · · · s N2 ⎥
⎢ ⎥
SSAM = ⎢ .. .. .. ⎥ . (6)
⎣ . . ··· . ⎦
s1SS As SS As
s2 · · · s NSS As
A Hybrid Salp Swarm Algorithm with β-Hill Climbing Algorithm … 137

Also in this step, the global best slap location Sbest is memorized where
Sbest = s 1 .
Phase 3: Intensification of the current slap population. To model the SSA chains,
the population divided into two sets which are leader and followers. The
leader is the slap at the front of the chain, whereas the rest of the slaps
are considered as slaps. Similarly to other swarm-based techniques, the
position of slaps is defined in an n-dimensional search space where n is
the number of variables of a given problem. Therefore, the position of all
slaps are stored in a two-dimensional matrix called s. It is also assumed that
there is a food source called F in the search space as the swarm’s target. To
update the position of the leader the following equation is proposed:
⎧
⎨ F( j) + c1 × ((ub j − lb j ) × c2 + lb j )Bigr  c3 ≥ 0,
si = 
j
(7)
⎩ F( j) − c1 × ((ub j − lb j ) × c2 + lb j )Bigr  c3 < 0

j
where si shows the position of the first slap (leader) in the jth dimension,
F j is the position of the food source in the jth dimension, ub j indicates
the upper bound of jth dimension, lb j indicates the lower bound of jth
dimension, c1 , c2 , and c3 are random numbers.
Phase 4: Updating the position of the leading slap (best solution ( Sbest )). During
for each iteration in SSA procedure, the global best slap location Fbest will
be updated if f (s j ) < f (Sbest ).
Phase 5: Stop condition. SSA repeats phase 3 and phase 4 till the stop criterion is
met. The stop criterion is normally met depends on two criteria, such as the
quality of the final outcomes or the number of iterations.

5 β-Hill Climbing Algorithm

The idea of β-hill climbing algorithm extended from a simple trajectory-based algo-
rithm called hill climbing algorithm, where the idea of the hill climbing algorithm
for finding the optimal solution for any problem. Firstly it starts to generate a random
solution then step by step try to improve this solution by finding the better parameters
from the search space. This trajectory approach will be repeated until reaches the
optimal solution. The main problem with a trajectory-based algorithm such as hill
climbing it is always looking for uphill movements to finding the optimal solutions.
For that, it leads to getting easily stuck in the local optima [8]. For solving the stuck
in the local optima problem and accept the downhill movements Al-Betar in 2017
introduced a new extension of the hill climbing algorithm which called β-hill climb-
ing algorithm [8]. The main idea behind the β-hill climbing algorithm is using a
simple operator to make a balance between both exploration and exploitation during
the search.
138 A. K. Abasi et al.

As aforementioned, the β-hill climbing algorithm is a trajectory search algorithm

which starts with single random solution, S = (s1 , s2 , . . . , s N ). After number of iter-
ation in the searching space, a new solution, s = (s1 , s2 , . . . , s N ), will be generated
by improving the current solution s. The improvement will happen based on two
operators, namely, N -operator and β-operator, where these operators represent the
sources for exploitation and exploration, respectively. Specifically, the N -operator
works as neighborhood search, while β-operator works as similar to mutation oper-
ator. At each iteration, the new solution can be improved by N -operator stage or
β-operator stage until the optimal solution is reached.
The algorithm begins to generate the solution randomly, then the solution is eval-
uated using the objective function f (s). The solution is then modified using N -
operator, which employs the impr ove(N (s)) function within a random range of its
neighbors. The solution s is as follows:

si = si ± U (0, 1) × bw ∃i ∈ [1, N ]

where i is randomly selected from the space range, i ∈ [1, 2, . . . , N ]. The parameter
bw representees the bandwidth between the current value and the new value.
In β-operator, within the β range where β ∈ [0, 1], variables of new solution will
be assigned based on selected randomly from available range or from the existing
values of the current solution as follows:

sr r nd ≤ βhc
si ←
si other wise,

where r nd generates a uniform random number between 0 and 1 and sr ∈ Si is the

possible range for the decision variable si .
Finally, the β-hill climbing has successfully achieved optimal results in many
global problems such as EEG and ECG signal denoising, EEG channel selection,
feature selection, and sudoku problem [7, 11, 18, 20–22].

6 The Proposed Hybrid Salp Swarm Algorithm for TDC

This section provides the introduced method (H-SSA) in detail. The introduced SSA
algorithm’s hybrid strategy consists of two main stages. They include improving the
quality of the initial candidate solutions in stage one and improving the best solution
in stage two that SSA provides at each iteration. Figure 2 displays the steps of the
introduced hybrid strategy and every single phase is described as follows.
A Hybrid Salp Swarm Algorithm with β-Hill Climbing Algorithm … 139

6.1 Improving the Initial Candidate Solutions Quality Using

β-Hill Climbing

The SSA algorithm is characterized by being too sensitive towards early condition
solutions. Therefore, the starting point of the algorithm may have a major impact on
the efficiency of the ultimate optimum solution’s [64].
Two or even more algorithms have been hybridized in recently conducted studies
to obtain the optimal solutions for tackling the problems of optimization [42, 63]. The
main hybridization drivers avoid local optimization and enhance initial solutions.
This results in enhancing global (diversification) searchability and it can be more
essential in problems of multi-dimension. In accordance with this perspective, the
introduced hybrid algorithm can combine the local search algorithm’s main features
to improve the initial solution. Thus, the result is directed as an SSA initial solution.
Following the step of preprocessing, the H-SSA can start initializing a few param-
eters; it can randomly generate solutions. Every single solution can be associated with
one salp. In other words, the number of solutions, as well as the number of salps is
equal. It is important to mention that the original SSA’s initialization phase (Sect. 4
can be adopted to cluster with specific modifications, which are associated with the
problem’s variables nature, considering that the problem of clustering is discrete;
the SSA algorithm was initially utilized to tackle continuous problems of structural
optimization [56]. The SSA has been introduced with the aim of dealing with the
decision variables’ discrete values of every single TDC solution via utilizing a round-
ing function so that continuous values are converted to discrete values. These changes
can be applied as follows:

1. Generation of initial solution function Eq. (6) is modified to:


j
xi = rand()%(ub j − lb j ))Bigr  ∀i ∈ (1, 2, . . . , n) ∧ ∀ j ∈ (1, 2, . . . , d),
(8)
whereby n signifies the candidate number of solutions and d signifies the number
of documents, lb j = 1, ub j signifies the number of clusters, and rand() signifies
a specific function, through which an arbitrary distribution number is generated
within the range [0.0, 1.0]. Bigr is a function that is approximately the value of
j j
xi to the closest integer; it is less than or it is equal to xi ,
2. The updates’ equations of the leader’s position and the follower’s position is
changed in Eq. 9 to
⎧
⎨ F( j) + c1 × ((ub j − lb j ) × c2 + lb j )Bigr  c3 < 0.5,
xi = 
j
(9)
⎩ F( j) − c1 × ((ub j − lb j ) × c2 + lb j )Bigr  c3 ≥ 0.5,

j
whereby xi demonstrates the first salp position (leader) in jth dimension and F j
demonstrates the food source position in jth dimension, as well as ub j , which
demonstrates the jth dimension’s upper bound, whereas lb j demonstrates the jth
140 A. K. Abasi et al.

dimension’s lower bound, in addition to c1 , c2 , and with c3 , which represent

arbitrary numbers.
1
j j j−1
xi = (xi + xi ), (10)
2
j
whereby i ≥ 2 and xi demonstrates the ith follower salp position in the dimension
of jth.
A solutions’ set can be randomly produced in the phase of initialization utilizing
Eq. 8. In this phase, the BHC algorithm can consider solutions in the form of an
input to enhance each solution and recalculate the solution’s quality using Eq. 5.
In case the generated quality of the solution can outperform the quality of the old
solution, it can memorize the solution that is generated; otherwise, it will continue
with an old solution. This process can be repeated considering the entire solutions
in a population (see, Algorithm 1).

Algorithm 1 Enhancing the quality of initial candidate solutions algorithm

Input: Population matrix U of size n solutions
Output: Improving Population matrix U of size n solutions by β-hill climbing algorithm
for Each solution xi in Population matrix U do
Step1: calculate the solution xi quality using Eq. 5
Step2: make the solution xi instead of the first step in β-hill climbing algorithm.
Step3: run β-hill climbing algorithm
Step4: calculate the solution xi quality using Eq. 5

if The solution xi improved then

Step5a: replace the solution xi with the old solution.
end
else
Step5b: the solution xi remains unchanged.
end
end

6.2 Improving the Best Solution of Salp Swarm Algorithm

with β-Hill Climbing

Similar to the remaining metaheuristic algorithms, SSA can be designed to accom-

plish two phases, including exploration and exploitation phases. In these phases, the
algorithm must be equipped with specific mechanisms so that an extensive search can
be carried out for a search space. The promising regions of the search space can be
identified in the exploration phase. On the other hand, the local search is highlighted
in the exploitation phase, as well as convergence to reach in an exploration phase
promising areas. The key adequate solution in both phases is the leading solution.
In the exploration phase, it is highly possible that the solutions learn from the lead-
A Hybrid Salp Swarm Algorithm with β-Hill Climbing Algorithm … 141

Fig. 2 The proposed methodology

ing solution. Furthermore, In the exploitation phase, it focuses only on the leading
solution. A liberal bias for improving the leading solution may share good attributes
along the searching path for the global optima, as well as lead the search so that better
solutions are found. At every single iteration, the leading solution, which SSA pro-
duces, is the β-hill climbing initial state. The stage intensifies the probability of the
existence of useful attributes in the optimal solution. Therefore, the optimal solution
can send these to the existing solutions and assist them in improving the fitness value.
The enhancement process of the optimal solution is provided in detail in Algorithm
2. The optimization framework of the proposed method is demonstrated in Fig. 2, as
well as pseudocode in Algorithm 3.

Algorithm 2 Enhance the best solution of SSA

Input: The best solution best S I
Output: Improving solution best S I by β-hill climbing algorithm
Step1: make the solution best S I instead of the first step in β-hill climbing algorithm.
Step2: run β-hill climbing algorithm
Step3: calculate the solution best S I quality using Eq. 5
if The solution best S I improved then
Step4a: replace the solution best S I with the old solution.
end
else
Step4b: the solution best S I remains unchanged.
end
142 A. K. Abasi et al.

Algorithm 3 Pseudocodes and general steps of SSA algorithm

1: Create SSA population considering ub and lb.
2: run Algorithm 1
3: while the end criterion is not satisfied do
4: Evaluate the objective function for all solutions.
5: Update the Best solution vector.
6: run Algorithm 2
7: Update c1
8: for each solutions indexed by i do
9: if i == 1 then
10: Update the position of the leading salpby Eq. 9
11: else
12: Update the position of the leading salpby Eq. 10
13: end if
14: end for
15: Amend the solutions based on the upper and lower bounds of variables.
16: end while
17: Produce the best Solution.

7 Results and Discussion

This section describes the experimental architecture, experimental datasets, state-of-

the-art algorithms, evaluation measures of the system, and parameter descriptions.
The convergence rate is also given in this study, and the findings are addressed and
analyzed. The datasets are described in the next section.

7.1 The Experimental Design

This section discusses the proposed method and involves a comparison of the intro-
duced H-SSA method with the remaining techniques of the state of the art. To provide
better insights into this work, the performance of the introduced algorithm is eval-
uated utilizing three artificial datasets types, involving scientific articles datasets
(SAD), data clustering datasets (DCD), in addition to text clustering datasets (TCD).
To achieve the results’ uniformity, the introduced algorithm has been conducted
30 times along with all the datasets, utilizing similar initial solutions for the meta-
heuristic algorithms. This specific number has been selected based on the literature.
Therefore, the introduced method is appropriately validated. Also, an almost reason-
able comparison that involved the entire competing algorithms has been conducted.
The algorithms that are local-based for the clustering technique can run 100 iterations
each run time, and 100 iterations are experimentally appropriate for convergence of
the intensification search algorithm. Also, 1000 iterations can be appropriate for con-
vergence of the diversification search algorithm of the population-based algorithms.
Seven measures of an external evaluation have been implemented as has always
been conventionally done, including accuracy and recall, as well as precision and
F-measure, in addition to purity and entropy criteria, along with the intra-cluster
distances’ sum, which is a measure of internal quality. For conducting a comparative
A Hybrid Salp Swarm Algorithm with β-Hill Climbing Algorithm … 143

assessment, the obtained findings by utilizing these evaluation measures were com-
pared as opposed to the obtained findings by utilizing ten state-of-the-art algorithms,
which include the original SSA K-mean and K-mean++, as well as DBSCAN and
Agglomerative, in addition to Spectral, KHA, HS, PSO, GA, CMAES, COA, along
with MVO by using a similar objective function.

7.2 Experimental Benchmark Datasets and Parameter

Setting

To validate the efficacy of the introduced H-SSA algorithm, three artificial datasets
types have been selected to test the clustering problem. Five DCD and two SAD, as
well as six TCD based on the literature, have been chosen as testbeds for the conducted
comparison. The datasets are benchmark datasets in the field of clustering. These
were commonly applied to examine the recently developed algorithms’ performance
including: CMC and Iris, as well as Seeds and Glass, in addition to Wine. TCD
is obtainable at the computational intelligence laboratory (LABIC) as a numerical
form after the term extraction. The characteristics of the TCD are described as in the
following:

CSTR. The Centre for Speech Technology Research (CSTR) is a research center
that is interdisciplinary. In CSTR, informatics and linguistics as well as the English
language are linked. In 1984, this center was established to foster research in
various areas, including information access, as well as speech recognition. The
selected dataset comprises 299 documents. These documents are included in four
classes, including theory and artificial intelligence, as well as robotics, in addition
to systems.
20Newsgroups. It comprises 19,997 articles. They are included in 20 classes and
collected from various Usenet newsgroups. Thus, the initial 100 documents for
the experiment have been chosen from the superior three classes within the dataset
as follows: comp_windows_x, talk_politics_misc, and rec_autos classes.
Tr12 , Tr41, and Wap (Datasets from Karypis Lab). These have been selected
from different sources to ensure diversity. Tr12 and Tr41, as well as Wap, com-
prise 313, 878, and 1560 documents. They are included in 8 and 10, as well as 20
classes, respectively. The details can be accessible in [69].
Classic4. It contains 2000 documents along with four classes: CACM and CRAN,
as well as CISI, in addition to MED (i.e., 500 documents in each class). This
dataset originally comprises 7,095 documents. They are included in any of four
previous categories, whereas Classic3, as well as Classic4, represent versions of
a similar dataset.

SAD includes scientific articles that have been published in internationally recog-
nized conferences such as (NIPS 2015, as well as AAAI 2013). The SAD’s features
are discussed herein:
144 A. K. Abasi et al.

NIPS 2015. Kaggle site has provided this dataset, which comprises 403 articles.
These articles were published in the conference of Neural Informa-
tion Processing Systems (NIPS). In the machine learning domain, this
conference is a fundamentally ranked conference. The topics involved
deep learning and computer vision, as well as cognitive science, in
addition to reinforcement learning. The NIPS 2015 dataset contains
the paper ID along with the paper title, as well as the type of event
(poster and oral or spotlight presentation). It also contains the pdf file
name, the abstract, and the paper text. However, the title of the paper,
the abstract, and the text are utilized in the experimentation of this
paper. Most articles are associated with machine learning, as well as
natural language processing.
AAAI 2013. The UCI has repository has provided this dataset. There are 150 articles
in this dataset. These articles have been accepted by another confer-
ence, which is a fundamentally ranked conference. It is the AI domain
(known as AAAI 2013). Each of the papers contains information that
includes the title of the paper, topics (i.e., keywords that are author-
selected low-level from the provided list of the conference), keywords
(i.e., keywords that are author-generated), the paper abstract, as well as
keywords of high-level (i.e., keywords that are author-selected high-
level from the provided list of the conference). Many articles revolved
around artificial intelligence like multi-agent systems and reasoning,
as well as machine learning like data mining, in addition to knowledge
discovery, and so on.

Table 1 shows related information given in each dataset. This information involves
the number of the dataset, the dataset’s name, objects or the number of documents,
the clusters’ number, in addition to the number of features. Thus, it is important
that the introduced H-SSA’s parameter values are identified, in addition to additional
comparative algorithms. Table 2 shows a comparison between all the algorithmic
parameters of algorithms.

7.3 Evaluation Measures

Considering the methods of text clustering,based, seven measures of the external

evaluation were normally utilized. These measures included accuracy and recall, as
well as precision, in addition to F-measure, purity, along with entropy [29]. The
utilized measures have been calculated after the findings have been achieved. The
calculation of these measures is provided in detail herein.
A Hybrid Salp Swarm Algorithm with β-Hill Climbing Algorithm … 145

Table 1 Datasets description

Type ID Datasets Number of Number of Number of
objects/Documents clusters (K) features (t)
DCD
DS1 CMC 1473 3 9
DS2 Iris 150 3 4
DS3 Seeds 210 3 7
DS4 Glass 214 7 9
DS5 Wine 178 3 13
TCD
DS6 CSTR 299 4 1725
DS7 20Newsgroups 300 3 2275
DS8 tr12 313 8 5329
DS9 tr41 878 10 6743
DS10 Wap 1560 20 7512
DS11 Classic4 2000 4 6500
SAD
DS12 NIPS 2015 403 2 22888
DS13 AAAI 2013 150 4 1897

Table 2 Parametric values for all algorithms being compared

Parameters Algorithm Value
Population size All Optimization algorithms 20
Maximum number of iteration All Optimization algorithms 1000
runs All Optimization algorithms 30
WEP Max proposed method (H-MVO) 1.00
WEP Min proposed method (H-MVO) 0.20
p proposed method (H-MVO) 6.00
Crossover probability GA 0.80
Mutation probability GA 0.02
Maximum inertia weight PSO 0.90
Minimum inertia weight PSO 0.20
C1 PSO 2.00
C2 PSO 2.00
Vf KHA 0.02
Dmax KHA 0.002
N Max KHA 0.05
P A R Min HS 0.45
P A R Max HS 0.90
bwMin HS 0.10
bwMax HS 1.00
HMCR HS 0.90
146 A. K. Abasi et al.

7.3.1 Accuracy

This measure has been used to compute objects or the percentage of text documents
assigned to correct clusters [30, 48] and calculation can be done utilizing Eq. (11).

1
k
Ac = ni j , (11)
n j=1

whereby n refers to objects or the number of text documents, k refers to the number
of clusters, whereas n i j represents the correct number of i text documents in the j
cluster.

7.3.2 Precision

The assignment ratio of objects or the text documents is assigned by using this
measure of the best clusters for the total number of cluster documents [35]. By using
Eq. (12), the precision of the class i in the cluster j can be calculated.
n i, j
P(i, j) = , (12)
nj

whereby n j signifies the number of the total documents in the cluster j, and ni, j
represents the number of the correctly assigned objects or text documents of the class
i in the cluster j.

7.3.3 Recall

Recall provides the correctly assigned objects’ ratio or text documents’ ratio to the
overall number of related documents (namely, the overall number of documents in
the specific class). By formulating Eq. (13), the recall measure can be calculated.
n i, j
R(i, j) = , (13)
ni

whereby ni, j signifies the number of the correctly assigned objects or text documents
of the class i in the cluster j, and n i signifies the total number of documents in the
class i.
A Hybrid Salp Swarm Algorithm with β-Hill Climbing Algorithm … 147

7.3.4 F-Measure

This measure is external, which represents a harmonic combination between preci-

sion, as well as recall. The closer the F-measure value to 1, the stronger the clustering
solution will be [3]. For computing the F-measure, Eq. (14) is used.
2 × P(i, j) × R(i, j)
F(i, j) = , (14)
P(i, j) + R(i, j)

whereby P(i, j) signifies the class i precision in the cluster j, and R(i, j) signifies
the class i recall in the cluster j. Equation (15) shows the F-measure calculation of
the entire clusters.


k
nj
F= max F(i, j) (15)
i=1
n

7.3.5 Purity Measure

This measure computes the percentage of each cluster in a large class via assigning
each cluster to the most recurrent class [67]. Accordingly, the best-obtained value of
purity is close to 1, given that a large-size rate of class in each cluster is computed
for agreement with a specific estimated cluster’s size. By using Eq. (16), the purity
measure of the entire clusters can be calculated.

1
k
purit y = max(i, j), (16)
n i=1

whereby max(i, j) refers to a large class i size in the cluster j, k signifies the number
of clusters, as well as n, which refers to the entire number of documents in this dataset.

7.3.6 Entropy Measure

The objects or text documents’ distribution for a single class to clusters is analyzed
by this measure. For every single class, this measure assesses the distribution of
documents to the correct clusters. To calculate the cluster’s entropy, a couple of steps
are followed: (i) the distribution of documents for clusters for every single class can
be calculated, (ii) all the entropies, which are obtained by the first step, are used to
calculate the clusters’ entropy. The cluster j’s entropy can be computed by utilizing
Eq. (17).

E(k j) = − p(i, j)log( pi , j), (17)
i=1
148 A. K. Abasi et al.

whereby E(k j) refers to the cluster j’s entropy, as well as p(i, j), which refers to
the objects or text documents’ probability in the cluster i that belongs to the class j.
Equation (18) is used to calculate entropy for the entire clusters.


K
ni
Entr opy = − E(k j), (18)
i=1
n

whereby K refers to the number of clusters, n i refers to the number of documents in

the cluster i, whereas n refers to the total number of documents in datasets.

7.4 Analysis of Results

To measure the obtained clusters via the introduced H-SSA algorithm, as well as other
algorithms, six measures of the evaluation were utilized. The obtained experimental
results showed that the introduced H-SSA2 algorithm has effectively worked to tackle
the TDC problem. The evaluation has been performed based on the performance
of the algorithm, being compared with current comparative algorithms. Thus, the
results of Five DCD and six TCD, as well as two SAD, were achieved according
to accuracy and recall, as well as precision, in addition to F-measure and purity,
along with entropy as shown in Table 3, Table 4, and Table 5, respectively. The
optimal results were highlighted with bold font. These results were reached by the
introduced H-SSA2 in all datasets according to the measure of accuracy, that is,
(DS1 and DS2), then by the H-SSA1 in DCD, as well as TCD, along with H-PSO in
SAD. In accordance with the measure of evaluation, which represents one of the most
commonly utilized standard measures in the text clustering domain, the improvement
of the initial solutions’ starting point and the optimal solution at every single iteration
were found to be effective in comparison with the original SSA algorithm and the
remaining comparative algorithms.

7.5 Convergence Analysis

The convergence behavior shows the effectiveness of various versions of H-SSA

(namely, H-SSA1, as well as H-SSA2) compared with current state-of-the-art meth-
ods. The convergence rate of the clustering algorithm indicates an evaluation crite-
rion, which aims to find the best solution. Figures 3 and 4, as well as Fig. 5 demonstrate
the behaviors of convergence of HS, GA, PSO, KHA, SSA, H-PSO, H-GA, H-SSA1,
as well as H-SSA2 on specific datasets; 30 runs were conducted for each dataset.
An average value has been computed depending on the convergence behavior of
each algorithm. Also, the documents’ average distance to the values of the cluster’s
centroid (ADDC) was plotted in comparison with 1000 iterations on 13 datasets.
Table 3 Results of entropy, purity, F-measure, recall, precision, and accuracy for five data clustering datasets
Dataset Measure Clustering techniques Optimization algorithms
Spectral K-mean K- Agglomerative DBSCAN BHC MVO KHA PSO H-PSO HS GA H-GA H-SSA1 H-SS2
mean++
DS1 Accuracy 0.5956 0.6412 0.5937 0.6464 0.6167 0.6434 0.6500 0.6244 0.6190 0.6774 0.5904 0.5990 0.6064 0.6695 0.6854
Precision 0.6226 0.7281 0.6596 0.6934 0.6266 0.7296 0.6929 0.6984 0.6709 0.6514 0.6193 0.6599 0.6673 0.7210 0.7309
Recall 0.5797 0.6593 0.6178 0.6855 0.5848 0.6747 0.6750 0.6905 0.6381 0.6315 0.5865 0.6411 0.6025 0.6531 0.6769
F- 0.5909 0.6789 0.6213 0.6739 0.5893 0.6876 0.6678 0.6847 0.6440 0.6687 0.5864 0.6314 0.6674 0.6499 0.6620
measure
Purity 0.6054 0.7004 0.6454 0.7017 0.6124 0.7158 0.6917 0.7049 0.6631 0.6529 0.6138 0.6631 0.6228 0.6891 0.7087
Entropy 0.5029 0.5087 0.5040 0.5170 0.4723 0.4998 0.4553 0.5119 0.4237 0.4549 0.5276 0.5094 0.4216 0.4162 0.4100
Rank 14 5 12 7 13 3 4 6 9 8 15 11 10 2 1
DS2 Accuracy 0.6768 0.6738 0.6677 0.7242 0.6866 0.6708 0.7204 0.6993 0.6722 0.7203 0.6908 0.7190 0.7278 0.7221 0.7314
Precision 0.7353 0.6940 0.6919 0.7703 0.7055 0.6949 0.7624 0.7747 0.7343 0.7417 0.7481 0.8056 0.7741 0.7876 0.8013
Recall 0.7140 0.6501 0.6639 0.7558 0.6584 0.6699 0.7456 0.7676 0.6921 0.7476 0.7154 0.7372 0.7624 0.7630 0.7797
F- 0.7094 0.6613 0.6562 0.7476 0.6659 0.6749 0.7447 0.7549 0.6946 0.7629 0.7215 0.7530 0.7965 0.7776 0.7513
measure
Purity 0.7410 0.6813 0.6835 0.7768 0.6854 0.6915 0.7681 0.7772 0.7153 0.7442 0.7395 0.7723 0.7065 0.7595 0.7668
Entropy 0.5284 0.4188 0.4163 0.5573 0.3918 0.4144 0.4916 0.5467 0.4628 0.5337 0.5280 0.4778 0.4630 0.4430 0.4570
Rank 10 15 14 7 12 13 5 6 11 8 9 3 4 2 1
DS3 Accuracy 0.6975 0.7247 0.7020 0.7779 0.6026 0.7038 0.7577 0.7328 0.7231 0.7858 0.6298 0.7021 0.7588 0.7722 0.7916
A Hybrid Salp Swarm Algorithm with β-Hill Climbing Algorithm …

Precision 0.6786 0.7290 0.7604 0.8146 0.6973 0.6851 0.7921 0.7976 0.7721 0.7996 0.7196 0.7658 0.7816 0.8038 0.8249
Recall 0.6860 0.6922 0.7461 0.8267 0.6927 0.6934 0.7783 0.7958 0.7420 0.7518 0.6876 0.7180 0.7516 0.7861 0.8069
F- 0.6557 0.6884 0.7373 0.8145 0.6738 0.6692 0.7719 0.7842 0.7501 0.7722 0.6881 0.7215 0.7713 0.6303 0.8179
measure
Purity 0.7138 0.7152 0.7682 0.8339 0.7013 0.7287 0.7916 0.8101 0.7697 0.8001 0.7171 0.7521 0.7081 0.7749 0.8374
Entropy 0.5053 0.4680 0.4818 0.4119 0.4474 0.4638 0.5326 0.4757 0.4148 0.3697 0.4638 0.4328 0.4028 0.3853 0.3779
Rank 14 11 9 2 15 12 7 4 8 3 13 10 6 5 1
(continued)
149
 150

Table 3 (continued)
Dataset Measure Clustering techniques Optimization algorithms
Spectral K-mean K- Agglomerative DBSCAN BHC MVO KHA PSO H-PSO HS GA H-GA H-SSA1 H-SS2
mean++
DS4 Accuracy 0.4966 0.5191 0.5424 0.5826 0.5499 0.5937 0.5345 0.5230 0.5066 0.5490 0.5056 0.4964 0.5446 0.5699 0.5876
Precision 0.5620 0.5366 0.6310 0.6108 0.6028 0.6293 0.6024 0.5812 0.5689 0.6032 0.5296 0.5220 0.6716 0.6463 0.6600
Recall 0.5231 0.4901 0.5652 0.6341 0.5938 0.6442 0.5922 0.5438 0.5432 0.6228 0.4820 0.4970 0.6340 0.6300 0.6752
F- 0.5302 0.4912 0.5859 0.6149 0.5827 0.6196 0.5865 0.5472 0.5392 0.5902 0.4967 0.4914 0.5337 0.5929 0.5971
measure
Purity 0.5554 0.5179 0.6072 0.6392 0.6105 0.6524 0.6090 0.5664 0.5659 0.6214 0.5122 0.5198 0.6562 0.6101 0.6603
Entropy 0.4114 0.4189 0.5202 0.4592 0.5036 0.5095 0.5661 0.4704 0.4389 0.4364 0.3782 0.4148 0.4262 0.4255 0.4100
Rank 12 14 8 4 7 2 9 10 11 6 13 15 5 3 1
DS5 Accuracy 0.4496 0.4755 0.5016 0.4877 0.5107 0.5204 0.5345 0.4782 0.4638 0.5672 0.4645 0.4481 0.5055 0.5287 0.5770
Precision 0.5169 0.4899 0.5912 0.5633 0.5570 0.5597 0.5665 0.5415 0.5245 0.5415 0.4829 0.4799 0.5629 0.5777 0.6127
Recall 0.4744 0.4463 0.5194 0.5504 0.5452 0.5455 0.5897 0.5014 0.4957 0.5734 0.4355 0.4550 0.5125 0.5479 0.5749
F- 0.4823 0.4512 0.5445 0.5465 0.5388 0.5510 0.5662 0.5015 0.4925 0.5545 0.4573 0.4444 0.5393 0.5673 0.5917
measure
Purity 0.5122 0.4786 0.5667 0.5644 0.5681 0.5682 0.5998 0.5267 0.5230 0.5667 0.4719 0.4752 0.5679 0.5638 0.6161
Entropy 0.5222 0.5186 0.6067 0.5664 0.6071 0.5978 0.6188 0.5877 0.5970 0.5607 0.5109 0.5032 0.5269 0.5087 0.4943
Rank 11 13 8 7 9 6 4 10 12 3 14 15 5 2 1
Average 12.2 11.6 10.2 5.4 11.2 7.2 5.8 7.2 10.2 5.6 12.8 10.8 6.0 2.8 1.0
Final 14 13 9 3 12 7 5 7 9 4 15 11 6 2 1
Rank
Note The lowest ranked algorithm is the best one
A. K. Abasi et al.
Table 4 Results of entropy, purity, F-measure, recall, precision, and accuracy for six text clustering datasets
Dataset Measure Clustering techniques Optimization algorithms
Spectral K-mean K- Agglomerative DBSCAN BHC MVO KHA PSO H-PSO HS GA H-GA H-SSA1 H-SS2
mean++
DS6 Accuracy 0.4319 0.3573 0.4355 0.4360 0.4005 0.4365 0.4593 0.3649 0.4356 0.5494 0.4464 0.3399 0.5056 0.5648 0.5717
Precision 0.3597 0.4091 0.3953 0.4423 0.3393 0.3766 0.5715 0.4213 0.5340 0.6065 0.4235 0.4417 0.6130 0.6385 0.6317
Recall 0.4925 0.3092 0.4076 0.4666 0.4256 0.5063 0.4829 0.5355 0.4360 0.5600 0.5060 0.3418 0.4950 0.5026 0.5439
F- 0.3971 0.3460 0.3546 0.3266 0.3046 0.4133 0.5244 0.4139 0.4819 0.5577 0.3377 0.3886 0.5859 0.5668 0.5864
measure
Purity 0.4485 0.3525 0.4096 0.4816 0.4076 0.4524 0.5685 0.3874 0.4953 0.6135 0.4355 0.4050 0.5743 0.5012 0.5827
Entropy 0.4893 0.8201 0.5246 0.5076 0.4586 0.8018 0.5207 0.4344 0.6199 0.5076 0.4786 0.7170 0.5104 0.4967 0.5016
Rank 10.0000 15.0000 11.0000 9.0000 12.0000 13.0000 5.0000 7.0000 6.0000 2.0000 8.0000 14.0000 4.0000 3.0000 1.0000
DS7 Accuracy 0.3633 0.3180 0.3784 0.4055 0.3038 0.3695 0.4044 0.3216 0.3498 0.4442 0.3122 0.3676 0.4308 0.5070 0.5185
Precision 0.3424 0.3121 0.3652 0.3990 0.3094 0.3526 0.4392 0.3829 0.4134 0.4821 0.3601 0.4209 0.4814 0.4915 0.5456
Recall 0.3280 0.3100 0.3662 0.3576 0.3017 0.3340 0.3842 0.3136 0.3497 0.4220 0.3170 0.3676 0.4887 0.4292 0.4753
F- 0.3136 0.3406 0.3619 0.3548 0.3193 0.3164 0.4109 0.2996 0.3803 0.4454 0.3214 0.3936 0.4433 0.5127 0.5180
measure
Purity 0.3110 0.3741 0.4134 0.4417 0.3027 0.3281 0.4344 0.3421 0.4097 0.4725 0.3355 0.4081 0.4837 0.5077 0.4426
Entropy 0.6125 0.8028 0.6611 0.5953 0.7473 0.7940 0.7121 0.6767 0.7723 0.6631 0.6481 0.7547 0.6833 0.6456 0.6475
Rank 10 14 7 5 15 13 6 12 9 4 11 8 3 2 1
DS8 Accuracy 0.3373 0.2971 0.3795 0.4481 0.3045 0.3881 0.4485 0.3357 0.4075 0.4796 0.3776 0.3677 0.5205 0.5267 0.5448
A Hybrid Salp Swarm Algorithm with β-Hill Climbing Algorithm …

Precision 0.4508 0.3522 0.4215 0.4923 0.3218 0.4379 0.5075 0.3748 0.4298 0.5355 0.4385 0.4128 0.4798 0.5895 0.6234
Recall 0.3055 0.2944 0.3778 0.4341 0.3234 0.3812 0.4398 0.3099 0.4264 0.3893 0.3453 0.3549 0.4974 0.5123 0.5095
F- 0.3781 0.3222 0.4176 0.4090 0.4048 0.4258 0.4706 0.2916 0.4278 0.5860 0.4470 0.3826 0.5428 0.5671 0.5797
measure
Purity 0.4132 0.3908 0.4808 0.5553 0.4728 0.4926 0.5448 0.3803 0.4878 0.5876 0.4986 0.4513 0.6038 0.5929 0.5894
Entropy 0.4932 0.7138 0.5094 0.3978 0.5010 0.7116 0.5224 0.5183 0.5720 0.5757 0.4517 0.6233 0.6380 0.5697 0.5618
Rank 11 15 9 5 13 10 6 14 8 4 7 12 3 2 1
(continued)
151
Table 4 (continued)
152

Dataset Measure Clustering techniques Optimization algorithms

Spectral K-mean K- Agglomerative DBSCAN BHC MVO KHA PSO H-PSO HS GA H-GA H-SSA1 H-SS2
mean++
DS9 Accuracy 0.3787 0.4126 0.4260 0.4610 0.3704 0.3728 0.4630 0.4054 0.4870 0.4740 0.3590 0.4320 0.5940 0.5524 0.6130
Precision 0.3630 0.3945 0.3769 0.3445 0.4242 0.4325 0.4569 0.4191 0.4505 0.5010 0.3710 0.4140 0.5245 0.5361 0.5557
Recall 0.3480 0.3813 0.3559 0.3487 0.4291 0.4391 0.4419 0.4579 0.4497 0.4768 0.3888 0.4008 0.5127 0.5019 0.5020
F- 0.3691 0.3876 0.4299 0.4187 0.4152 0.4342 0.4569 0.4137 0.4497 0.4701 0.3701 0.4071 0.5647 0.5074 0.5164
measure
Purity 0.4866 0.4108 0.5471 0.4760 0.6025 0.6208 0.6081 0.6063 0.5790 0.5273 0.4863 0.5603 0.6420 0.6361 0.6476
Entropy 0.5386 0.5874 0.4745 0.4821 0.4776 0.5850 0.5355 0.4041 0.5391 0.5029 0.4539 0.5469 0.5371 0.4890 0.4808
Rank 14 15 11 12 8 9 6 5 7 4 13 10 2 3 1
DS10 Accuracy 0.4905 0.5012 0.4937 0.4933 0.4846 0.5038 0.5291 0.5191 0.5623 0.5532 0.5032 0.5316 0.6523 0.5812 0.6674
Precision 0.4454 0.4626 0.4913 0.4479 0.5004 0.4571 0.5213 0.4773 0.5249 0.5450 0.4630 0.5314 0.5749 0.5900 0.6213
Recall 0.3423 0.4011 0.3508 0.4471 0.4176 0.3601 0.4496 0.4106 0.4811 0.4907 0.4447 0.4706 0.5671 0.5807 0.5758
F- 0.4141 0.4315 0.4462 0.4487 0.4878 0.4292 0.4831 0.4131 0.5017 0.5521 0.4611 0.4998 0.6187 0.5888 0.5940
measure
Purity 0.5142 0.4759 0.5797 0.5445 0.4007 0.5146 0.6069 0.5939 0.6125 0.6939 0.5589 0.4917 0.7315 0.6795 0.6938
Entropy 0.5468 0.7044 0.5961 0.4875 0.5706 0.6858 0.6625 0.5595 0.5765 0.6310 0.5470 0.6216 0.6585 0.6216 0.6200
Rank 13 15 11 8 12 14 6 10 5 4 9 7 2 3 1
DS11 Accuracy 0.6326 0.5858 0.6799 0.6683 0.6206 0.6992 0.7043 0.5761 0.6363 0.6853 0.5963 0.6621 0.7273 0.7126 0.7184
Precision 0.6620 0.5699 0.6880 0.6449 0.5708 0.6929 0.6919 0.6246 0.6604 0.7354 0.6064 0.6726 0.7774 0.7066 0.7049
Recall 0.6528 0.5259 0.7028 0.6224 0.5796 0.7109 0.6844 0.5550 0.6164 0.6754 0.5674 0.6320 0.6944 0.6721 0.7032
F- 0.6764 0.5472 0.6920 0.6281 0.6173 0.7074 0.6881 0.5529 0.6377 0.6297 0.5787 0.6519 0.6887 0.6816 0.6987
measure
Purity 0.6352 0.5938 0.6227 0.5812 0.6089 0.6272 0.6742 0.5550 0.6243 0.6413 0.5983 0.6320 0.6733 0.6943 0.7302
Entropy 0.4002 0.5601 0.3540 0.4033 0.4854 0.5504 0.5113 0.5321 0.5306 0.5656 0.4946 0.5781 0.6166 0.5624 0.4347
Rank 7 15 2 9 12 6 4 14 11 8 13 10 3 5 1
Average 10.8 14.8 8.5 8.0 12.0 10.8 5.5 10.3 7.7 4.3 10.2 10.2 2.8 3.0 1.0
Final 12 15 8 7 14 12 5 11 6 4 9 9 2 3 1
Rank
Note The lowest ranked algorithm is the best one
A. K. Abasi et al.
Table 5 Results of entropy, purity, F-measure, recall, precision, and accuracy for two scientific articles datasets
Dataset Measure Clustering techniques Optimization algorithms
Spectral K-mean K- Agglomerative DBSCAN BHC MVO KHA PSO H-PSO HS GA H-GA H-SSA1 H-SS2
mean++
DS12 Accuracy 0.5593 0.5433 0.5545 0.5472 0.5281 0.5646 0.6081 0.6306 0.5802 0.6556 0.6211 0.5503 0.6272 0.6356 0.6576
Precision 0.4929 0.4489 0.4994 0.5340 0.5443 0.4962 0.6113 0.5914 0.6440 0.6384 0.5413 0.6209 0.6246 0.6585 0.6675
Recall 0.3911 0.5211 0.3913 0.4797 0.4706 0.3937 0.5756 0.4986 0.5717 0.6075 0.5156 0.5571 0.5901 0.6119 0.6661
F- 0.4547 0.5197 0.4701 0.4661 0.4251 0.4681 0.5941 0.5298 0.6461 0.6195 0.4991 0.5797 0.5965 0.6114 0.6604
measure
Purity 0.5685 0.5735 0.5262 0.5839 0.6459 0.5689 0.6899 0.5127 0.7089 0.7877 0.6919 0.5015 0.7589 0.6325 0.7279
Entropy 0.5555 0.5415 0.5218 0.5030 0.5565 0.4876 0.5005 0.5556 0.5190 0.4536 0.5915 0.5085 0.4544 0.4534 0.4260
Rank 15 11 14 10 12 13 6 9 5 2 8 7 3 4 1
DS13 Accuracy 0.4188 0.4887 0.4896 0.5635 0.4892 0.5039 0.6238 0.5811 0.6242 0.6811 0.5945 0.5395 0.6685 0.6766 0.6874
Precision 0.4865 0.6323 0.5459 0.5219 0.6175 0.6455 0.6625 0.6139 0.5635 0.7289 0.6125 0.5429 0.6439 0.6699 0.6878
Recall 0.3977 0.6304 0.5734 0.4415 0.5324 0.6381 0.6087 0.5860 0.6414 0.6100 0.6447 0.5255 0.6083 0.6310 0.6478
F- 0.4805 0.6295 0.5043 0.4983 0.5615 0.6376 0.6395 0.6084 0.5875 0.6554 0.6172 0.5103 0.6342 0.6393 0.6529
measure
Purity 0.4280 0.6464 0.5512 0.5299 0.6127 0.6593 0.6380 0.6502 0.6187 0.6572 0.6758 0.5629 0.6456 0.6749 0.6794
Entropy 0.6190 0.6054 0.5402 0.5579 0.6267 0.5506 0.5780 0.5632 0.5917 0.5532 0.5268 0.5859 0.5016 0.5504 0.5340
A Hybrid Salp Swarm Algorithm with β-Hill Climbing Algorithm …

Rank 15 10 12 14 11 7 6 8 9 2 5 13 4 3 1
Average 15.0 10.5 13.0 12.0 11.5 10.0 6.0 8.5 7.0 2.0 6.5 10.0 3.5 3.5 1.0
Final 15 11 14 13 12 9 5 8 7 2 6 9 3 3 1
Rank
Note The lowest ranked algorithm is the best one
153
154 A. K. Abasi et al.

DS1 DS2
0.405

0.4 0.29

0.395
0.28
0.39
Fitness value (ADDC)

Fitness value (ADDC)

0.27
0.385

0.38 0.26

0.375
HS 0.25 HS
0.37 GA GA
PSO PSO
KHA 0.24 KHA
0.365
MVO MVO
H-PSO H-PSO
0.36 H-GA
0.23 H-GA
H-SSA1 H-SSA1
0.355 H-SSA2 H-SSA2
0.22
100 200 300 400 500 600 700 800 900 1000 100 200 300 400 500 600 700 800 900 1000
No. of iteration No. of iteration

DS3 DS4
0.42 0.42

0.4 0.4

0.38
0.38
Fitness value (ADDC)

Fitness value (ADDC)

0.36
0.36
0.34
HS HS
GA 0.34 GA
0.32 PSO PSO
KHA KHA
MVO 0.32 MVO
0.3 H-PSO H-PSO
H-GA H-GA
H-SSA1 H-SSA1
0.28 H-SSA2 0.3 H-SSA2

100 200 300 400 500 600 700 800 900 1000 100 200 300 400 500 600 700 800 900 1000
No. of iteration No. of iteration

DS5
0.4

0.38

0.36
Fitness value (ADDC)

0.34

0.32 HS
GA
PSO
KHA
0.3 MVO
H-PSO
H-GA
H-SSA1
0.28
H-SSA2

100 200 300 400 500 600 700 800 900 1000
No. of iteration

Fig. 3 Optimization algorithms convergence rate for five data clustering datasets

The convergence methods of state of the art were faster compared to H-SSA1, as
well as H-SSA2. Nevertheless, H-SSA2 was found to be more effective in compari-
son with the original SSA regarding the performance of the algorithm and the time
of execution. Thus, a better clustering quality has been generated compared to the
algorithms that are most popular. Even though the basic SSA has slowly approached
local optima compared to hybrid versions, namely, H-SSA1, as well as H-SSA2, it
has more efficiently converged to the best solution. H-SSA2 has achieved the optimal
A Hybrid Salp Swarm Algorithm with β-Hill Climbing Algorithm … 155

DS6 DS7
0.21

0.25
0.205

0.2 0.24

0.195
Fitness value (ADDC)

Fitness value (ADDC)

0.23
0.19

0.185 0.22
HS HS
0.18 GA GA
PSO 0.21 PSO
KHA KHA
0.175 MVO MVO
H-PSO H-PSO
H-GA 0.2 H-GA
0.17
H-SSA1 H-SSA1
H-SSA2 H-SSA2

100 200 300 400 500 600 700 800 900 1000 100 200 300 400 500 600 700 800 900 1000
No. of iteration No. of iteration

DS8 DS9
0.205
0.8
0.2
0.78
0.195
0.76
0.19
Fitness value (ADDC)

Fitness value (ADDC)

0.74
0.185

0.72
0.18

0.175 HS 0.7 HS
GA GA
PSO PSO
0.17 KHA 0.68 KHA
MVO MVO
0.165 H-PSO H-PSO
H-GA
0.66 H-GA
H-SSA1 H-SSA1
0.16 H-SSA2 H-SSA2
0.64
100 200 300 400 500 600 700 800 900 1000 100 200 300 400 500 600 700 800 900 1000
No. of iteration No. of iteration

DS10
0.95 DS11
0.6

0.9
Fitness value (ADDC)

0.85 0.55
Fitness value (ADDC)

0.8
HS
HS GA
GA 0.5 PSO
0.75 PSO KHA
KHA MVO
MVO H-PSO
H-PSO H-GA
0.7 H-GA
H-SSA1
H-SSA1
H-SSA2
H-SSA2 0.45

100 200 300 400 500 600 700 800 900 1000 100 200 300 400 500 600 700 800 900 1000
No. of iteration No. of iteration

Fig. 4 Optimization algorithms convergence rate for six text clustering datasets

solution compared to the comparative algorithms. Also, high-quality clusters were

achieved compared to different comparative algorithms. By examining the conver-
gence behavior of the introduced versions, it was found that the proposed H-SSA2
has achieved the most optimal performance and has quickly achieved optimal results
as well. However, its convergence was found to be quicker than the convergence of
the versions of the SSA algorithm along with known clustering algorithms.
156 A. K. Abasi et al.

DS12 DS13
0.8 0.68

0.66
0.78
0.64
Fitness value (ADDC)

Fitness value (ADDC)

0.76 0.62

0.6
0.74
0.58
HS HS
GA GA
0.72 0.56
PSO PSO
KHA KHA
MVO 0.54 MVO
0.7 H-PSO H-PSO
H-GA H-GA
H-SSA1 0.52 H-SSA1
H-SSA2 H-SSA2
0.68
100 200 300 400 500 600 700 800 900 1000 100 200 300 400 500 600 700 800 900 1000
No. of iteration No. of iteration

Fig. 5 Optimization algorithms convergence rate for two scientific articles datasets

8 Conclusion and Future Works

The problem of text clustering is a major concern. Many researchers have therefore
tried to solve this problem. The population-based SSA algorithm is a novel optimiza-
tion algorithm. It aims at solving numerous global optimization problems. SSA aims
to simultaneously provide good exploration for the search space’s various regions
to find the best solution at the exploitation cost. The aim of this work is to tackle
two critical issues related to the SSA algorithm, involving the initial value of the
objective function for the candidate solutions, along with the most optimal solution
that SSA can produce at every single iteration.
Therefore, a new hybrid algorithm has been developed in this work with the aim
of solving the text clustering problem based on the k-means clustering combination
along with the SSA algorithms to tackle critical issues. Thus, an enhanced SSA
version has been created to enhance the quality of the initial candidate solutions and
enhance the optimal solution as well. By utilizing such a hybridization, H-SSA has
carried out a more efficient and more effective search; it has converged quickly to
optimal solutions.
Considering the assessment of the novel H-SSA, seven evaluation measures were
applied, involving error rate and accuracy, as well as precision, in addition to recall
and F-measure, along with purity, entropy, and the behavior of convergence, along
with statistical analysis. These measures are the most commonly utilized evaluation
criteria in the data and text mining domain, which aim to evaluate the novel cluster-
ing method. Thus, the novel H-SSA can produce the best-recorded findings for the
entire benchmark datasets that are used compared with current versions along with
several clustering methods and techniques that are successfully implemented in the
literature. Therefore, the SSA algorithm with the k-means is an efficient approach
for clustering techniques. Consequently, many success stories are expected in the
domain of data and text clustering. The findings of this work showed that the newly
introduced hybridization can be active, which is an effective method for tackling
A Hybrid Salp Swarm Algorithm with β-Hill Climbing Algorithm … 157

the problems of clustering. The experimental findings were compared with current
comparative algorithms. The novel SSA (H-SSA) hybridization has properly solved
the clustering problems regarding data and text. H-SSA can, therefore, contribute
to the domain of clustering. Also, different clustering problems are expected to be
explored in future research to confirm the capability of the introduced algorithm
in this domain. Accordingly, a more powerful local search can be hybridized to
achieve further enhancements of the SSA exploitation capability. In conclusion, the
introduced algorithm can be further investigated regarding the benchmark function
datasets in future works.

References

1. Abasi, Ammar Kamal, Ahamad Tajudin Khader, Mohammed Azmi Al-Betar, Syibrah Naim,
Zaid Abdi Alkareem Alyasseri, and Sharif Naser Makhadmeh. 2020. An ensemble topic extrac-
tion approach based on optimization clusters using hybrid multi-verse optimizer for scientific
publications. Journal of Ambient Intelligence and Humanized Computing 1–37.
2. Abasi, Ammar Kamal, Ahamad Tajudin Khader, Mohammed Azmi Al-Betar, Syibrah Naim,
Zaid Abdi Alkareem Alyasseri, and Sharif Naser Makhadmeh. 2020. A novel hybrid multi-verse
optimizer with k-means for text documents clustering. Neural Computing & Applications.
3. Abasi, Ammar Kamal, Ahamad Tajudin Khader, Mohammed Azmi Al-Betar, Syibrah Naim,
Sharif Naser Makhadmeh, and Zaid Abdi Alkareem Alyasseri. 2019. An improved text feature
selection for clustering using binary grey wolf optimizer. In Proceedings of the 11th national
technical seminar on unmanned system technology 2019, 503–516. Springer.
4. Abasi, Ammar Kamal, Ahamad Tajudin Khader, Mohammed Azmi Al-Betar, Syibrah Naim,
Sharif Naser Makhadmeh, and Zaid Abdi Alkareem Alyasseri. 2019. A text feature selection
technique based on binary multi-verse optimizer for text clustering. In 2019 IEEE Jordan
international joint conference on electrical engineering and information technology (JEEIT),
1–6. IEEE.
5. Abasi, Ammar Kamal, Ahamad Tajudin Khader, Mohammed Azmi Al-Betar, Syibrah Naim,
Sharif Naser Makhadmeh, and Zaid Abdi Alkareem Alyasseri. 2020. Link-based multi-verse
optimizer for text documents clustering. Applied Soft Computing 87: 106002.
6. Abed-alguni, Bilal H., and Faisal Alkhateeb. 2018. Intelligent hybrid cuckoo search and β-hill
climbing algorithm. Journal of King Saud University-Computer and Information Sciences.
7. Abualigah, Laith Mohammad, Ahamad Tajudin Khader, Mohammed Azmi Al-Betar, and Zaid
Abdi Alkareem Alyasseri. 2017. β-hill climbing technique for the text document clustering.
In New trends in information technology NTIT2017 conference, Amman, Jordan, 1–6. IEEE.
8. Al-Betar, Mohammed Azmi. 2017. β-hill climbing: An exploratory local search. Neural Com-
puting and Applications 28 (1): 153–168.
9. Al-Betar, Mohammed Azmi, Ibrahim Aljarah, Mohammed A Awadallah, Hossam Faris, and
Seyedali Mirjalili. 2019. Adaptive β-hill climbing for optimization. Soft Computing 1–24.
10. Al-Betar, Mohammed Azmi, Mohammed A. Awadallah, Iyad Abu Doush, Emad Alsukhni, and
Habes ALkhraisat. 2018. A non-convex economic dispatch problem with valve loading effect
using a new modified β-hill climbing local search algorithm. Arabian Journal for Science and
Engineering.
11. Al-Betar, Mohammed Azmi, Mohammed A. Awadallah, Asaju Laaro Bolaji, and Basem O.
Alijla. 2017. β-hill climbing algorithm for sudoku game. In Second Palestinian international
conference on information and communication technology (PICICT 2017), Gaza, Palestine,
1–5. IEEE.
158 A. K. Abasi et al.

12. Al-Betar, Mohammed Azmi, Mohammed A. Awadallah, Ahamad Tajudin Khader, Asaju Laaro
Bolaji, and Ammar Almomani. 2018. Economic load dispatch problems with valve-point load-
ing using natural updated harmony search. Neural Computing and Applications 29 (10): 767–
781.
13. Al-Betar, Mohammed Azmi, and Ahamad Tajudin Khader. 2012. A harmony search algorithm
for university course timetabling. Annals of Operations Research 194 (1): 3–31.
14. Alomari, Osama Ahmad, Ahamad Tajudin Khader, Mohammed Azmi Al-Betar, and Zaid
Abdi Alkareem Alyasseri. 2018. A hybrid filter-wrapper gene selection method for cancer
classification. In 2018 2nd international conference on biosignal analysis, processing and
systems (ICBAPS), 113–118. IEEE.
15. Alomari, Osama Ahmad, Ahamad Tajudin Khader, Mohammed Azmi Al-Betar, and
Mohammed A. Awadallah. 2018. A novel gene selection method using modified MRMR and
hybrid bat-inspired algorithm with β-hill climbing. Applied Intelligence 48 (11): 4429–4447.
16. Alsukni, Emad, Omar Suleiman Arabeyyat, Mohammed A. Awadallah, Laaly Alsamarraie,
Iyad Abu-Doush, and Mohammed Azmi Al-Betar. 2019. Multiple-reservoir scheduling using
β-hill climbing algorithm. Journal of Intelligent Systems 28 (4): 559–570.
17. Alweshah, Mohammed, Aram Al-Daradkeh, Mohammed Azmi Al-Betar, Ammar Almomani,
and Saleh Oqeili. 2019. β-hill climbing algorithm with probabilistic neural network for clas-
sification problems. Journal of Ambient Intelligence and Humanized Computing 1–12.
18. Alyasseri, Zaid Abdi Alkareem, Ahamad Tajudin Khader, and Mohammed Azmi Al-Betar.
2017. Optimal electroencephalogram signals denoising using hybrid β-hill climbing algorithm
and wavelet transform. In Proceedings of the international conference on imaging, signal
processing and communication, 106–112. ACM.
19. Alyasseri, Zaid Abdi Alkareem, Ahamad Tajudin Khader, Mohammed Azmi Al-Betar, Ammar
Kamal Abasi, and Sharif Naser Makhadmeh. 2019. EEG signals denoising using optimal
wavelet transform hybridized with efficient metaheuristic methods. IEEE Access 8: 10584–
10605.
20. Alyasseri, Zaid Abdi Alkareem, Ahamad Tajudin Khader, Mohammed Azmi Al-Betar, and
Laith Mohammad Abualigah. 2017. ECG signal denoising using β-hill climbing algorithm and
wavelet transform. In ICIT 2017, the 8th international conference on information technology,
1–7.
21. Alyasseri, Zaid Abdi Alkareem, Ahamad Tajudin Khader, Mohammed Azmi Al-Betar, and
Osama Ahmad Alomari. 2020. Person identification using EEG channel selection with hybrid
flower pollination algorithm. Pattern Recognition 107393.
22. Alyasseri, Zaid Abdi Alkareem, Ahamad Tajudin Khader, Mohammed Azmi Al-Betar, and
Mohammed A Awadallah. 2018. Hybridizing β-hill climbing with wavelet transform for
denoising ECG signals. Information Sciences 429: 229–246.
23. Alyasseri, Zaid Abdi Alkareem, Ahamad Tajudin Khader, Mohammed Azmi Al-Betar,
Mohammed A. Awadallah, and Xin-She Yang. 2018. Variants of the flower pollination algo-
rithm: A review. In Nature-inspired algorithms and applied optimization, 91–118. Springer.
24. Alzaidi, Amer Awad, Musheer Ahmad, Mohammad Najam Doja, Eesa Al Solami, and M.M.
Sufyan Beg. 2018. A new 1D chaotic map and beta-hill climbing for generating substitution-
boxes. IEEE Access 6: 55405–55418.
25. Bharti, Kusum Kumari, and Pramod Kumar Singh. 2016. Chaotic gradient artificial bee colony
for text clustering. Soft Computing 20 (3): 1113–1126.
26. Boley, Daniel, Maria Gini, Robert Gross, Eui-Hong Sam Han, Kyle Hastings, George Karypis,
Vipin Kumar, Bamshad Mobasher, and Jerome Moore. 1999. Document categorization and
query generation on the world wide web using webace. Artificial Intelligence Review 13 (5–6):
365–391.
27. Bouras, Christos, and Vassilis Tsogkas. 2012. A clustering technique for news articles using
wordnet. Knowledge-Based Systems 36: 115–128.
28. Cura, Tunchan. 2012. A particle swarm optimization approach to clustering. Expert Systems
with Applications 39 (1): 1582–1588.
A Hybrid Salp Swarm Algorithm with β-Hill Climbing Algorithm … 159

29. Deepa, M., P. Revathy, and P. Student. 2012. Validation of document clustering based on
purity and entropy measures. International Journal of Advanced Research in Computer and
Communication Engineering 1 (3): 147–152.
30. Del Buono, Nicoletta, and Gianvito Pio. 2015. Non-negative matrix tri-factorization for co-
clustering: An analysis of the block matrix. Information Sciences 301: 13–26.
31. Ekinci, Serdar, and Baran Hekimoglu. 2018. Parameter optimization of power system stabilizer
via salp swarm algorithm. In 2018 5th international conference on electrical and electronic
engineering (ICEEE), 143–147. IEEE.
32. Elaziz, Mohamed Abd, Lin Li, KPN Jayasena, and Shengwu Xiong. 2019. Multiobjective big
data optimization based on a hybrid salp swarm algorithm and differential evolution. Applied
Mathematical Modelling.
33. Faris, Hossam, Seyedali Mirjalili, Ibrahim Aljarah, Majdi Mafarja, and Ali Asghar Heidari.
2020. Salp swarm algorithm: Theory, literature review, and application in extreme learning
machines. In Nature-inspired optimizers, 185–199. Springer.
34. Figueiredo, Elliackin, Mariana Macedo, Hugo Valadares Siqueira, Clodomir J. Santana Jr,
Anu Gokhale, and Carmelo J.A. Bastos-Filho. 2019. Swarm intelligence for clustering a sys-
tematic review with new perspectives on data mining. Engineering Applications of Artificial
Intelligence 82: 313–329.
35. Forsati, Rana, Andisheh Keikha, and Mehrnoush Shamsfard. 2015. An improved bee colony
optimization algorithm with an application to document clustering. Neurocomputing 159: 9–26.
36. Forsati, Rana, Mehrdad Mahdavi, Mehrnoush Shamsfard, and Mohammad Reza Meybodi.
2013. Efficient stochastic algorithms for document clustering. Information Sciences 220: 269–
291.
37. Hossein, Gandomi Amir, and Amir Hossein Alavi. 2012. Krill herd: A new bio-inspired opti-
mization algorithm. Communications in Nonlinear Science and Numerical Simulation 17 (12):
4831–4845.
38. Hegazy, AhE, M.A. Makhlouf, and GhS El-Tawel. 2019. Feature selection using chaotic salp
swarm algorithm for data classification. Arabian Journal for Science and Engineering 44 (4):
3801–3816.
39. Huang, Anna. 2008. Similarity measures for text document clustering. In Proceedings of
the sixth New Zealand computer science research student conference (NZCSRSC2008),
Christchurch, New Zealand, 49–56.
40. Ibrahim, Rehab Ali, Ahmed A. Ewees, Diego Oliva, Mohamed Abd Elaziz, and Lu Songfeng.
2019. Improved salp swarm algorithm based on particle swarm optimization for feature selec-
tion. Journal of Ambient Intelligence and Humanized Computing 10 (8): 3155–3169.
41. Ismael, Sherif M., Shady H.E. Abdel Aleem, Almoataz Y. Abdelaziz, and Ahmed Faheem
Zobaa. 2018. Practical considerations for optimal conductor reinforcement and hosting capacity
enhancement in radial distribution systems. IEEE Access 6: 27268–27277.
42. Jangir, Pradeep, Siddharth A. Parmar, Indrajit N. Trivedi, and R.H. Bhesdadiya. 2017. A novel
hybrid particle swarm optimizer with multi verse optimizer for global numerical optimiza-
tion and optimal reactive power dispatch problem. Engineering Science and Technology, an
International Journal 20 (2): 570–586.
43. Jensi, R., and G. Wiselin Jiji. 2014. A survey on optimization approaches to text document
clustering. arXiv:1401.2229.
44. Karaa, Wahiba Ben Abdessalem, Amira S. Ashour, Dhekra Ben Sassi, Payel Roy, Noreen
Kausar, and Nilanjan Dey. 2016. Medline text mining: An enhancement genetic algorithm
based approach for document clustering. In Applications of intelligent optimization in biology
and medicine, 267–287. Springer.
45. Karaboga, Dervis, and Celal Ozturk. 2011. A novel clustering approach: Artificial bee colony
(ABC) algorithm. Applied Soft Computing 11 (1): 652–657.
46. Katrawi, Anwar H., Rosni Abdullah, Mohammed Anbar, and Ammar Kamal Abasi. 2020.
Earlier stage for straggler detection and handling using combined CPU test and LATE method-
ology. International Journal of Electrical & Computer Engineering 10. ISSN: 2088-8708.
160 A. K. Abasi et al.

47. Kaveh, A., and M. Khayatazad. 2012. A new meta-heuristic method: Ray optimization. Com-
puters & Structures 112: 283–294.
48. Lin, Yung-Shen, Jung-Yi Jiang, and Shie-Jue Lee. 2014. A similarity measure for text clas-
sification and clustering. IEEE Transactions on Knowledge and Data Engineering 26 (7):
1575–1590.
49. Makhadmeh, Sharif Naser, Ahamad Tajudin Khader, Mohammed Azmi Al-Betar, and Syibrah
Naim. 2018. An optimal power scheduling for smart home appliances with smart battery using
grey wolf optimizer. In 2018 8th IEEE international conference on control system, computing
and engineering (ICCSCE), 76–81. IEEE.
50. Makhadmeh, Sharif Naser, Ahamad Tajudin Khader, Mohammed Azmi Al-Betar, and Syi-
brah Naim. 2019. Multi-objective power scheduling problem in smart homes using grey wolf
optimiser. Journal of Ambient Intelligence and Humanized Computing 10 (9): 3643–3667.
51. Makhadmeh, Sharif Naser, Ahamad Tajudin Khader, Mohammed Azmi Al-Betar, Syibrah
Naim, Ammar Kamal Abasi, and Zaid Abdi Alkareem Alyasseri. 2019. Optimization methods
for power scheduling problems in smart home: Survey. Renewable and Sustainable Energy
Reviews 115: 109362.
52. Makhadmeh, Sharif Naser, Ahamad Tajudin Khader, Mohammed Azmi Al-Betar, Syibrah
Naim, Zaid Abdi Alkareem Alyasseri, and Ammar Kamal Abasi. 2019. Particle swarm opti-
mization algorithm for power scheduling problem using smart battery. In 2019 IEEE Jordan
international joint conference on electrical engineering and information technology (JEEIT),
672–677. IEEE.
53. Mirjalili, Seyedali. 2015. The ant lion optimizer. Advances in Engineering Software 83: 80–98.
54. Mirjalili, Seyedali. 2016. Dragonfly algorithm: A new meta-heuristic optimization technique
for solving single-objective, discrete, and multi-objective problems. Neural Computing and
Applications 27 (4): 1053–1073.
55. Mirjalili, Seyedali, Amir H. Gandomi, Seyedeh Zahra Mirjalili, Shahrzad Saremi, Hossam
Faris, and Seyed Mohammad Mirjalili. 2017. Salp swarm algorithm: A bio-inspired optimizer
for engineering design problems. Advances in Engineering Software 114: 163–191.
56. Mirjalili, Seyedali, Seyed Mohammad Mirjalili, and Abdolreza Hatamlou. 2016. Multi-verse
optimizer: A nature-inspired algorithm for global optimization. Neural Computing and Appli-
cations 27 (2): 495–513.
57. Niknam, Taher, and Babak Amiri. 2010. An efficient hybrid approach based on PSO, ACO and
k-means for cluster analysis. Applied Soft Computing 10 (1): 183–197.
58. Ozturk, Celal, Emrah Hancer, and Dervis Karaboga. 2015. Dynamic clustering with improved
binary artificial bee colony algorithm. Applied Soft Computing 28: 69–80.
59. Pan, Wen-Tsao. 2012. A new fruit fly optimization algorithm: Taking the financial distress
model as an example. Knowledge-Based Systems 26: 69–74.
60. Park, Hae-Sang, and Chi-Hyuck Jun. 2009. A simple and fast algorithm for k-medoids clus-
tering. Expert Systems with Applications 36 (2): 3336–3341.
61. Patel, Monika Raghuvanshi Rahul. 2017. An improved document clustering with multiview
point similarity/dissimilarity measures. International Journal of Engineering and Computer
Science 6 (2).
62. Sahoo, G., et al. 2017. A two-step artificial bee colony algorithm for clustering. Neural Com-
puting and Applications 28 (3): 537–551.
63. Sayed, Gehad Ismail, Ashraf Darwish, and Aboul Ella Hassanien. 2017. Quantum multiverse
optimization algorithm for optimization problems. Neural Computing and Applications 1–18.
64. Sayed, Gehad Ismail, Ashraf Darwish, and Aboul Ella Hassanien. 2018. A new chaotic multi-
verse optimization algorithm for solving engineering optimization problems. Journal of Exper-
imental & Theoretical Artificial Intelligence 30 (2): 293–317.
65. Shahnaz, Farial, Michael W. Berry, V. Paul Pauca, and Robert J. Plemmons. 2006. Document
clustering using nonnegative matrix factorization. Information Processing & Management 42
(2): 373–386.
66. Shelokar, P.S., Valadi K. Jayaraman, and Bhaskar D. Kulkarni. 2004. An ant colony approach
for clustering. Analytica Chimica Acta 509 (2): 187–195.
A Hybrid Salp Swarm Algorithm with β-Hill Climbing Algorithm … 161

67. Wei, Tingting, Lu Yonghe, Huiyou Chang, Qiang Zhou, and Xianyu Bao. 2015. A semantic
approach for text clustering using wordnet and lexical chains. Expert Systems with Applications
42 (4): 2264–2275.
68. Zaw, Moe Moe, and Ei Ei Mon. 2015. Web document clustering by using PSO-based cuckoo
search clustering algorithm. In Recent advances in swarm intelligence and evolutionary com-
putation, 263–281. Springer.
69. Zhao, Ying, and George Karypis. 2001. Criterion functions for document clustering: Experi-
ments and analysis.
Controlling Population Diversity of
Harris Hawks Optimization Algorithm
Using Self-adaptive Clustering Approach

Hamza Turabieh and Majdi Mafarja

Abstract Harris Hawks optimization (HHO) algorithm is a new adaptive search

method that can be employed as approximation heuristic for complex optimization
problems. HHO algorithm processes a population of search space with two opera-
tions: Soft besiege and Hard besiege. One of main problems in the use of population-
based algorithms is premature convergence. A premature stagnation of the search
creates a shortage of diversity, which affects the relationship between exploration and
exploitation processes. Here, we propose a self-adaptive clustering approach based
on k-means and density for all solutions in the population pool. The k-means clus-
tering method will examine the distribution (i.e., locations) of the solutions over the
search space. Once most of solutions (more than 90%) are related to one cluster, an
exploitation process will gain a higher priority compared to exploration one. Redis-
tributing the solutions over several clusters by a redistribution process will keep the
ratio between exploration and exploitation stable over all iterations. The proposed
approach will be tested over several mathematical test functions. The obtained results
show that the proposed modification can enhance the performance of HHO algorithm
and prevent premature convergence.

Keywords HHO · Harris Hawks Optimization · Data clustering · Optimization ·

Evolutionary computation · Nature-inspired algorithms · Swarm intelligence ·
Meta-heuristics · Unsupervised learning

H. Turabieh
Information Technology Department, CIT College, Taif University, Taif, Saudi Arabia
e-mail: [email protected]
M. Mafarja (B)
Department of Computer Science, Birzeit University, Birzeit, West Bank, Palestine
e-mail: [email protected]

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer 163
Nature Singapore Pte Ltd. 2021
I. Aljarah et al. (eds.), Evolutionary Data Clustering: Algorithms and Applications, Algo-
rithms for Intelligent Systems,
https://doi.org/10.1007/978-981-33-4191-3_7
164 H. Turabieh and M. Mafarja

1 Introduction

Population-based algorithms usually have two criticisms while optimizing complex

problems which are trapped in local optima, and computationally inefficient while
searching without contributing to the best solution found [3]. Low population diver-
sity is a major factor for premature convergence of population-based algorithms [5,
11]. Finding a certain level of population diversity is an acceptable solution that will
help curb premature convergence. In general, every optimization algorithm should
address two criteria: exploration and exploitation of a search space. Exploration refers
to visit new areas of the search space, while exploitation refers to visit those areas
within neighborhood points [5, 20]. As a result, the population-based search algo-
rithm should find a balanced ratio between exploration and exploitation to prevent
the premature convergence state and low population diversity.
In general, the overall behavior of population-based algorithms such as genetic
algorithms (GAs) [6], ant colony optimization (ACO) [8], evolutionary strategies
(ESs) [2], evolutionary programming (EP) [9], particle swarm optimization (PSO)
[14],brain storm optimization (BSO) [17], Harris Hawks optimization (HHO) [10],
to name recent will-known algorithms, is determined by the relationship between
exploitation and exploration throughout the run [16]. Several researchers believe
that population-based algorithms are effective due to their excellent ratio between
both operations (i.e., exploration and exploitation).
Up to date, researchers are still working to enhance the performance of population-
based algorithms. Recent development and enhancements can be found in [4, 12, 21].
Kelly et al. [11] proposed a novel selection method called knobelty, to dynamic con-
trol the exploration and exploitation for GP algorithm by generating mixed population
of parents. The proposed approach overcomes the low population diversity. Asiain
et al. [1] proposed a controller exploitation-exploration (CEE) with reinforcement
learning (RL) approach. The proposed CEE consists of three parts: controller, fast-
tracked learning, and the actor-critic. The proposed CEE is able to find a good ratio
between exploitation and exploration and prevent premature convergence. Nguyen
et al. [13] proposed a dynamic sticky binary PSO, which controls the PSO parame-
ters to improve the exploration and exploitation processes for feature selection and
knapsack problems. Zhang et al. [24] highlighted the importance of exploitation and
exploration for evolutionary algorithms with a different adaptive variation for local
fine-tuning while searching for global optima. Interested readers about recently pub-
lished papers related to exploration and exploitation for population-based algorithms
can read the following cited papers [15, 18, 19, 23].
In 2019, Asghar et al. [10] proposed HHO algorithm, that mimics the Harris Hawks
hunting nature. This chapter presents a self-adaptive learning method based on k-
means clustering method to control the population diversity of HHO algorithm. The
motivation of this chapter is to prevent the premature convergence of HHO once the
majority of solutions belong to one cluster throughout the run and no improvements
occur on the best so far solution. The proposed approach will redistribute half of
the current population to enhance the ratio for exploration process and discover new
areas.
Controlling Population Diversity of Harris Hawks Optimization … 165

The rest of this chapter is organized as follows: Sect. 2 presents the proposed
enhanced of HHO algorithm by hybridizing it with k-means method. Section 3
presents the obtained results and comparison with the original HHO algorithm.
Section 4 presents the conclusion and future works.

2 Proposed Approach

A self-adaptive controlling method to control the population diversity of Harris

Hawks Optimization Algorithm is proposed in this chapter. In the beginning, HHO
generates randomly a set of initial solutions (Rabbit positions) distributed on the
search space as shown in Fig. 1. The original HHO algorithm comes with two main
phases for exploration and exploitation processes. The exploration process that sim-
ulates the nature of Harris’ hawks tracks the expected prey (solution) with their
powerful eyes. Harris’ hawks use two strategies to perch a prey. The first strategy
depends on the location of other Harris’ hawks and the expected prey location. While
the second strategy depends on monitoring random locations in their eyes range, both
strategies have an equal chance. Equation (1) demonstrates the exploration process
for HHO algorithm.


X rand (t) − r1 |X rand (t) − 2r2 X (t)| q ≥ 0.5
X (t + 1) = (1)
(X rabbit (t) − X m (t)) − r3 (L B + r4 (U B − L B)) q < 0.5

where X refers to a position vector, the variable X (t + 1) presents the hawks position
in the following iteration t, while X rabbit (t) presents the location of the prey (i.e.,
rabbit) in the current iteration t. X rand (t) is a hawk that is selected randomly from the
current population. X m is the average position of the current population of hawks. r1 ,
r2 , r3 , r4 , and q are random numbers between 0 and 1, which are generated randomly
at each iteration. L B and U B are the upper and lower bounds of variables. The
average position for all hawks is calculated based on Eq. (2), where X i (t) presents
the location of each hawk in iteration t, and N refers to the total number of hawks
in the population.

1 
N
X m (t) = X i (t) (2)
N i=1

One of the main advantages of HHO is the switching process between exploration
to exploitation. This switching process depends on the escaping energy of the prey.
In general, the energy decreases in escaping behavior. Equation (3) presents the
decreasing value of the prey energy, where E is the escaping energy of the prey, T
denotes the maximum number of iterations, and E 0 is a random initial energy of
the prey within the range (−1, 1) that is generated at each iteration. The switching
166 H. Turabieh and M. Mafarja

Fig. 1 Initial distribution of solutions over the search space

process between exploration to exploitation is controlled by the value of E. The

exploration process is executed when |E| ≥1, while exploitation process is executed
when |E| <1.
 
t
E = 2E 0 1 − (3)
T

The exploitation process occurs based on the prey energy status E. The exploration
process inside HHO has two strategies called soft and hard besiege processes. Let
a variable r presents the chance of rabbit in successfully escaping (r < 0.5) or not
successfully escaping (r ≥ 0.5) before surprise pounce. The soft besiege occurs when
r ≥ 0.5 and |E| ≥ 0.5. Since |E| ≥ 0.5, this indicates that the prey still have a good
enough energy to escape. The hawk keeps flying until the energy of rabbit (i.e., prey)
starts decrease. Equation (4) demonstrates the soft besiege process, where ΔX (t)
refers to the difference between the current location in the search space and the
position vector at iteration t. J = 2(1 − r5 ) indicates a simulation of random jump
of rabbit while escaping process, where r5 is a random value between 0 and 1.

X (t + 1) = ΔX (t) − E |J X rabbit (t) − X (t)| (4)

ΔX (t) = X rabbit (t) − X (t) (5)

Equation (6) presents the hard besiege process, when both |E| and r <0.5. The
hawks reduce the average positions distance with the escaping prey where both
variables Y and Z are obtained based on Eqs. (8) and (9), where X m (t) is achieved
based on Eq. (2). The hawks update their position while executing soft besiege
process is based on Eq. (6), where both variables Y and Z are evaluated based on two
Equations that are Eqs. (8) and (9)), respectively. Algorithm 1 shows the pseudocode
of the original HHO algorithm.
Controlling Population Diversity of Harris Hawks Optimization … 167


Y i f F(Y ) < F(X (t))
X (t + 1) = (6)
Z i f F(Z ) < F(X (t))

X (t + 1) = X rabbit (t) − E |ΔX (t)| (7)

Y = X rabbit (t) − E |J X rabbit (t) − X (t)| (8)

Z = Y + S × L F(D) (9)

Algorithm 1 The original HHO.

Defines:
Set population size = N
Set number of iterations T
Set Iter =1
Output: Best solution (i.e., Rabbit location)
Generate population randomly X i (1, 2, . . . , N )
while (Iter < T ) do
Calculate the fitness values for each hawk
Set X rabbit as the location of rabbit (best location)
for (each hawk (X i )) do
Update the initial energy E 0
Update the jump strength J
Update the value of E using Eq. (3)
if (|E| ≥ 1) then
Update the location vector using Eq. (1)
if (|E| < 1) then
if (r ≥0.5 and |E| ≥ 0.5 ) then
Update the location vector using Eq. (4)
else if (r ≥0.5 and |E| < 0.5 ) then
Update the location vector using Eq. (7)
else if (r <0.5 and |E| ≥ 0.5 ) then
Update the location vector using Eq. (6)
else if (r <0.5 and |E| < 0.5 ) then
Update the location vector using Eq. (6)
Iter=Iter+1
Return X rabbit

In general, HHO and other population-based algorithms may suffer from prema-
ture convergence after a set of iterations. Figure 2 demonstrates a pictorial presen-
tation of premature convergence case, where all solutions may have high similarity
(i.e., cluster) and stuck in local optima. To overcome this problem, an enhanced ver-
sion of HHO has been proposed by hybridizing it with k-means clustering approach.
The proposed idea will check the distribution (i.e., location) of the solutions (preys)
168 H. Turabieh and M. Mafarja

Fig. 2 Premature convergence case

Fig. 3 Redistribution of the solutions after numbers of iterations

over the search space. If the majority of preys at certain iterations starts belong
to one cluster and no improvement occurs on the best solution for M iterations, a
redistribution process will generate random solutions and update the current popula-
tion. Figure 3 shows the redistribution process for solutions. This process will keep
the diversity of the population at higher levels and prevent premature convergence.
Figure 2 demonstrates the modified HHO algorithm.
The proposed approach will distribute half of the solutions in the current popula-
tion randomly over the search space. This mechanism will keep population diversity
high, which enhance the exploration process and prevent the premature convergence.
Adding the k-means clustering method will increase the execution time for the orig-
inal HHO algorithm. However, it will enhance the overall performance of HHO
algorithm.
Controlling Population Diversity of Harris Hawks Optimization … 169

Algorithm 2 The modified HHO.

Defines:
Set population size = N
Set number of iterations = T
Set Iter =1
Output: Best solution (i.e., Rabbit location)
Generate population randomly X i (1, 2, . . . , N )
while (Iter < T ) do
Calculate the fitness values for each hawk
Define X rabbit as the location best location.
for (each hawk (X i )) do
Update the initial energy E 0 and jump strength J
Update the energy (E) based on Eq. (3)
if (|E| ≥ 1) then
Update the location vector based on Eq. (1)
if (|E| < 1) then
if (r ≥0.5 and |E| ≥ 0.5 ) then
Update the location vector based on Eq. (4)
else if (r ≥0.5 and |E| < 0.5 ) then
Update the location vector based on Eq. (7)
else if (r <0.5 and |E| ≥ 0.5 ) then
Update the location vector based on Eq. (6)
else if (r <0.5 and |E| < 0.5 ) then
Update the location vector based on Eq. (6)
Cluster= k-means(X rabbit ,2)  only two clusters will be used
if (the majority of X rabbit belong to one cluster and no improvement for M iterations) then
generate N /2 new random positions X rabbit .
Update the current population.
Iter=Iter+1
Return X rabbit

3 Results and Analysis

To evaluate the enhanced HHO, we employed the modified HHO on different exam-
ples of mathematical test functions with known global optima. We employed 13
mathematics benchmark functions and compared the obtained results with the orig-
inal HHO [7, 10, 22]. These mathematical test functions are categorized into three
sets: unimodal (UM) test functions, multimodal (MM) test functions, and compo-
sition (CM) test functions. Functions with a single global optimum from (F1–F7)
are UM, while functions with multi global optimum from (F8–F13). Tables 1 and
2 demonstrate each mathematical functions for two categories (i.e., UM and MM).
Each mathematical function was simulated for 21 independent runs with dimension
30 and 100. All experiments are evaluated using Matlab 2019b, 2.81 GHz, and 16
GB RAM. Table 3 presents the parameters setting used in this chapter.
Table 4 demonstrates the obtained average and standard deviation for each func-
tion. In all experiments, we employed 500 iterations for original and modified HHO
algorithm. It is clear that for UM functions, the proposed modified approach outper-
170 H. Turabieh and M. Mafarja

Table 1 Details of unimodal (UM) benchmark test functions

Test function Dimensions Range f min
n
f 1 (x) = i=1 xi2 30 [−100, 100] 0
n n
f 2 (x) = i=1 |xi |+ i=1 |xi | 30 [−10, 10] 0
n i 2
f 3 (x) = i=1 j−1 x j 30 [−100, 100] 0
f 4 (x) = maxi {|xi |, 1 ≤ i ≤ n} 30 [−100, 100] 0
n−1 2
f 5 (x) = i=1 100 xi+1 − xi2 + (xi − 1)2 30 [−30, 30] 0
n
f 6 (x) = i=1 ([xi + 0.5])2 30 [−100, 100] 0
n
f 7 (x) = i=1 i xi4 + random[0, 1) 30 [−128, 128] 0

form the original HHO in five functions based on the average (i.e., f 1 , f 2 , f 3 , f 4 ,
and f 7 ) under the same settings. While for MM functions, the proposed approach
outperform in one function (i.e., f 8 ) and gain same results in 3 mathematical func-
tions (i.e., f 9 , f 10 , and f 11 ). Moreover, the proposed modification can enhance the
overall performance of the original HHO and prevent the premature convergence.
Table 5 shows the number of calling time for population redistribution method for
the modified HHO. It is clear that the original HHO has a fast premature convergence
since all solutions (preys) belong to one cluster at the early iterations. Moreover, the
original HHO exploitation ratio is more higher than exploration one.
Figure 4 shows the convergence performance of f 3 for the original and modified
HHO algorithm with Dim equals 30. It is obvious that the original HHO has a fast
convergence compared to the modified one. From Table 5, the average redistribution
function called 372.95 times out of 500 iterations. Moreover, the reported results
for average value in Table 4 for f 3 show that the modified HHO can outperform the
original one.
Figure 5 shows the performance of f 10 with Dim equals 100. It is clear that the
original HHO has a fast convergence, while the modified approach tries to explore
more search space areas by updating the population once no improvement occurs.
Moreover, from Table 5, the updating function is called 432 times out of 500, it is
clear that the original HHO has a high exploitation process compared to exploration
one.
In general, HHO algorithm exploitation ratio is higher than exploration one. Form
Table 5 more than 70% of the total iterations, the redistribution method is called.
A premature convergence occurs once most of the solutions belong to one cluster
and HHO can not enhance the current best solution. The reported results show that
controlling the population diversity and having a good ratio between exploration and
exploitation process will reduce the probability of premature convergence of HHO
algorithm.
Table 2 Details of multimodal (MM) benchmark test functions
Test function Dimensions Range f min
n √
f 8 (x) = i=1 −xi sin |xi | 30 [−500, 500] −418.9829 ×
n
n  
f 9 (x) = i=1 xi2 − 10 cos (2π xi ) + 10 30 [−5.12, 5.12] 0
 
n n
f 10 (x) = −20 exp(−0.2 n1 i=1 xi2 ) − exp n1 i=1 cos (2π xi ) + 20 + e 30 [−32, 32] 0
n 
1 n 2 x
f 11 (x) = 4000 i=1 xi − i=1 cos √i + 1 30 [−600, 600] 0
 i 
n−1  
f 12 (x) = πn 10 sin (π y1 ) + i=1 (yi − 1)2 1 + 10 sin2 (π yi+1 ) + (yn − 1)2 30 [−50, 50] 0
n
+ i=1 u(xi , 10, 100, 4) ⎧
⎪ m xi > a
⎨k(xi − a)
xi +1
yi = 1 + 4 u(xi , a, k, m) = 0 − a < xi < a
⎪
⎩
k(−xi − a)m xi < −a
Controlling Population Diversity of Harris Hawks Optimization …

 2 n    
f 13 (x) = 0.1 sin (3π x1 ) + i=1 (xi − 1)2 1 + sin2 (3π xi + 1) + (xn − 1)2 1 + sin2 (2π xn ) +
 30 [−50, 50] 0
n
i=1 u(x i , 5, 100, 4)
171
172 H. Turabieh and M. Mafarja

Table 3 Parameters setting for original and modified HHO

Parameter Value
Number of iterations (T ) 500
Population size (N ) 50

Table 4 Comparison between original and modified HHO algorithm

Original HHO Modified HHO Original HHO Modified HHO
Dim 30 Dim 30 Dim 100 Dim 100
Avg. Std. Avg. Std. Avg. Std. Avg. Std.
f1 2.93E−95 1.22E−94 0.00E+00 0.00E+00 4.02E−94 1.82E−93 0.00E+00 0.00E+00
f2 9.11E−52 3.09E−51 4.73E−94 2.13E−93 3.94E−50 1.79E−49 6.64E−95 2.51E−94
f3 1.41E−78 6.44E−78 0.00E+00 0.00E+00 2.19E−66 9.79E−66 0.00E+00 0.00E+00
f4 3.20E−49 7.47E−49 3.75E−94 1.14E−93 5.20E−49 1.44E−48 1.95E−92 5.01E−92
f5 0.006925 0.007514 0.195261 0.294316 0.03484 0.044785 0.003743 0.04682
f6 0.000172 0.000303 0.000679 0.000859 0.000457 0.000577 0.016174 0.015386
f7 0.000213 0.000336 9.07E−05 7.85E−05 0.000213 0.000178 1.42E−04 1.48E−04
f8 −12569 0.687329 −12568.4 1.09969 −41851.6 205.618 −41889.4 12.77664
f9 0 0 0 0 0 0 0 0
f 10 8.88E−16 0 8.88E−16 0 8.88E−16 0 8.88E−16 0
f 11 0 0 0 0 0 0 0 0
f 12 6.98E−06 1.07E−05 7.85E−05 7.14E−05 4.86E−06 9.50E−06 1.80E−04 1.40E−04
f 13 0.000128 0.000206 0.000662 0.000593 0.000146 0.000213 0.013219 0.020858

Table 5 Number of calling times for updating method

Dim 30 Dim100
Avg. call Avg. call
f1 368.57 436
f2 373.19 448
f3 372.95 457
f4 371.33 426
f5 372.67 433
f6 374.43 437
f7 369.76 447
f8 370.67 429
f9 371.81 444
f 10 369.24 432
f 11 371.33 441
f 12 371.19 436
f 13 373.05 430
Controlling Population Diversity of Harris Hawks Optimization … 173

105
2
Original HHO
1.8 Modified HHO

1.6

1.4
Fitness Value

1.2

0.8

0.6

0.4

0.2

0
5 10 15 20 25 30
Iterations

Fig. 4 Convergence diagram for f 3 with Dim = 30

25
Modified HHO
Original HHO

20
Fitness Value

15

10

0
5 10 15 20 25 30 35 40 45 50
Iterations

Fig. 5 Convergence diagram for f 10 with Dim = 100

174 H. Turabieh and M. Mafarja

4 Conclusion and Future Works

In this chapter, we introduced a modified version of HHO algorithm by hybridizing it

with k-means method. The k-means is employed to examine the population diversity
of the population. Once the majority of solutions belong to one cluster and HHO
can find any improvement for predetermined iterations (i.e., premature convergence
case), we update the current population by replacing half of the current population
with new solutions randomly. This updating process will redistribute the solutions
over the search space, which will increase the exploration process by discovering
new locations for the HHO. The obtained results show that the proposed redistribu-
tion approach has been called more than 70% of the overall number of iterations.
Our future works will examine the proposed approach on several real NP-hard opti-
mization problems.

References

1. Asiain, Erick, Julio B. Clempner, and Alexander S. Poznyak. 2019. Controller exploitation-
exploration reinforcement learning architecture for computing near-optimal policies. Soft Com-
puting 23 (11): 3591–3604.
2. Bäck, Thomas. 1996. Evolutionary algorithms in theory and practice: Evolution strategies,
evolutionary programming, genetic algorithms. New York, NY, USA: Oxford University Press,
Inc.
3. Bongard, Josh C., and Gregory S. Hornby. 2010. Guarding against premature convergence while
accelerating evolutionary search. In Proceedings of the 12th annual conference on genetic and
evolutionary computation, GECCO ’10, 111–118. New York, NY, USA: ACM.
4. Bujok, Petr, Josef Tvrdk, and Radka Polkovi. 2019. Comparison of nature-inspired population-
based algorithms on continuous optimisation problems. Swarm and Evolutionary Computation
50: 100490.
5. Črepinšek, Matej, Shih-Hsi Liu, and Marjan Mernik. 2013. Exploration and exploitation in
evolutionary algorithms: A survey. ACM Computing Surveys 45 (3): 35:1–35:33
6. DeJong, Kenneth A., and Kenneth A. De Jong. 2002. Evolutionary computation. The MIT
Press.
7. Digalakis, J.G., and K.G. Margaritis. 2001. On benchmarking functions for genetic algorithms.
International Journal of Computer Mathematics 77 (4): 481–506.
8. Dorigo, Marco, and Gianni Di Caro. 1999. Ant colony optimization: A new meta-heuristic. In
Proceedings of the 1999 congress on evolutionary computation-CEC99 (Cat. No. 99TH8406),
vol. 2, 1470–1477. IEEE.
9. Fogel, Lawrence J. 1999. Intelligence through simulated evolution: Forty years of evolutionary
programming. New York, NY, USA: Wiley, Inc.
10. Heidari, Ali Asghar, Seyedali Mirjalili, Hossam Faris, Ibrahim Aljarah, Majdi Mafarja, and
Huiling Chen. 2019. Harris hawks optimization: Algorithm and applications. Future Genera-
tion Computer Systems 97: 849–872.
11. Kelly, Jonathan, Erik Hemberg, and Una-May O’Reilly. 2019. Improving genetic programming
with novel exploration—Exploitation control. In Genetic programming, ed. Lukas Sekanina,
Hu Ting, Nuno Lourenço, Hendrik Richter, and Pablo García-Sánchez, 64–80. Cham: Springer
International Publishing.
12. Mahdavi, Sedigheh, Shahryar Rahnamayan, and Abbas Mahdavi. 2019. Majority voting for
discrete population-based optimization algorithms. Soft Computing 23 (1): 1–18.
Controlling Population Diversity of Harris Hawks Optimization … 175

13. Nguyen, B.H., B. Xue, P. Andreae, and M. Zhang. 2019. A new binary particle swarm opti-
mization approach: Momentum and dynamic balance between exploration and exploitation.
IEEE Transactions on Cybernetics 1–15.
14. Olorunda, O., and Engelbrecht, A.P. 2008. Measuring exploration/exploitation in particle
swarms using swarm diversity. In 2008 IEEE congress on evolutionary computation (IEEE
world congress on computational intelligence), 1128–1134, June 2008.
15. Rezapoor Mirsaleh, Mehdi, and Mohammad Reza Meybodi. 2018. Balancing exploration and
exploitation in memetic algorithms: A learning automata approach. Computational Intelligence
34 (1): 282–309.
16. Sharma, Harish, Jagdish Chand Bansal, and K.V. Arya. 2014. Self balanced differential evolu-
tion. Journal of Computational Science 5 (2): 312–323 (Empowering Science through Com-
puting + BioInspired Computing).
17. Shi, Yuhui. 2011. Brain storm optimization algorithm. In International conference in swarm
intelligence, 303–309. Springer.
18. Shojaedini, Ehsan, Mahshid Majd, and Reza Safabakhsh. 2019. Novel adaptive genetic algo-
rithm sample consensus. Applied Soft Computing 77: 635–642.
19. Stanovov, Vladimir, Shakhnaz Akhmedova, and Eugene Semenkin. 2019. Selective pressure
strategy in differential evolution: Exploitation improvement in solving global optimization
problems. Swarm and Evolutionary Computation 50: 100463.
20. Tan, K.C., S.C. Chiam, A.A. Mamun, and C.K. Goh. 2009. Balancing exploration and exploita-
tion with adaptive variation for evolutionary multi-objective optimization. European Journal
of Operational Research 197 (2): 701–713.
21. Wu, Guohua, Rammohan Mallipeddi, and Ponnuthurai Nagaratnam Suganthan. 2019. Ensem-
ble strategies for population-based optimization algorithms—A survey. Swarm and Evolution-
ary Computation 44: 695–711.
22. Yao, Xin, Yong Liu, and Guangming Lin. 1999. Evolutionary programming made faster. IEEE
Transactions on Evolutionary Computation 3 (2): 82–102.
23. Yelghi, Aref, and Cemal Kse. 2018. A modified firefly algorithm for global minimum opti-
mization. Applied Soft Computing 62: 29–44.
24. Zhang, Hu, Jianyong Sun, Tonglin Liu, Ke Zhang, and Qingfu Zhang. 2019. Balancing explo-
ration and exploitation in multiobjective evolutionary optimization. Information Sciences 497:
129–148.
A Review of Multiobjective Evolutionary
Algorithms for Data Clustering Problems

Ruba Abu Khurma and Ibrahim Aljarah

Abstract Clustering is a challenging machine learning problem that deals with

unlabeled data instances. It divides them into homogeneous groups according to
similarity/dissimilarity between these instances. This process is called unsupervised
learning. Detecting the data structure is not an easy task, especially if there are no
previous assumptions to guide the clustering process. Multi-objective evolutionary
algorithms have been used as alternative solutions in place of traditional cluster-
ing and single-objective evolutionary methods. By using multi-objective evolution-
ary algorithms, multiple objectives can be optimized simultaneously. This chapter
introduces the clustering problem, multi-objective optimization, and reviews sev-
eral multi-objective clustering techniques. It focuses on five subjects: chromosome
encoding methods, fitness functions for optimization, evolutionary operators, storing
current non-dominated solutions, and selection of the final solution from the non-
dominated front. Finally, various applications of multi-objective clustering algo-
rithms are discussed in the areas of image processing, bioinformatics and social
network, etc. The chapter concludes by highlighting some of the open issues that
could be investigated in the future.

Keywords Data clustering · Multi-objective optimization · Evolutionary

computation · Nature-inspired algorithms · Swarm intelligence · Meta-heuristics
unsupervised learning · Pareto optimality

R. Abu Khurma · I. Aljarah (B)

King Abdullah II School for Information Technology, The University of Jordan, Amman, Jordan
e-mail: [email protected]
R. Abu Khurma
e-mail: [email protected]

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer 177
Nature Singapore Pte Ltd. 2021
I. Aljarah et al. (eds.), Evolutionary Data Clustering: Algorithms and Applications, Algo-
rithms for Intelligent Systems,
https://doi.org/10.1007/978-981-33-4191-3_8
178 R. Abu Khurma and I. Aljarah

1 Introduction

Recently, the amount of data stored in data repositories such as databases, data
warehouses, and metadata has increased the need for specialized data analysis tools
[28]. One of the new technologies is unsupervised techniques. The merit of these
methods is that they are data driven. This means that the learning process is not guided
by labeled training data. However, unsupervised methods are inductive methods
that, unlike supervised methods, do not rely on examples of training. Therefore,
unsupervised learning is based on the distinct structure of the data. It uses some
criteria to detect the hidden pattern of data points in the data space [35].
Clustering is a popular unsupervised method of learning from a massive amount of
unlabeled data items. It has drawn the attention of researchers in the fields of bioin-
formatics, medicine, and pattern recognition. The clustering principle is to obtain
homogeneity within clusters by grouping similar data items in the same cluster. Fur-
thermore, it posses heterogeneity between clusters by distributing dissimilar data
items into different clusters [5–7].
Some aspects considered challenges for the clustering process [67]. First, the same
dataset may lead to different clustering solutions. This is because each clustering
algorithm has its method of grouping data elements and each one may use different
properties of these elements to determine their corresponding clusters. Thus, it is very
difficult to obtain stable clustering results for the dataset using different clustering
algorithms. For example, the K-means algorithm and graph-based partitioning algo-
rithm formulate the clusters and determines the elements membership using different
methods.
The second challenge for clustering algorithms is to optimize the cluster param-
eters like the number of clusters and cluster centers [40]. The clustering algorithm
may fix this dilemma after spending a long time determining these numbers but still
have a chance of failing because it is considering a single evaluation function to
estimate the quality of a cluster such as the spatial separation, compactness, and
connectedness.
The final challenge is that the traditional clustering algorithms such as K-means
have a local minima [37]. This is due to the high sensitivity towards the initially
selected centroids. The other major problem is that these methods apply a greedy
search to optimize a single criterion. In the case of K-means, the algorithm tries to
achieve the largest similarity between data instances in the cluster. This objective is
called cluster compactness (intra-cluster cohesion).
In the literature, various evolutionary algorithms such as Genetic Algorithms (GAs)
have been used as alternative solutions to address the problem of local minima [64].
Although these methods can mitigate the local minima problem and approach near
global best, they are unable to deal with the different characteristics of the dataset
simultaneously [1]. This is because the single-objective evolutionary algorithm has
a fitness function that tries to optimize a single validity index [2]. Therefore, there has
been a growing trend towards using multi-objective algorithms to simultaneously pro-
cess different data characteristics. Considering different characteristics of the dataset
will bring more benefits to the process of clustering optimization [9, 10, 58, 66].
A Review of Multiobjective Evolutionary Algorithms … 179

The Multi-Objective Optimization Problem (MOOP) depends on a number of

conflicting objectives. At least, two competing criteria must be optimized. Unlike
scalar-objective optimization, the MOOP does not have a single solution to make
optimization for all the evaluation functions simultaneously. However, there will be
trade-off solutions generated from the set of all possible solutions [45].
Evolutionary Algorithms (EAs) have a major role in tackling the MOOPs and
have been proposed in the literature to address various MOOPs in many applications.
Due to the promising results achieved by Multi-Objective Evolutionary Algorithms
(MOEAs), they are useful for use in the clustering problem [45]. Dealing with mul-
tiple objectives simultaneously may help consider different data properties that may
improve the robustness of the clustering method and enhance clustering results.
In general, the MOEAs have two different methods for multi-objective optimiza-
tion: the first is to convert the multi-objective problem into a single-objective problem
based on a formula that gives a rank to each objective based on its importance. The
second is the Pareto-based method of generating a group of non-dominated solutions
by simultaneous optimization of the objectives. The obtained solutions are trade-off
solutions that can’t be further optimized based on one objective without adversely
affecting other competing objectives.
In the clustering problem, there is no advance information about available objec-
tives so that their relative importance can’t be detected in advance [3]. Thus, the first
method that relies on the order of objectives in a single-objective function based on
their weights, then improving this function may degrade the clustering results [4, 7,
24, 25, 59].
The second multi-objective optimization is more convenient for clustering because
it optimizes objectives simultaneously instead of merging them into a single-objective
function. The obtained trade-off solutions can then be compared based on user needs
and problems at hand.
This chapter reviews the subject of evolutionary multi-objective clustering. Due to
the low coverage of this research area, we have directed our efforts to systematically
study many of the previous works in this area. The chapter discusses the cluster-
ing problem, multi-objective optimization, and poses a framework for evolutionary
multi-objective clustering. Besides, a detailed discussion is provided about the differ-
ent multi-objective clustering methods and finally, the applications of multi-objective
clustering are described.
The chapter is structured as follows: Sect. 2 presents the clustering problem.
Section 3 identifies clustering as an optimization problem. Section 4 explains the con-
cepts of multi-objective optimization. Section 5 discusses the importance of multi-
objective clustering. Section 6 poses a framework of evolutionary multiple objectives
clustering. Section 7 discusses various multi-objective clustering techniques. Section
8 discusses the applications of multi-objective evolutionary clustering. Finally, Sect.
9 contains the conclusion of the chapter and paves the way for future studies.
180 R. Abu Khurma and I. Aljarah

2 Clustering Problem

Clustering is a main machine learning process to determine the hidden structure of

a group of unlabeled data instances [8]. The major challenge of clustering is that the
learning process is unsupervised because there is no knowledge available about the
structural characteristics of the dataset such as distribution, volume, shapes, etc. A
simple definition for Clustering is that it is a process of finding the intrinsic grouping
of a set of instances and distribute them into homogeneous sets based on certain
grouping criteria. Accordingly, there are two clustering methods: distance clustering
and conceptual clustering. The former one uses the similarity measure (distance)
to collect close instances in the same cluster. The conceptual clustering collects
instances in clusters based on a common descriptive concept.
In general, no best criterion can be determined in an absolute way. The goal of
clustering is user-dependent which means that the user can define the clustering
criterion based on his/her needs. For example, the user may be interested in finding
similar behavior or unusual behavior (outliers detection) in a customer database.
Formally speaking, clustering is an unsupervised learning method for dividing
space into K areas {C1 , C2 , . . . , C K } based on some criteria. K is the number of
clusters that maybe not given in advance. Given a set of data points {X 1 , X 2 , . . . , X n }
that are distributed in a multi-dimensional space, the purpose of clustering is to
produce a partition matrix of size K × n. The partition matrix is named by U = [u i j ]
where i ∈ [1, K ] and j ∈ [1, n]. The element of the partition matrix denoted by u i j
means that an instance x j is a member of the cluster Ci . Based on this definition, there
two types of clustering: partitional and overlapping [36]. In partitional clustering,
there are hard partition of objects into mutually disjoint groups so that Cr 1 ∩ Cr 2 = φ
where r 1 = r 2 and C1 ∪ C2 ∪ ... ∪ C K = X where Ci = φ. In other words, hard
partition clustering has a condition that u i j = 1 iff x j ∈ Ci and u i j = 0 iff x j ∈
/ Ci.
The other clustering method is the overlapping clustering. It relaxes the partitional
clustering constraints. Hence, the data instance may belong entirely to one or more
clusters such as in soft clustering or may belong to different clusters with different
degrees such as in fuzzy clustering. Thus, the condition in fuzzy clustering is 0 ≤
u i j ≤ 1.
Figure 1 illustrates the results of clustering for some clustering algorithms on
toy datasets. The used clustering algorithms are MiniBatchKMeans, AffinityPropa-
gation, MeanShift, SpectralClustering, Ward, AgglomerativeClustering, DBSCAN,
OPTICS’, Birch and GaussianMixture. The used datasets are noisy-circles, noisy-
moons, varied, aniso and no-structure. This example illustrates the diverse behavior
of different clustering algorithms and gives some intuition on their performance
across 2D datasets. The distribution of points shows that in most cases the clustering
results were good. On the other hand, the last row in Fig. 1 shows that the dataset is
fairly homogeneous so the clustering results were not good.
A Review of Multiobjective Evolutionary Algorithms … 181

Fig. 1 Clustering example

3 Formulating Clustering as an Optimization Problem

Clustering is formalized as an optimization problem because it searches to determine

the best grouping of data points in such a way the grouping criteria are optimized.
These grouping criteria are called validity indices. Consequently, the objective func-
tion of an optimization algorithm may rely on optimizing one cluster validity index
(i.e., single-objective algorithm) or a set of validity indices (i.e., multi-objective algo-
rithm). The search space of the clustering problem represents the set of all possible
groupings of data points. Thus, searching to find the best grouping of data points in
this large search space makes the clustering problem an NP-hard problem. Formally,
the clustering problem is formulated as an optimization problem [32].
Let X a finite set of data points, a distance measure δ(i, j) ∈ + 0 for i, j ∈ X ,
a positive integer K and a criterion function J (C, δ(., .)) on a K-partition C =
C1 , . . . , C K of X and measure δ(., .). An optimization clustering problem seeks to
find the partition of X into disjoint sets {C1 . . . C K } that maximizes the expression
J (C, δ(., .)). The number R(K , N ) that represents the possible solutions for the
division of a dataset of size N into K partitions is given by Eq. 1:
 
1 
K
K KN
R(K , N ) = (−1) K −i (i) N ≈ (1)
K ! i=1 i K!

The search space of the clustering problem increases in exponential manner even
if K number of partitions is fixed. This means that exhaustive search methods will
not be able to traverse all the possible solutions even for medium-sized problems.
182 R. Abu Khurma and I. Aljarah

7 7 7
6 6 6
5 5 5
4 4 4
3 3 3
2 2 2
1 1 1
0 0 0
−1 −1 −1
−2 −2 −2
−3 −3 −3
−4 −2 0 2 4 6 8 10 −4 −2 0 2 4 6 8 10 −4 −2 0 2 4 6 8 10

(a) GA clustering (b) PSO clustering (c) DE clustering

540 450 600
520 440
500 550
430
480
Best Cost

Best Cost

Best Cost
420 500
460
410
440
400 450
420
400 390
400
380 380
360 370 350
0 50 100 150 200 0 50 100 150 200 0 50 100 150 200
Iteration Iteration Iteration

(d) GA convergence (e) PSO convergence (f) DE convergence

Fig. 2 Evolutionary clustering

Figure 2 shows an example for evolutionary clustering. In this example, Genetic algo-
rithm (GA), Particle Swarm Optimization (PSO), and Differential Evolution (DE)
are used to optimize the clustering problem. Clusters generated using these evolu-
tionary algorithms differ in their convergence behavior. As seen that PSO achieved
the best cost then GA and finally DE.

4 Multi-objective Optimization Problem

The Multi-Objective Optimization Problem (MOPs) optimizes more than one objec-
tive. These objectives are usually conflicting which indicates that no solution can be
found to optimize all these objectives simultaneously [45]. Instead, trade-off solu-
tions called Pareto optimal solutions are found among different objectives which are
called Pareto Front (PF) in the objective space and Pareto Set (PS) in the solution
space. Many Pareto optimal solutions can be generated [69]. However, the task is to
find those solutions that converge closer to PF and distribute uniformly among them
[72]. Equation2 formally describes the multi-objective optimization problem.

min F(x) = ( f 1 (x), f 2 (x), . . . , f k (x))T

(2)
subject to x = (x1 , x2 , . . . , xn )T ∈ Ω

where F(x) consists of k objective functions and f k (x) represents the kth objective
function. x ∈ Ω is a decision variable vector, Ω is the decision space, and n is
the dimension number of the decision variable x. It is said that the decision vector
x A ∈ Ω dominate the decision vector x B ∈ Ω If and only if the Eq. 3 is satisfied.
A Review of Multiobjective Evolutionary Algorithms … 183

1.4

1.2

1
Objective

0.8

0.6
nd
2

0.4

0.2

0
0 0.2 0.4 0.6 0.8 1
1st Objective

Fig. 3 Near-optimal Pareto solutions generated by MOPSO

∀i = 1, 2, . . . , k, f i (x A ) f i(x B )
(3)
∧ ∃ j = 1, 2, . . . , k f j (x A ) ≺ f j (x B )

A vector x ∗ ∈ Ω is Pareto optimal if no vector x ∈ Ω can dominate it. In other

words, x ∗ is a non-dominated solution. The Pareto Optimal Set (PS) consists of all
non-dominated vectors. The set composed of all the Pareto optimal objective vectors
is called the Pareto Front (PF). The definition of PF can be written as in Eq. 4:

P F = F(x  ) = ( f 1 (x  ), f 2 (x  ), . . . , f k (x  ))T | x  ∈ P S. (4)

All the potential solutions for the MOP can be found in PS. Furthermore, the MOP
can be defined as a search for a set of non-dominated vectors that can approximate
the PF.
Figure 3 shows the Near-optimal Pareto solutions generated by a Multi-Objective
Particle Swarm Optimization Algorithm (MOPSO).

5 Importance of Multi-objective Clustering

As data dimensions increase and data items distribute sparsely in the data space, the
results of traditional clustering methods become weak. The main problem is that
considering a single criterion for clustering cannot fit all clusters in the entire feature
space. There may be some clusters of mysterious shapes hidden in sub-spaces of the
184 R. Abu Khurma and I. Aljarah

original feature space [27]. Thus, reliance on a single criterion may result in clusters
of similar shapes and have a close data density. Obviously, a good clustering method
should carefully choose the appropriate optimization criterion for clustering.
There are several valid clustering criteria for a dataset, many of which are con-
flicting proprieties. Furthermore, data distribution is usually unknown at the begin-
ning of the clustering process. For these reasons, the simultaneous optimization of
multiple complementary objectives will solve most of the clustering problems and
obtain more reliable solutions. A multi-objective clustering algorithm applies the
Pareto optimization which considers multiple clustering qualities simultaneously.
The algorithm focuses on optimizing a group of clustering objectives and creating a
Pareto front set consisting of a set of trade-off solutions.

6 Framework of Evolutionary Multiple Objectives

Clustering

To build a general evolutionary multiple objectives clustering framework, there is a

need to review some aspects of evolutionary single-objectives clustering. To design
an evolutionary single-objectives clustering, there are three aspects to consider [50]:
• The encoding scheme of the solution. This describes the representation of a clus-
tering solution using a chromosome.
• The selection of appropriate evolutionary operators to manage the update of solu-
tions at each evolutionary cycle.
• Choose the fitness function that guides the reproductive mechanism by assessing
the closeness of the candidate solution is to the optimal level.
In the case of evolutionary multi-objective clustering, there are some additional issues
to consider:

• First, there is more than one fitness function that must be optimized simultaneously.
The fitness functions should be carefully defined in a manner that takes into account
each specific clustering validity index.
• Second, there must be a specific tracking strategy to store the non-dominated
solutions (PF).
• Third, there should be an appropriate strategy for determining the clustering solu-
tion from the non-dominated Pareto near-optimal solutions.

All the aforementioned points must be taken into account when designing a new
MOEA. Algorithm 1 is the Pseudocode for an evolutionary multi-objective clustering
in general.
A Review of Multiobjective Evolutionary Algorithms … 185

Algorithm 1 Pseudocode for a general evolutionary multi-objective clustering

Input: Dataset, the parameters of the algorithm.
Output: Clustering results of dataset
i = 0;
Pi=Initialize Population(Dataset,Algorithm parameters);
Fiti=Evaluate Population(Pi);
while Stopping condition is not met do
i = i + 1;
Mati=Select(Pi-1, Fiti-1);
Crossi=Crossover(Mati);
Muti=Mutation(Crossi);
Fiti=Evaluate Population(Muti);
Pi=Update Population(Pi-1, Fiti-1,Muti, Fiti);
C=Select From Pareto Front(Pi);
end while
return C

7 Overview of Multi-objective Clustering Techniques

This section provides a detailed information about the major multi-objective evolu-
tionary methods. These methods were suggested in the literature to solve the clus-
tering problem. An overview of the key studies will be provided based on the points
discussed in Sect. 6.

7.1 Encoding Schemes of Evolutionary Clustering Solution

This subsection summarizes some of the different encoding schemes that have been
adopted in the literature to represent the evolutionary algorithm solution to the clus-
tering problem. In general, there are two common ways to represent a clustering
solution: prototype-based method and point-based method. In the prototype-based
method, the clustering solution encodes some points representing the cluster like
medoids, centroids, and modes. On the other hand, the point-based method encodes
the complete clustering solution by specifying for each point in the data space its asso-
ciated cluster. Table 1 summarizes these methods based on their encoding scheme,
advantages, and disadvantages and provides some references to the studies that have
adopted each method.

7.2 Objective Functions of MOEA for Clustering Problem

Because the clustering problem is multi-objective, there is a set of objective functions

that have to be optimized simultaneously. In the literature, different combinations
Table 1 Summary of the encoding schemes for the MOEA clustering solution
186

Encoding scheme Strategy Pros Cons References

Prototype-based Centroid-based The chromosome has real values that represent 1. The lengths of chromosomes are small 1. Cannot capture the clusters with inconsis- [12, 13, 43, 44, 46,
the coordinates of the cluster centers. If K is the 2. Less time to apply genetic operators tent shapes 47]
number of cluster centers and d is the number of 3. The algorithm converges quickly 2. Variable lengths of chromosomes will make
dimensions then the length of the chromosome 4. Determine the overlapping clusters the design of the evolutionary operators very
will be d × K complex
3. Unsuitable for high dimensional data
because this will increase the chromosome
length
Medoid-based Encode in the chromosomes the cluster medoids 1. Medoids differs from centroids in that they do Finding the medoid is computationally [48, 61]
or the indices of the points representing the not suffer from the presence of outliers expensive when the number of points is large (not commonly used
cluster medoids. The optimization process tries to 2. Medoids can be computed based on a distance in MOEA
minimize the sum of distances of the cluster matrix without the information of the coordinates clustering)
points to the medoid. The chromosome length of the points
equals the number of clusters K 3. Used with categorical dataset
Mode-based Cluster modes are used to represent the clusters. 1. Has advantages for the datasets with categorical Cluster modes may have noise or outliers [49, 51]
In chromosome representation, K represents the datatype because the centroid cannot be computed
most frequent value for that feature (cluster 2. The cluster modes are computed much faster
modes). The length of a chromosome will be than cluster medoids
K ×d
Point-based Cluster Label-based All the data points are encoded instead of only the 1. Easy decoding step because the data points are 1. The chromosome length equals to the num- [21, 29, 33, 54, 55,
representatives/prototypes of the clusters. In this already given the cluster labels in the encoding ber of points. Thus, it becomes slower for 57, 57]
representation, the length of a chromosome is 2. Can capture clusters with arbitrary shape large datasets (Most common
equal to the number of points in the dataset and 3. Suitable for any data types such as continuous 2. It produces a large number of redundant technique)
each position denotes the cluster label of the and categorical solutions
associated point. If position i of the chromosome 4. The lengths of the chromosomes are fixed and 3. Applying crossover/mutation may cause
is assigned a value k, then the i th data point is put because they do not rely on the number of clusters the clustering structure to get damaged
in cluster k. Thus, a chromosome can be encoded in the chromosome 4. Does not support fuzzy clustering since
composed of only integer values 1, 2, . . . , K , each data point is assigned to exactly one
where K is the maximum number of clusters cluster so that overlapping clusters can not
be detected
Locus-based In this representation, a chromosome has n 1. The number of clusters is encoded in the chro- 1. The chromosome length is equal to the [14, 30, 31]
Adjacency Graph positions for the number of data items. If position mosome automatically number of data items so it is not practical with
i has an integer value j, it means that i and j are 2. The number of redundant chromosomes is small large datasets
linked. These data items are added to the same 3. Can catch clusters with arbitrarily shaped 2. Clustering process does not support
cluster in the clustering solution. Thus, each fuzziness
chromosome is represented as a graph, the data
items are the vertices and the links between two
data points are the edges in the graph
R. Abu Khurma and I. Aljarah
A Review of Multiobjective Evolutionary Algorithms … 187

of clustering validity indices were used to formulate the objective functions of the
evolutionary clustering algorithm. The clustering validity indices represent differ-
ent clustering criteria such as compactness, separation, and cluster density. These
indices are either related to cluster prototypes or cluster labels. This subsection
presents some of the validity indices used in MOEA to solve the clustering problem.
Table 2 summarizes some of the MOEA objective functions used in the literature for
the clustering problem.

7.3 Design of Evolutionary Operators

An important aspect to design a multi-objective evolutionary framework is to pay

attention to the evolutionary operators such as selection, crossover, and mutation and
how to refine them to work in the multi-objective space.
Selection operators are used in the evolutionary algorithm to determine which
chromosomes will be more reproduced to generate new children for the next gen-
eration. In literature, there are many strategies for applying selection operators in
multi-objective space. The main purpose of the selection strategy is to obtain differ-
ent possible trade-offs among the fitness functions to get a variety of solutions in the
non-dominated set.
A region-based selection strategy was used to divide the objective space into a set
of hyperboles and each solution is put in a certain hyper box according to its fitness
values [17]. Thus the selection is done for a hyper box instead of an individual
solution. The main goal is to choose the hyper box with a fewer number of solutions.
Another selection strategy used in the literature is a crowding distance-based
binary tournament selection strategy [20]. A population and the child population are
combined in this strategy then a non-dominated sorting is used to give a rank for
every solution. The P top solutions are then selected (P is the population size) using
a crowding factor. The crowding factor represents the density of the area in which
the solution locates in the multi-objective space. For two solutions, the one that will
be selected is the one with a higher rank. However, if the two solutions have the
same rank then the one that will be selected is the one from a less crowded area.
Crowding distance is computed by determining the position of the solution in the
objective space and drawing a hyper box around this solution which includes other
succeeding and preceding solutions. The volume of the hyper box determines its
crowding distance. The larger crowding distance means the solution is from a less
crowded region. The idea is to have a higher diversity of solutions in the Pareto front.
Crossover is another evolutionary operator that exchanges the genes between the
chromosomes to generate new children to be used in the next generation. For the
multi-objective clustering algorithms, the crossover operator may be a single-point
or two points operation. In prototype-based chromosomes, the crossover point is
selected to fall between two cluster centroids/modes to avoid sudden disturbances
of the cluster centroids/modes [12, 49]. Two points crossover is used to deal with
Table 2 MOEA objective functions for clustering problem
Validity index Objective function Description Mathematical formula References

K  n
2
188

Based on cluster prototype Jm index The global fuzzy cluster variance Jm = um ki D (z k , xi ) (5) [33]
k=1 i=1
where D(z k , xi ) is the distance of the i th data point xi from the kth
cluster center z k . n is the number of data points, K is the number of clusters
and u mki is the membership degree of the i th data point to the kth cluster

Overall cluster deviation The crisp version of Jm index Dev(C) = D(z k , xi ) (6) [19]
Ck ∈C xi ∈Ck
k
Davies-Bouldin Index Represents a function of the ratio of the sum of 1  R (7) [71]
DB = K i
within-cluster scatter to between-cluster separation i=1  
Si +S j
Ri = max j, j =i d{i j} (8)
where Si and S j is the scatter within the i th , jth clusters, respectively. di j
is defined as the distance between the centers of the two clusters

K  n
Xie-Beni Index Defined as a function of the ratio of the total fuzzy σ = u2 2 (9) [62]
ki D (z k , xi )
cluster variance σ to the minimum separation sep of k=1 i=1 
the clusters 2
sep = min k =l D (z k , zl ) (10)
K 1
Intra-Cluster Entropy Used to measure the average purity of clusters without H= [(1 − H (Ci ).g(z i ))] k (11) [39]
using the class labels of data points i=1
where
H (Ci ) = −[g(z i ) log2 g(z i ) + (1 − g(z i ))log2 (1 − g(z i )))] (12)
g(z i ) is the average similarity between a cluster center z i and the data
points belonging to cluster i that is defined in terms of cosine distance
2 
K 
Cluster Separation Cluster Separation Sep or Inter-cluster distances is the Sep(C) = D 2 (z i , z j ) (13) [42]
K (K −1)
average distance among the cluster centers i=1 j =i
where C is the set of all clusters
K

Average Between Group Similar to the cluster separation measure that is ABG SS = n i .D 2 (z i , z) (14) [56]
Sum of Squares (ABGSS) computed as the average distance of the cluster centers i=1
where C is the set of all clusters
from the centroid of the dataset
  
δ(Ci ,C j )
Based on cluster labels Dunn Index Computes the ratio of the distance between two D N = min 1≤i≤K min 1≤ j≤K , j =i (15) [23]
max1≤k≤K Δ(Ck )
clusters δ(Ci , C j ) to the maximum diameter of the where δ(Ci , C j ) = min x ∈C ,x ∈C D(xi , x j ) (16)
clusters max1≤k≤K Δ(Ck ) i i j j
and Δ(Ci ) = max x ,x ∈C D(xi , xk ) (17)
i k i 
n   L
Cluster Connectedness Describes the connectivity level of clusters that can be Conn(C) = ki,nn (18) [33]
i=1 j=1 i, j
measured using the connectedness index between the
data point and its neighboring points where
ki,nn = 1j if Ck : i ∈ Ck ∧ nn i, j ∈ Ck and ki,nn = 0 otherwise.
ij ij
nn i j is the jt h nearest neighbor of data point i. L is a user defined
parameter that decides the number of neighbors and impacts the
R. Abu Khurma and I. Aljarah

connectedness measure

(continued)
Table 2 (continued)
Validity index Objective function Description Mathematical formula References
n
 
Edge Index Measures the distance between boundaries of the Edge(C) = − ξi j (19) [65]
clusters i=1 j∈I Fi
where ξi j = D(xi , x j )
Silhouette Index It is the ratio between ai that represents the average The silhouette width si can be formulated as [63]
distance of a point xi from the other points in same bi −ai
si = max{a (20)
cluster, and bi is the minimum of the average distances i ,bi }
and Silhouette index S is defined as follows:
of xi from the points of the other clusters n

S = n1 Si (21)
i=1
where −1 ≤ S ≤ 1
 
k / i D(x j ,xk )
Min-Max Cut It is the ratio of the distances among the points within a Min MaxCut (C) =  j∈Ci k ∈C (22) [21]
A Review of Multiobjective Evolutionary Algorithms …

cluster and the distances of the points from different i=1 j∈Ci k∈Ci D(x j ,xk )
clusters
d K  

n   1 
Total Within Cluster Describes the variance between points of the same T WCV = xi2j − nk xi j (23) [41]
Variance cluster i=1 j=1 k=1 xi ∈Ck
xi j denotes the jth feature value of the i th data point, and n k denotes the
number of points in cluster Ck
189
190 R. Abu Khurma and I. Aljarah

the variable lengths of chromosomes when the centroid is encoded in chromosomes

[43].
The mutation is an operator that causes small changes in the chromosomes to
enforce diversity in the population. There are many strategies for applying mutation
on the multi-objective clustering algorithm. These strategies can be classified based
on the encoding methods of the solution.
An example of applying mutation is generating a real value between 0 and 1 and
from a uniform distribution and use it to shift the position of a randomly selected
centroid [49]. In general, for prototype-based encoding such as medoid-based, mode-
based and centroid based, the mutation operators need to change only in few positions
because the length of the chromosomes is small so that so small changes may heavily
change the cluster structure. For point-based encoding, mutation operation takes a
longer time because the chromosome has a larger length and the mutation is done
in one or two positions in the chromosomes to update the clustering structure. Low
mutation probability does not result in much diversity in the population resulting in
a low convergence rate. To mitigate this problem, a mutation in one position results
in mutations in neighboring points also [33].

7.4 Maintaining Non-dominated Solutions

One important issue that the MOEA has to take into consideration is the preservation
of the non-dominated solutions. Storing the group of current non-dominated solutions
can be done by two ways: either by storing them in the population itself or keeping
them outside the population in a separate archive.

7.4.1 Storing Non-dominated Solutions Within Population

In this method, a non-dominated sorting is done for the current population and the
generated child population to give different ranks for the solutions. Elitism is used
then to select half of the population for the next iteration. NSGA-II-based [20] is
a multi-objective clustering algorithm that stores the non-dominated solution in the
same population. This method was used by many MOEAs including [12, 16, 61].
The advantage of this method is that the multi-objective algorithm and the single-
objective algorithm are similar. Furthermore, there is no need for a separate archive
to maintain the non-dominated front which requires additional overhead to update
the external archive in each iteration. The disadvantage of this method is that the
crowding distance criterion may be used to truncate the non-dominated solutions
when they are going beyond the population size. This means that this method cannot
be applied in some cases that require generating a large number of trade-off solutions.
For example, when the dataset is complex or the number of clusters is large.
A Review of Multiobjective Evolutionary Algorithms … 191

7.4.2 Storing Non-dominated Solutions in Archive

Archive is another method for storing non-dominated solutions. PESA-II is an exam-

ple of this method which has two populations: internal and external. The non-
dominated solutions in the internal population enter the external population (archive)
when they are non-dominated by any solution in the archive. All the other dominated
solutions are truncated using a crowding factor, when the non-dominated solutions
are added to the archive. Examples on MOEAs that adopted this method to store the
non-dominated solutions include [33, 60, 65].
The advantage of this method is that the evolutionary operators can be applied to
the internal population only. Because the archive does not include the application of
evolutionary operators, population size can be increased without increasing the com-
plexity of time. The disadvantage of this method is the requirement for maintaining
the archive and updating it at each iteration.

7.5 Deciding the Final Clustering Solution

The outcome of an MOEA for clustering is a set of trade-off solutions, that cannot be
optimized more. This is because the improvement of a solution for one objective may
lead to degrading another objective. Thus, there should be a specific method to be
applied by the MOEA clustering algorithm to obtain the final clustering solution from
the resulting non-nominated solutions. In literature, there are three main methods for
doing this based on an independent objective, finding the knee of the non-dominated
front, and using the ensemble of the non-dominated solutions.
In the first method, the final clustering solution is obtained from the set of non-
dominated front independently based on a specific validity index which was not
considered in the objective function of the MOEA [12, 46]. This is a simple method
that requires a little time complexity. However, it may be biased to the selected validity
index. In the second method, the interesting final solution is one for which the change
in one objective value results in maximizing the values of the other objectives. This
solution is very important for the non-nominated Pareto front [14, 30].
The first and second methods are based on some metrics or proprieties of the Pareto
optimal front. The major problem is that selecting one solution may deny important
clustering information that was gathered by other discarded non-dominated solutions.
To address this problem, a different ensemble method has been proposed to combine
all the non-dominated solutions in one final solution [44, 47]. The idea is to integrate
a set of MOEAs in one clustering framework and use a supervised classifier to select
among the solutions in the non-dominated set. The instances that are usually clustered
together by the majority voting of different MOEAs are most likely to be clustered
correctly. These instances are used to train the supervised classifier to cluster the
remaining instances that are not yet clustered. The positive of this method is that it
treats all the non-nominated solutions equally and gives the same importance.
192 R. Abu Khurma and I. Aljarah

8 Applications of Multi-objective Evolutionary Clustering

The MOEAs have been used to solve the clustering problem and find optimized
solutions in several domains. This section discusses some major real-life applica-
tions that have used MOEAs for clustering purposes including image processing,
bioinformatics, social networks, and more.

8.1 Image Applications

Image segmentation is a major image application that has extensively used MOEAs
to divide the image into several disjointed regions with the same features. Image
segmentation serves various fields which concern image understanding and pattern
recognition.
One of the image segmentation methods is the image clustering. In 2008, a Multi-
Objective Evolutionary Clustering Ensemble Algorithm (MECEA) was proposed
for texture image segmentation [60]. The image segmentation was applied in two
phases where MECEA was applied in the first phase to trade-off compactness within
a cluster and connectedness between clusters. The second phase used the MCLA
to select the last segmentation solution from the group of Pareto solutions obtained
from the first phase.
One year later, a new method for evolutionary multi-objective clustering image
segmentation was proposed. Two objectives were optimized simultaneously: the
overall deviation and edge value. They have a significant role in image segmentation.
A new multi-objective nature-inspired clustering (MoNiC) algorithm was pro-
posed for image segmentation [70]. The study explained identified multiple objec-
tives associated with the image segmentation problem.
In [73], a Multi-Objective Evolutionary Fuzzy Clustering (MOEFC) algorithm
was proposed. It used the decomposition strategy to divide the problem into sub-
problems where each sub-problem incorporates local image information and repre-
sents a fuzzy clustering problem. Opposition-based learning was used to improve the
search process. The results showed that the proposed clustering algorithm was able
to achieve a trade-off between two necessary complementary objectives for image
segmentation which are preserving image details and removing noise.
Later on, a Multi-Objective Evolutionary Intuitionistic Fuzzy Clustering Algo-
rithm (NR-MOEIFC) was proposed for noisy image segmentation [74]. The Intu-
itionistic Fuzzy Set (NR-IFS) was able to overcome the noise in the corrupted image.
NR-IFS considered the membership and non-membership degree to describe the
blurred features in the contaminated image. The results revealed that NR-MOEIFC
improved noise removal and segmentation performance, and at the same time reduced
the time.
For more information about this multi-objective clustering with meta-heuristic in
image segmentation, a reader can refer to a survey [15].
A Review of Multiobjective Evolutionary Algorithms … 193

8.2 Bioinformatics

MOEAs clustering algorithms have been used in wide applications in bioinformatics.

These applications include grouping coexpressed genes according to their expres-
sions because genes that have close patterns for their expressions may have similar
functionality [51]. Furthermore, microarray technology was used to monitor the lev-
els of a large number of gene expression [48]. A recent application of MOEA in
bioinformatics is the identification of complex protein such as the clustering protein
interaction networks [52].
In [48], the main objective was to search for a suitable set of cluster centers to
optimize the validity index of the cluster. For achieving this purpose, fuzzy parti-
tioning and Genetic algorithm were utilized to perform the clustering. Simultaneous
optimization of several validity indices was considered. The performance of the algo-
rithm was compared with partitional and hierarchical methods on Yeast Sporulation
and Human Fibroblasts Serum datasets.
In [68], a Differential Evolution (DE) algorithm was used to apply fuzzy clustering
Multi-Objective Optimization (MO). The target was to simultaneously optimize two
fuzzy validity indices. The user can select from the resulting Pareto optimal set from
the most promising solutions based on the problem at hand. The multi-objective DE-
variants were compared with two common MO clustering, algorithms namely the
Non-Dominated Sorting Genetic Algorithm (NSGAII) and Multi-Objective Cluster-
ing with an unknown number of Clusters K (MOCK). The results proved that DE is
a promising algorithm for MO clustering schemes.
Recently, an interactive version of MOEA has been developed to interact with
human decision-makers and periodically adjust the objective functions in an inter-
active way to cope with new gene expression data.

8.3 Social Networks

The social network is a graph that represents the interaction between individuals.
There is much useful information that can be obtained from social networks related
to social behavior, marketing, and web characteristics. Clustering plays an important
role in social networks where each cluster shares common characteristics. In [38],
proposed the integration of Dynamic and Multi-Objective Evolutionary Algorithms
(DMOEA) for online social network clustering in dynamic environments.
In [16], a new dynamic clustering multi-objective evolutionary algorithm was
proposed. The proposed algorithm was different from traditional static clustering
methods in that the length of a chromosome was variable which allowed the algo-
rithm to search for both optimal cluster center positions and cluster numbers. The
cluster number was optimized during dynamically run time instead of being pre-
specified in advance. NSGA-II algorithm was adopted to simultaneously optimize
the compactness and connectedness.
194 R. Abu Khurma and I. Aljarah

The multi-objective optimization was used to discover the evolving communities

in dynamic [26]. The proposed algorithm optimized two main objectives related to
the accuracy of the resulting clusters and the difference between successive steps.
Experiments on synthetic datasets showed improved performance of the method.
Authors in [11] proposed the multi-objective optimization to solve the community
detection problem based on a combination of the Harmony Search Algorithm (HSA)
and Chaotic Local Search (CLS). The external repository was used to save non-
dominated solutions found during the search process and a fuzzy clustering method
was used to control the size of the repository. The results from applying the proposed
algorithm on synthetic and real networks showed that more accurate community
structures were discovered.
The community detection problem was redefined in [34] using the multi-objective
model to get the intra- and inter-community structures simultaneously. MOEA/D
and the community detection model were applied in complex networks to detect the
overlapped community sets.

8.4 Other Applications

MOEAs were used in various other applications such as web recommender systems,
software module clustering, security vulnerability assessment, and classification of
actions of 3D human models.
Web recommender systems is an application used to cluster data that represents
the user sessions [22]. Sequence clustering is used to draw a weighted, undirected
graph that shows the visited webpages by a user sequentially. Similar pages are
connected via an edge. Hence, the clustering problem is actually becoming a graph
partitioning problem which is considered as NP-hard problem and needs some kinds
of MOEA to be solved. The primary objective of the paper was to determine the most
efficient MOEA to cluster sequence data by conducting multiple comparisons with
other methods in the literature on sample datasets and the best method is selected.
In [57], two novel multi-objective algorithms were proposed to automatically
solve the software module clustering problem. The main target was to improve the
program structure by utilizing an MOEA to organize the software units into modules.
The MOEA task is to identify the module boundaries by search procedure using
an objective function. The proposed objective function worked to optimize two con-
flicting objectives, namely, cohesion and coupling. The required objectives were to
increase cohesion and decrease coupling. This paper introduces the software module
clustering problem using two new multi-objective formulations, in which several dif-
ferent objectives are represented independently. The experiments were performed on
17 real-world module clustering problems. The results showed that a multi-objective
approach produced more promising solutions than the single-objective approach.
A new MOEA approach was proposed in [18] to cluster data of security assess-
ments. The clustered devices share their vulnerabilities. The results showed that the
A Review of Multiobjective Evolutionary Algorithms … 195

quality of the obtained clusters by the proposed MOEA was better than those obtained
by single-objective method.
An automatic clustering algorithm Multi-objective Immunized Particle Swarm
Optimization (MOIMPSO) [53]. The algorithm provides a Pareto optimal archive for
clustering aggressive actions of 3D human models. It was shown that the MOIMPSO
algorithm achieved a ratio of 100% classification accuracy for the physical actions
of human models.

9 Conclusion

This chapter sheds light on two main problems: the clustering problem and the
multi-objective problem. Furthermore, it reviews several multi-objective evolution-
ary clustering algorithms proposed in the literature to provide optimized solutions to
the problem of multi-objective clustering problem. The lack of systematic reviews
on this subject has motivated us to undertake this effort to prepare this work.
The chapter discusses the clustering problem and identifies it as an optimization
problem. Furthermore, it focuses on the basic concepts of multi-objective optimiza-
tion and its importance to solve the clustering problem. Subsequently, it presents
a framework for evolutionary multi-objective clustering. A detailed overview of
multi-objective clustering techniques is provided with a focus on five subjects: chro-
mosome encoding methods, objective functions for optimization, evolutionary oper-
ators, storing current non-dominated solutions, and selecting the final solution from
the non-dominated front.
Finally, a discussion is presented on the use of multi-objective clustering algo-
rithms in different applications including image segmentation, bioinformatics, social
networks, etc.
For the future, there is plenty of researches that are planned in the area of multi-
objective evolutionary clustering. One of the main challenges is the scalability prob-
lem. This problem requires the development of efficient evolutionary algorithms that
are able to cope with large scale clustering problem. These algorithms are scalable
variants for the classical multi-objective clustering algorithms that can simultane-
ously handle many conflicting validity indices and generate near-optimal Pareto front
in a large multi-objective space.

References

1. Abu Khurma, Ruba, Ibrahim Aljarah, and Ahmad Sharieh. 2020. An efficient moth flame
optimization algorithm using chaotic maps for feature selection in the medical applications.
In Proceedings of the 9th international conference on pattern recognition applications and
methods—volume 1: ICPRAM, 175–182. INSTICC, SciTePress.
196 R. Abu Khurma and I. Aljarah

2. Abu Khurma, Ruba, Ibrahim Aljarah, Ahmad Sharieh, and Seyedali Mirjalili. 2020. EvoloPy-
FS: An open-source nature-inspired optimization framework in Python for feature selection.
In Evolutionary machine learning techniques, 131–173. Springer.
3. Al-Madi, Nailah, Ibrahim Aljarah, and Simone A. Ludwig. 2014. Parallel glowworm swarm
optimization clustering algorithm based on mapreduce. In 2014 IEEE symposium on swarm
intelligence, 1–8. IEEE.
4. Al Shorman, Amaal, Hossam Faris, and Ibrahim Aljarah. 2020. Unsupervised intelligent system
based on one class support vector machine and grey wolf optimization for IoT botnet detection.
Journal of Ambient Intelligence and Humanized Computing 11 (7): 2809–2825.
5. Aljarah, Ibrahim, and Simone A. Ludwig. 2012. Parallel particle swarm optimization clustering
algorithm based on mapreduce methodology. In 2012 fourth world congress on nature and
biologically inspired computing (NaBIC), 104–111. IEEE.
6. Aljarah, Ibrahim, and Simone A. Ludwig. 2013. Mapreduce intrusion detection system based
on a particle swarm optimization clustering algorithm. In 2013 IEEE congress on evolutionary
computation, 955–962. IEEE.
7. Aljarah, Ibrahim, and Simone A. Ludwig. 2013. A new clustering approach based on glowworm
swarm optimization. In 2013 IEEE congress on evolutionary computation, 2642–2649. IEEE.
8. Aljarah, Ibrahim, and Simone A. Ludwig. 2013. Towards a scalable intrusion detection sys-
tem based on parallel PSO clustering using mapreduce. In Proceedings of the 15th annual
conference companion on Genetic and evolutionary computation, 169–170.
9. Aljarah, Ibrahim, Majdi Mafarja, Ali Asghar Heidari, Hossam Faris, and Seyedali Mir-
jalili. 2020. Clustering analysis using a novel locality-informed grey wolf-inspired clustering
approach. Knowledge and Information Systems 62 (2): 507–539.
10. Aljarah, Ibrahim, Majdi Mafarja, Ali Asghar Heidari, Hossam Faris, and Seyedali Mirjalili.
2020. Multi-verse optimizer: Theory, literature review, and application in data clustering. In
Nature-inspired optimizers, 123–141. Springer.
11. Amiri, Babak, Liaquat Hossain, and John Crowford. 2012. A multiobjective hybrid evolutionary
algorithm for clustering in social networks. In Proceedings of the 14th annual conference
companion on Genetic and evolutionary computation, 1445–1446.
12. Bandyopadhyay, Sanghamitra, Ujjwal Maulik, and Anirban Mukhopadhyay. 2007. Multiob-
jective genetic clustering for pixel classification in remote sensing imagery. IEEE Transactions
on Geoscience and Remote Sensing 45 (5): 1506–1511.
13. Bandyopadhyay, Sanghamitra, Anirban Mukhopadhyay, and Ujjwal Maulik. 2007. An
improved algorithm for clustering gene expression data. Bioinformatics 23 (21): 2859–2865.
14. Bezdek, James C. 1981. Objective function clustering. In Pattern recognition with fuzzy objec-
tive function algorithms, 43–93. Springer.
15. Bong, C.W., and M. Rajeswari. 2012. Multiobjective clustering with metaheuristic: Current
trends and methods in image segmentation. IET Image Processing 6 (1): 1–10.
16. Chen, Enhong, and Feng Wang. 2005. Dynamic clustering using multi-objective evolution-
ary algorithm. In International conference on computational and information science, 73–80.
Springer.
17. Corne, David W., Nick R. Jerram, Joshua D. Knowles, and Martin J. Oates. 2001. Pesa-II:
Region-based selection in evolutionary multiobjective optimization. In Proceedings of the 3rd
annual conference on genetic and evolutionary computation, 283–290.
18. Corral, Guiomar, Alvaro Garcia-Piquer, Albert Orriols-Puig, Albert Fornells, and Elisabet
Golobardes. 2009. Multiobjective evolutionary clustering approach to security vulnerability
assesments. In International conference on hybrid artificial intelligence systems, 597–604.
Springer.
19. Davies, David L., and Donald W. Bouldin. 1979. A cluster separation measure. IEEE Trans-
actions on Pattern Analysis and Machine Intelligence 2: 224–227.
20. Deb, Kalyanmoy, Amrit Pratap, Sameer Agarwal, and Tamt Meyarivan. 2002. A fast and elitist
multiobjective genetic algorithm: NSGA-II. IEEE Transactions on Evolutionary Computation
6 (2): 182–197.
A Review of Multiobjective Evolutionary Algorithms … 197

21. Demir, G.Nildem, A. Şima Uyar, and Şule Gündüz-Öğüdücü. 2010. Multiobjective evolution-
ary clustering of web user sessions: A case study in web page recommendation. Soft Computing
14 (6): 579–597.
22. Demir, Gül Nildem, A. Sima Uyar, and Sule Gündüz Ögüdücü. 2007. Graph-based sequence
clustering through multiobjective evolutionary algorithms for web recommender systems. In
Proceedings of the 9th annual conference on genetic and evolutionary computation, 1943–
1950.
23. Dunn, Joseph C. 1974. Well-separated clusters and optimal fuzzy partitions. Journal of Cyber-
netics 4 (1): 95–104.
24. Faris, Hossam, Ibrahim Aljarah, and Ja’far Alqatawna. 2015. Optimizing feedforward neural
networks using Krill Herd algorithm for e-mail spam detection. In 2015 IEEE Jordan conference
on applied electrical engineering and computing technologies (AEECT), 1–5. IEEE.
25. Faris, Hossam, Ibrahim Aljarah, Seyedali Mirjalili, Pedro A. Castillo, and Juan Julián Merelo
Guervós. 2016. EvoloPy: An open-source nature-inspired optimization framework in Python.
In IJCCI (ECTA), 171–177.
26. Folino, Francesco, and Clara Pizzuti. 2010. A multiobjective and evolutionary clustering
method for dynamic networks. In 2010 international conference on advances in social networks
analysis and mining, 256–263. IEEE.
27. Gzara, Mariem, and Abdelbasset Essabri. 2011. Balanced explore-exploit clustering based
distributed evolutionary algorithm for multi-objective optimisation. Studies in Informatics and
Control 20 (2): 97–106.
28. Han, Jiawei, Jian Pei, and Micheline Kamber. 2011. Data mining: Concepts and techniques.
Elsevier.
29. Handl, Julia, and Joshua Knowles. 2004. Evolutionary multiobjective clustering. In Interna-
tional conference on parallel problem solving from nature, 1081–1091. Springer.
30. Handl, Julia, and Joshua Knowles. 2005. Exploiting the trade-off—the benefits of multiple
objectives in data clustering. In International conference on evolutionary multi-criterion opti-
mization, 547–560. Springer.
31. Handl, Julia, and Joshua Knowles. 2005. Multiobjective clustering around medoids. In 2005
IEEE congress on evolutionary computation, vol. 1, 632–639. IEEE.
32. Handl, Julia, and Joshua Knowles. 2006. Multi-objective clustering and cluster validation. In
Multi-objective machine learning, 21–47. Springer.
33. Handl, Julia, and Joshua Knowles. 2007. An evolutionary approach to multiobjective clustering.
IEEE Transactions on Evolutionary Computation 11 (1): 56–76.
34. Hariz, Wisam A., Mayyadah F. Abdulhalim, et al. 2016. Improving the performance of evo-
lutionary multi-objective co-clustering models for community detection in complex social
networks. Swarm and Evolutionary Computation 26: 137–156.
35. Hastie, Trevor, Robert Tibshirani, and Jerome Friedman. 2009. Unsupervised learning. In The
elements of statistical learning, 485–585. Springer.
36. Hruschka, Eduardo Raul, Ricardo J.G.B. Campello, Alex A. Freitas, et al. 2009. A survey of
evolutionary algorithms for clustering. IEEE Transactions on Systems, Man, and Cybernetics,
Part C (Applications and Reviews) 39 (2): 133–155.
37. Kanungo, Tapas, David M. Mount, Nathan S. Netanyahu, Christine D. Piatko, Ruth Silverman,
and YWu Angela. 2002. An efficient k-means clustering algorithm: Analysis and implemen-
tation. IEEE Transactions on Pattern Analysis and Machine Intelligence 24 (7): 881–892.
38. Kim, Keehyung, Robert Ian McKay, and Byung-Ro Moon. 2010. Multiobjective evolutionary
algorithms for dynamic social network clustering. In Proceedings of the 12th annual conference
on genetic and evolutionary computation, 1179–1186.
39. Kirkland, Oliver, Victor J. Rayward-Smith, and Beatriz de la Iglesia. 2011. A novel multi-
objective genetic algorithm for clustering. In International conference on intelligent data engi-
neering and automated learning, 317–326. Springer.
40. Kodinariya, Trupti M., and Prashant R. Makwana. 2013. Review on determining number of
cluster in k-means clustering. International Journal 1 (6): 90–95.
198 R. Abu Khurma and I. Aljarah

41. Liu, Yimin, Tansel Özyer, Reda Alhajj, and Ken Barker. 2005. Integrating multi-objective
genetic algorithm and validity analysis for locating and ranking alternative clustering. Infor-
matica 29 (1).
42. Maulik, Ujjwal, and Sanghamitra Bandyopadhyay. 2002. Performance evaluation of some
clustering algorithms and validity indices. IEEE Transactions on Pattern Analysis and Machine
Intelligence 24 (12): 1650–1654.
43. Maulik, Ujjwal, Sanghamitra Bandyopadhyay, and Anirban Mukhopadhyay. 2011. Combining
pareto-optimal clusters using supervised learning. In Multiobjective genetic algorithms for
clustering, 123–145. Springer.
44. Maulik, Ujjwal, Anirban Mukhopadhyay, and Sanghamitra Bandyopadhyay. 2009. Combining
pareto-optimal clusters using supervised learning for identifying co-expressed genes. BMC
Bioinformatics 10 (1): 27.
45. Mirjalili, Seyedeh Zahra, Seyedali Mirjalili, Shahrzad Saremi, Hossam Faris, and Ibrahim
Aljarah. 2018. Grasshopper optimization algorithm for multi-objective optimization problems.
Applied Intelligence 48 (4): 805–820.
46. Mukhopadhyay, Anirban, Sanghamitra Bandyopadhyay, and Ujjwal Maulik. 2006. Clustering
using multi-objective genetic algorithm and its application to image segmentation. In 2006
IEEE international conference on systems, man and cybernetics, vol. 3, 2678–2683. IEEE.
47. Mukhopadhyay, Anirban, Sanghamitra Bandyopadhyay, and Ujjwal Maulik. 2009. Analysis
of microarray data using multiobjective variable string length genetic fuzzy clustering. In 2009
IEEE congress on evolutionary computation, 1313–1319. IEEE.
48. Mukhopadhyay, Anirban, and Ujjwal Maulik. 2007. Multiobjective approach to categorical
data clustering. In 2007 IEEE congress on evolutionary computation, 1296–1303. IEEE.
49. Mukhopadhyay, Anirban, Ujjwal Maulik, and Sanghamitra Bandyopadhyay. 2009. Multiob-
jective genetic algorithm-based fuzzy clustering of categorical attributes. IEEE Transactions
on Evolutionary Computation 13 (5): 991–1005.
50. Mukhopadhyay, Anirban, Ujjwal Maulik, and Sanghamitra Bandyopadhyay. 2015. A survey
of multiobjective evolutionary clustering. ACM Computing Surveys (CSUR) 47 (4): 1–46.
51. Mukhopadhyay, Anirban, Ujjwal Maulik, Sanghamitra Bandyopadhyay, and Carlos A. Coello
Coello. 2013. Survey of multiobjective evolutionary algorithms for data mining: Part II. IEEE
Transactions on Evolutionary Computation 18 (1): 20–35.
52. Mukhopadhyay, Anirban, Sumanta Ray, and Moumita De. 2012. Detecting protein complexes
in a PPI network: a gene ontology based multi-objective evolutionary approach. Molecular
BioSystems 8 (11): 3036–3048.
53. Nanda, Satyasai Jagannath, and Ganapati Panda. 2013. Automatic clustering algorithm based
on multi-objective immunized PSO to classify actions of 3D human models. Engineering
Applications of Artificial Intelligence 26 (5–6): 1429–1441.
54. Özyer, Tansel, Yimin Liu, Reda Alhajj, and Ken Barker. 2004. Multi-objective genetic algo-
rithm based clustering approach and its application to gene expression data. In International
conference on advances in information systems, 451–461. Springer.
55. Özyer, Tansel, Ming Zhang, and Reda Alhajj. 2011. Integrating multi-objective genetic algo-
rithm based clustering and data partitioning for skyline computation. Applied Intelligence 35
(1): 110–122.
56. Pakhira, Malay K., Sanghamitra Bandyopadhyay, and Ujjwal Maulik. 2004. Validity index for
crisp and fuzzy clusters. Pattern Recognition 37 (3): 487–501.
57. Praditwong, Kata, Mark Harman, and Xin Yao. 2010. Software module clustering as a multi-
objective search problem. IEEE Transactions on Software Engineering 37 (2): 264–282.
58. Qaddoura, Raneem, Hossam Faris, and Ibrahim Aljarah. 2020. An efficient clustering algorithm
based on the k-nearest neighbors with an indexing ratio. International Journal of Machine
Learning and Cybernetics 11 (3): 675–714.
59. Qaddoura, Raneem, Hossam Faris, Ibrahim Aljarah, and Pedro A. Castillo. 2020. Evocluster:
An open-source nature-inspired optimization clustering framework in Python. In International
conference on the applications of evolutionary computation (part of EvoStar), 20–36. Springer.
A Review of Multiobjective Evolutionary Algorithms … 199

60. Qian, Xiaoxue, Xiangrong Zhang, Licheng Jiao, and Wenping Ma. 2008. Unsupervised tex-
ture image segmentation using multiobjective evolutionary clustering ensemble algorithm. In
2008 IEEE congress on evolutionary computation (IEEE world congress on computational
intelligence), 3561–3567. IEEE.
61. Ripon, Kazi Shah Nawaz, and Mia Nazmul Haque Siddique. 2009. Evolutionary multi-objective
clustering for overlapping clusters detection. In 2009 IEEE congress on evolutionary compu-
tation, 976–982. IEEE.
62. Ripon, K.S. Nawaz, Chi-Ho Tsang, Sam Kwong, and Man-Ki Ip. 2006. Multi-objective evo-
lutionary clustering using variable-length real jumping genes genetic algorithm. In 18th inter-
national conference on pattern recognition (ICPR’06), vol. 1, 1200–1203. IEEE.
63. Rousseeuw, J.P.J. 1989. A graphical aid to the interpretation and validation of cluster analysis.
Journal of Computational Application Mathematics.
64. Sheikh, Rahila H., Mukesh M. Raghuwanshi, and Anil N. Jaiswal. 2008. Genetic algorithm
based clustering: A survey. In 2008 first international conference on emerging trends in engi-
neering and technology, 314–319. IEEE.
65. Shirakawa, Shinichi, and Tomoharu Nagao. 2009. Evolutionary image segmentation based on
multiobjective clustering. In 2009 IEEE congress on evolutionary computation, 2466–2473.
IEEE.
66. Shukri, Sarah, Hossam Faris, Ibrahim Aljarah, Seyedali Mirjalili, and Ajith Abraham. 2018.
Evolutionary static and dynamic clustering algorithms based on multi-verse optimizer. Engi-
neering Applications of Artificial Intelligence 72: 54–66.
67. Steinbach, Michael, Levent Ertöz, and Vipin Kumar. 2004. The challenges of clustering high
dimensional data. In New directions in statistical physics, 273–309. Springer.
68. Suresh, Kaushik, Debarati Kundu, Sayan Ghosh, Swagatam Das, Ajith Abraham, and Sang
Yong Han. 2009. Multi-objective differential evolution for automatic clustering with application
to micro-array data analysis. Sensors 9 (5): 3981–4004.
69. Wang, Rui, Wu Shiming Lai, Lining Xing Guohua, Ling Wang, and Hisao Ishibuchi. 2018.
Multi-clustering via evolutionary multi-objective optimization. Information Sciences 450: 128–
140.
70. Wei, Bong Chin, and Rajeswari Mandava. 2010. Multi-objective nature-inspired clustering
techniques for image segmentation. In 2010 IEEE conference on cybernetics and intelligent
systems, 150–155. IEEE.
71. Xie, Xuanli Lisa, and Gerardo Beni. 1991. A validity measure for fuzzy clustering. IEEE
Transactions on Pattern Analysis and Machine Intelligence 13 (8): 841–847.
72. Zhang, Hu, Shenmin Song, Aimin Zhou, and Xiao-Zhi Gao. 2014. A clustering based multi-
objective evolutionary algorithm. In 2014 IEEE congress on evolutionary computation (CEC),
723–730. IEEE.
73. Zhang, Mengxuan, Licheng Jiao, Wenping Ma, Jingjing Ma, and Maoguo Gong. 2016. Multi-
objective evolutionary fuzzy clustering for image segmentation with MOEA/D. Applied Soft
Computing 48: 621–637.
74. Zhao, Feng, Jiulun Fan, Hanqiang Liu, Rong Lan, and Chang Wen Chen. 2018. Noise robust
multiobjective evolutionary clustering image segmentation motivated by the intuitionistic fuzzy
information. IEEE Transactions on Fuzzy Systems 27 (2): 387–401.
A Review of Evolutionary Data
Clustering Algorithms for Image
Segmentation

Laila Al-Qaisi, Mohammad A. Hassonah, Mahmoud M. Al-Zoubi,

and Ala’ M. Al-Zoubi

Abstract Evolutionary algorithms have demonstrated their ability to solve complex

optimization problems throughout the past years. Recent studies took advantage of
such algorithms to enhance the performance of clustering algorithms in finding the
best values for necessary parameters. Image segmentation is one of the most rec-
ognized fields in clustering algorithms that extensively applied evolutionary algo-
rithms for different problems. This chapter presents a review on integrating several
evolutionary algorithms with clustering techniques to perform image segmentation.
We choose some of the most common applications on the topic, which are Medical,
Multi-objective, and Multilevel Thresholding image segmentation techniques. Then,
other applications in the field are also reviewed.

Keywords Data clustering · Image segmentation · Optimization · Evolutionary

computation · Nature-inspired algorithms · Swarm intelligence · Metaheuristics ·
Unsupervised learning

L. Al-Qaisi
Information Systems and Networks Department, Faculty of Information Technology, The World
Islamic Sciences and Education University, Amman, Jordan
e-mail: [email protected]
M. A. Hassonah
King Abdullah II School for Information Technology, The University of Jordan, Amman, Jordan
e-mail: [email protected]
M. M. Al-Zoubi
Department of Commutation and Information Technology, Yarmouk Water Company, Irbid
21110, Jordan
e-mail: [email protected]
A. M. Al-Zoubi (B)
School of Science, Technology and Engineering, University of Granada, Granada, Spain
e-mail: [email protected]; [email protected]

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer 201
Nature Singapore Pte Ltd. 2021
I. Aljarah et al. (eds.), Evolutionary Data Clustering: Algorithms and Applications, Algo-
rithms for Intelligent Systems,
https://doi.org/10.1007/978-981-33-4191-3_9
202 L. Al-Qaisi et al.

1 Introduction

As a result of the rapid growth of technology utilization all around our daily routine,
massive amount of data exists today [48]. Which needs to be recognized, categorized,
and classified in order to get the needed knowledge [4]. Therefore, techniques and
concepts of acquiring, storing and discovering are required to accomplish such tasks.
Data clustering is defined as concept which partition data into smaller several parts
called clusters. This increases the ability of user understanding for a dataset structure.
Therefore, a cluster is a set of objects that are internally similar but less or not similar
to objects belonging to other surrounding clusters. Both similarity and dissimilarity
are defined then assessed in terms of object attributes being examined [41].
Data clustering techniques can be applied to images, or image datasets in a process
called image segmentation. It is defined as partitioning an image into several sets
of pixels that are homogeneous in a certain way. Some attributes like intensity and
color are used to determine homogeneity of image segments. The concept of image
segmentation itself is considered as a preprocessing step for image processing, so it
has been applied through the literature on various fields [10].
Image segmentation techniques can be sorted as traditional and metaheuristic.
Metaheuristic techniques are described as flexible, more reliable and capable of
solving image segmentation problems more than traditional ones. Various advan-
tages such as the ability of human thinking simulation and working with their own-
ership functions along with having cost-effective, high performance and steadfast
solutions made metaheuristics more suitable for image segmentation applications.
In the literature, different algorithms and different versions of algorithms were used
to solve image segmentation problem such as Genetic Algorithm (GA) [2], Parti-
cle Swarm Optimization (PSO), Grey Wolf Optimization (GWO), and Grasshopper
Optimization Algorithm (GOA). [10, 46].
The rest of this chapter is organized as follows: Preliminary definitions for some
of the used terminologies including Clustering, Metaheuristics, and Image Segmen-
tation in Sect. 2. Section 3 reviews the work conducted on image segmentation using
metaheuristic clustering techniques. Finally, Sect. 4 concludes and summarizes the
work presented in this chapter.

2 Preliminaries

2.1 Clustering

Clustering or Cluster Analysis is the process of finding natural groupings of data

points without previous knowledge of their labels [51]. Clustering is known to be
applied and utilized in various areas, such as image processing, market segmentation,
document classification, and many other fields.
A Review of Evolutionary Data Clustering Algorithms for Image Segmentation 203

Data clustering Algorithms are categorized into four main categories: Partition-
based clustering algorithms, Fuzzy clustering, Model-based Clustering Algorithms,
and Density-based clustering algorithms [35].
Partition-based clustering algorithms are established on a general idea which is
An object belongs to exactly one group and each cluster has at least one object. In
this case, number of clusters is set by the user making the process a semi-supervised
process. While fuzzy clustering algorithms specify a set of membership coefficients
to each object which determines the degree of membership to a cluster.
Model-based clustering assumes that the given data is generated from various
multiple probability distributions such as Gaussian or multinomial, then Expectation
Maximization Algorithm is used to estimate their parameters mean and covariance
matrix. Moreover, the process of selecting optimal number of clusters may be per-
formed by Bayesian information or Akaike information criteria.
Finally, Density-based clustering algorithms make use of the fact that describes
the cluster as a connected dense component. This component may expand and grow
in any direction until it reaches a threshold. As a result, an automatic negligence of
outliers and identification of well-divided clusters of random shapes will occur.

2.2 Metaheuristic Algorithms

Nature-inspired metaheuristic algorithms have been used for optimization purposes

in various problems. They were described as powerful techniques for solving complex
optimization problems. The Main characteristic of metaheuristic algorithms is their
independence of the addressed problem. This made metaheuristic algorithms the
suitable option for optimizing discontinuous, multimodal, non-smooth, and non-
convex functions. They are used for the purpose of finding global optimal solutions
with limited computational effort [1].
Many different metaheuristic optimization techniques have been presented and
successfully applied to different problems. Examples include Genetic Algorithms
(GAs), Particle Swarm Optimization (PSO), Whale Optimization Algorithm (WOA)
[3], Salp Swarm Algorithm (SSA) [14, 15, 43], Multi-Verse Optimizer (MVO) [13,
19, 43], Water Cycle, Mine Blast and Ant Lion Optimizer (ALO), Firefly Algorithm,
etc.

2.3 Image Segmentation

Image segmentation is known as the process of decomposing an image into multiple

different parts, groups of pixels. These pixels in each part comparatively have similar
visual attributes that differ from the other portions. The purpose of segmentation is
to make the image representation simple and easy for analysis or in order to alter the
representation of the image itself. Segmentation has a wide range of applications,
204 L. Al-Qaisi et al.

such as in medical implementations (i.e., recognize the anatomical structure), satellite

imaging (i.e., identify civil areas), and so on. Image segmentation can be used for any
problem where the image needs further analysis alongside removing the noise without
the loss of important information from an image. Besides, there are five techniques
usually applied when performing the image segmentation, including, model-based,
threshold-based, region-based, feature-based clustering, and edge-based. Each tech-
nique has its negative and positive characteristics when used for a specific scenario,
application, or problem.

3 Applications of Clustering with Metaheuristic

Algorithms on Image Segmentation

3.1 Medical Image Segmentation

Medical image segmentation is one of the most common implementations, as noticed

through written works. Magnetic Resonance Imaging (MRI), Computed Tomogra-
phy (CT), X-Ray, Ultrasound, and many other types of medical images present the
sources, from which, better diagnostic insights can be obtained, hence the popularity.
Magnetic Resonance Imaging (MRI) is one of the most reliable techniques to
detect brain tumors, which are rare shaping of cells in the cerebrum, and can either
be benign (non-cancerous) or malignant (cancerous). Therefore, several image seg-
mentation methods were proposed to detect suspicious areas in MRI. Of these recent
methods is the metaheuristic clustering approach. For instance, [18] proposed an
intelligent system in 2010 to process MR images to diagnose brain tumors using
Fuzzy C-Means (FCM) alongside Genetic Algorithm (GA) and Particle Swarm Opti-
mization (PSO). While [16, 25] utilized Ant-Colony Optimizer (ACO) in the same
manner as an initial step in identifying the suspicious region in a brain MRI, and [21]
integrated Chaotic Firefly algorithm with FCM for brain tissue segmentation.
Another metaheuristic algorithm called Glowworm Swarm Optimization (GSO)
algorithm used by [23] to initialize the cluster centers for FCM algorithm and change
them through an iterative manner as well. The results showed improvements over
conventional FCM in terms of time to reach the optimal result.
On the other hand, [40] overcame the problems of being trapped into local min-
ima while selecting the cluster centers, and sensitivity to noise through proposing an
improved PSO integrated with kernelized fuzzy entropy clustering with spatial infor-
mation and bias correction. Their work was tested on simulated and real MRI brain
images and compared against 5 other well-recognized methods. The results showed
more effectiveness of the proposed method than the other methods. Similarly, Fuzzy
Firefly algorithm was used in [36] to increase the chance of finding the optimal
cluster centers in a shorter period of time by reducing the number of iterations. The
method was applied on Brain MRI and superior results were obtained.
A Review of Evolutionary Data Clustering Algorithms for Image Segmentation 205

Further work was conducted to segment five brain MRI images using K-means
with Galactic Swarm Optimization algorithm (GSOA) in [34]. Otsu’s thresholding
method was used as a fitness function for GSOA. Although the proposed required
extra time for processing the images due to the number of iterations for GSOA, it
showed more accuracy in segmenting the images than other benchmark algorithms
in different measures.
Recently, [8] proposed a clustering method using Flower Pollination Algorithm
(FPA) and Crow Search Algorithm (CSA) to process MRI images for brain tumors
and breast cancer. Their work demonstrated the ability to reach global optimal solu-
tions with the least error rates among traditional clustering methods including K-
means and FCM, and metaheuristic clustering algorithms including GA and PSO. It
is also worth mentioning that all metaheuristic clustering methods performed better
than the traditional methods.
Apart from MRI, Computed Tomography (CT) scan is also common in medical
imaging, which is often used to find irregularities in different parts of the body,
including lungs and liver. Recent metaheuristic clustering methods were applied on
CT images. In particular, Firefly algorithm was implemented in [37] to initialize
centroids for FCM algorithm to segment lung nodules in real-time CT images. The
results showed better results than in normal FCM and K-means. Ali et al. [5] applied
Grey Wolf Optimizer (GWO) with FCM for Abdominal CT Liver Parenchyma image
segmentation, where GWO is used to try to find the optimal cluster centers. In
addition, a second type of application using Ant Bee Colony (ABC) approach with
Region Growing technique for CT liver segmentation was employed as a clustering
technique itself in the same study. The best similarity index was obtained compared
to other well-known techniques.
Firefly, GA, and Teaching Learning-Based Optimization (TLBO) algorithms were
tested in [47] to initialize and update cluster centers in images produced by lung
CT scans. Nodular planes images are segmented into pixels represented by features,
which are cluster center candidates and selected or not by the metaheuristic algorithm
according to the corresponding solution.
A density-based algorithm for clustering called Density Peaks Clustering (DPC)
algorithm was applied in combination with Fruit fly Optimization Algorithm (FOA)
by [56]. The FOA was used to randomly determine the parameters of DPC such as
cluster centers. Image entropy was used a fitness measurement for FOA. FOA-DPC
performed competitively against robust methods such as k-means and GA-DPC.
Other medical issues require different forms of imaging. For example, X-ray
images are used to detect bone fracture or other abnormalities. However, in a number
of cases, the detection is delicate and requires accurate diagnosis. Therefore, many
studies applied image segmentation on X-ray as well. Das [11] used Bat Algorithm
with K-means algorithm in separate phases to enhance the x-ray image and then
produce final image, respectively.
Ultrasound image represents another source of medical images that is mainly
important for detecting a number of types of cancer. Samundeeswari et al. [45] seg-
mented ultrasound images to detect breast cancer. They used K-means with regular-
206 L. Al-Qaisi et al.

ization technique. ACO algorithm was implemented for cluster center initialization.
The results showed improvements over normal K-means algorithm.
Some of the most known types of imaging in medical field were presented in
this section. We explored recent works on image segmentation using MRI, CT, X-
ray, and Ultrasound images. It is found that metaheuristic algorithms are mostly
used for cluster center initialization, while some of them are utilized for setting
further parameters. Only a few studies tried to operate metaheuristics as clustering
mechanism themselves.

3.2 Multiobjective Image Segmentation

Solving multi-tasks or multi-mathematical optimization problems can be carried out

using a single-objective function; however, in the single-objective, the priority is usu-
ally given to one solution more than the others. Therefore, in the last decade, multiob-
jective optimization algorithms have been applied more often to overcome this issue
[32, 38, 50]. Multiobjective optimization is a non-dominant criterion to handle the
multiple optimization problems, simultaneously. These problems can range between
maximizing one problem while minimizing the other without affecting each other.
Multiobjective optimization has been utilized in different fields such as engineering
[30], economics [17, 33], science [12], and in our case, image segmentation.
Lately, multiobjective optimization for image segmentation attracted more atten-
tion in the literature, especially combined with evolutionary algorithms for cluster
problems. Fuzzy clustering is one of the most popular methods for image segmenta-
tion in multiobjective cases. The concept of fuzzy clustering (soft clustering) comes
from the probability of having data points that belong to more than one cluster,
unlike the non-fuzzy clustering, where each point belongs to distinct clusters [28].
The fuzzy clustering can be a useful tool in image segmentation and multiobjective
applications. For example, the work in [44] presented an improved multiobjective
differential evolution with fuzzy clustering to optimize different validity measures at
the same time. Their work generated a set of solutions with a non-dominated char-
acteristic for image segmentation that can easily be selected according to specific
requirements. Similar to the previous work, fuzzy clustering was also applied with a
multiobjective evolutionary algorithm and multiple images spatial to overcome the
noise in images using image segmentation [54]. The proposed approach consists of
two components, the first is to get the fuzzy information from an image, while the
second to apply the evolutionary multiobjective optimization. Furthermore, the intu-
itionistic fuzzy fitness function and evaluation are both adopted in order to optimize
and select the best solution, respectively. In a recent work from [53], the authors
proposed a fuzzy clustering and evolutionary multiobjective algorithm to preserve
significant details from images while removing the noise using decomposition tech-
niques for image segmentation. The decomposition strategy is designed to divide
the main problem into a set of sub-problems where each sub-problem depicts one
fuzzy clustering problem that corresponds to local information on the image. The
A Review of Evolutionary Data Clustering Algorithms for Image Segmentation 207

outcome of their proposed approach showed a trade-off between removing noise and
preserving details from the images for image segmentation.
In 2011, [52] presented an Artificial Immune System (AIS) for image segmen-
tation of the Synthetic Aperture Radar (SAR). The proposed approach consists of
the following phases; first, they combined the AIS and multiobjective optimization
algorithm in order to rank the clones and maintenance diversity. Secondly, the fuzzy
clustering validity integrated and then optimized. Moreover, the texture discrimina-
tion and representation feature generated and researched for an optimal solution. The
SAR with synthetic texture images was compared with several clustering algorithms
namely, Single-Objective Genetic Algorithm (SOGA), Spectral Clustering Ensem-
ble (SCE), Wavelet-Domain Hidden Markov models (HMTseg), Self-Organizing
Map (SOM), and Fuzzy C-Means (FCM). Experimental results for the proposed
method achieved better performance than the other algorithms in image segmen-
tation. Another aspect of using the fuzzy clustering and the evolutionary multiob-
jective optimization in image segmentation was presented in [55]. They proposed
a Kriging model to reduce the computational cost while applying their approach
(KRV-MRSFC) in removing noise from images. Furthermore, the angle penalized
distance is utilized in order to enhance the Kriging model accuracy rate by selecting
the optimal individuals in the exploration and exploitation phase. The results proved
that their algorithm achieved acceptable segmentation with low time cost.
On the other hand, many researchers in the literature adapted and solved image seg-
mentation problems apart from fuzzy clustering as their main approach. For instance,
Multiobjective Emperor Penguin Optimizer (MEPO) was used to automatically iden-
tify the best number of clusters in image segmentation [29]. The proposed MEPO has
two operations, calculate the cluster centroid and setting the threshold to tune these
centroids. Besides, the MEPO compared against other well-known approaches, nev-
ertheless, the proposed method outperforms all the other approaches. Another work
was conducted by [49] on image recognition using multiobjective evolutionary clus-
tering and image segmentation. In this work, two objectives were utilized, edge value
and overall deviation, and both of them were optimized using an evolutionary mul-
tiobjective algorithm to find several solutions. They tested their method on various
image segmentation problems and the performance showed good results regarding
the image segmentation technique. In [6], the authors provided a detailed review of
image segmentation problems using multiobjective clustering. They classified the
approaches into three folds, types of optimization, formulation problems, and the
kinds of datasets used. Another review proposed by [7], depicts the state-of-the-
art for image segmentation clustering through metaheuristic algorithm. Their study
discusses the non-evolutionary and evolutionary multiobjective clustering methods
alongside their requirements and challenges for image segmentation.
However, in this brief review, we covered and focused on the recent works that
utilized the evolutionary multiobjective clustering in image segmentation, unlike the
mentioned reviews.
208 L. Al-Qaisi et al.

3.3 Multilevel Thresholding Image Segmentation

Image thresholding is referred to dividing images into several class-based segments.

Bi-level thresholding is used to segment images into two classes: objects and back-
ground. Nevertheless, when images are complex and may contain several levels, bi-
level thresholding cannot segment them. In this case, multilevel thresholding tech-
nique is required. The main problem here is increased computational complexity
because of the increased levels within an image. Several studies used different meta-
heuristic algorithms to overcome drawbacks of increased levels in the problem of
multilevel thresholding [9, 22, 24, 26, 27, 31].
The study conducted by [27] used Grey Wolf Optimizer (GWO) for solving the
drawback of multilevel thresholding in image segmentation. The computation com-
plexity increases when number of thresholds increases. Two major functions were
used for the experiments, namely Kapur’s and Otsu’s entropy between class variance
functions. The proposed algorithm and functions were tested using set of benchmark
images. Afterwards, set of performance comparisons were conducted with PSO and
Bacterial Foraging Optimization (BFO). Experimental results showed that GWO
outperformed both PSO and BFO-based methods. Nevertheless, the proposed GWO
was slower than PSO, however, faster than BFO-based algorithms.
Moth-Flame Optimization (MFO) was used to overcome the problem of multilevel
thresholding in image segmentation by [26]. The threshold chosen selection criteria
were Kapur’s entropy and Otsu’s between class variance functions. The proposed
algorithm was tested and evaluated using benchmark images such as Lena, Peppers,
Baboon, Man, and Airplane. Segmentation analysis measure was Peak Signal-to-
Noise Ratio (PSNR) besides fitness function. Results proved that MFO can be used
as a solution to multilevel thresholding problem using both functions. The proposed
MFO was compared with both PSO and BFO-based methods by statistical analysis
using Anova test. It was observed that the proposed MFO outperformed both PSO and
BFO-based methods with increased stability and higher solution quality. Moreover,
multilevel thresholding computation cost was greatly reduced.
Another solution to multilevel thresholding problem was proposed by [31] using
Gravitational Search Algorithm (GSA). Basically, multilevel thresholding divides an
image into several non-overlapping regions. A novel (2D) histogram-based segmen-
tation method is used to enhance multilevel thresholding image segmentation. In the
proposed method, Kbest GSA was used to reach optimal thresholds. Furthermore, a
2D Rényi entropy was redefined for the optimization purposes. The proposed method-
ology was tested on Berkeley Segmentation Dataset and Benchmark (BSDS300) in
terms of both subjective and objective evaluations. The experimental results showed
that the proposed methodology outperformed other traditional (2D) histogram-based
segmentation methods.
Minimum Cross-Entropy Thresholding (MCET) was the area of study in [9]. The
problem was targeting the computation cost of MCET in image segmentation. A
novel Improved Particle Swarm Optimization (IPSO) algorithm was presented and
deployed to conduct multilevel thresholding image segmentation. The main idea
A Review of Evolutionary Data Clustering Algorithms for Image Segmentation 209

was that PSO generally has a problem of premature convergence. This was solved
by IPSO which decomposed high-dimensional swarm into multiple one-dimensional
swarms, then removed premature convergence from each one-dimensional swarm.
The proposed technique was tested on grey-scale images, and results showed sig-
nificantly better performance. It resulted in better MCET threshold values along
with fast and high convergence rate. Experimental results of the proposed approach
were compared with other state-of-the-art methods such as modified ABC, Cuckoo
search, Firefly, PSO, and GA algorithms. It was concluded that the proposed IPSO
approach performed better in terms of quantitative measures such as fitness function
value, CPU time and qualitative measures such as effective misclassification error,
peak signal-to-noise ratio, feature similarity index measurement, complex wavelet
structural similarity index measurement values.
A new multilevel thresholding method for color images based on Water Cycle
Algorithm (WCA) along with Masi entropy was suggested by [24] and the term
Masi-WCA was used for it. Other exploited optimization algorithms were Bat, PSO,
Wind-Driven Optimization (WDO), Monarch Butterfly Optimization (MBO), and
GOA. Two fitness functions were used along with the chosen algorithms, Tsallis
and Masi entropy methods for comparison purposes. Experiments tested the chosen
algorithm and proposed Masi-WCA in terms of quality, and statistical analysis was
provided in this regard. WCA registered a better performance in terms of quality and
fast convergence rate. Masi entropy was recommended as a fitness function to be
used for color image segmentation with WCA rather than Tsallis.
Hybrid GOA and Minimum Cross-Entropy (MCE) was proposed by [22] for mul-
tilevel threshold color image segmentation. The proposed model combined GOA with
self-adaptive differential evolution (jDE), and called GOA–jDE. This was done to
improve the efficiency of search and preserve diversity of population specifically
in last iterations. The proposed multilevel was compared with standard color image
thresholding methods along with advanced image thresholding techniques in differ-
ent criteria. Berkeley segmentation dataset was used for validating robustness of pro-
posed GOA–jDE. Both Friedman test and Wilcoxon’s rank sum test were conducted
for assessment purposes. The proposed GOA–jDE outperformed other techniques in
terms of average fitness function value, Peak Signal-to-Noise Ratio, Structural Simi-
larity Index, feature similarity index, Standard Deviation, convergence performance,
and computation time.
Overall, it can be seen that previous studies tried to solve the main drawbacks of
multilevel thresholding, specifically cost in addition of accuracy of the thresholds
with levels nearly greater than three. Some of the proposed and suggested solutions
by using different metaheuristic algorithms in recent studies are also listed, explained,
and summarized. Most of them had significantly improved performance.
210 L. Al-Qaisi et al.

3.4 Other Applications on Image Segmentation

Through the literature, several image segmentation methods were applied to various
systems to solve the problem of locating objects in different images such as Traffic
Control Systems, Self-Driving Cars, and Locating objects in satellite images.
Rodrigues et al. [42] conducted a study on image segmentation using Firefly
Algorithm (FFA). They applied five different functions for evaluating the proposed
mechanism: Cross-Entropy, Exponential Cross-Entropy, Modified Cross-Entropy,
Shannon and Tsallis entropy measurements. The selected images were taken from
previously conducted researches on image processing and synthetic data observed
from 2-, 3-, 4-, and 5-modal Gaussians. Conducting experiments on real images
showed that FFA is a great alternative, where Tsallis entropies outperformed the rest
tested functions as a kernel and proved to be better than cross-entropy which was
showed ability to have the best results in previous research conducted by [20].
A study was performed by [39] using K-means. They introduced a criterion to
decrease number of clusters while selecting the best segments; in order to overcome
the problem of depending on the number of clusters. The difference between intra-
class and inter-class was used to overcome the shortcoming of initialization to form
the optimal criterion. The proposed criterion was tested and evaluated using eight dig-
ital images. Results showed that segmented images using the proposed criterion had
stability and robustness. Also, the proposed criterion was found to be outperforming
the traditional K-means methods with increased consistency.
Eventually, this section demonstrates the broadness of image segmentation appli-
cations by presenting a number of studies that applied metaheuristic algorithms on
cluster analysis. The experiments were conducted on benchmark images. This indi-
cates that the same algorithms and techniques are open to be utilized in more fields
such as Traffic Control Systems, Self-Driving Cars, and Locating objects in satellite
images.

4 Conclusion and Future Directions

This chapter reviewed a number of studies for performing clustering mechanism

on image segmentation with integration of metaheuristic algorithms. Studies for
Medical, Multiobjective, and Multilevel Thresholding applications were addressed,
as well as studies of other applications were generally viewed.
It is well-observed that utilizing metaheuristic algorithms helped in overcoming
the many issues faced by clustering algorithms, primarily to find global optimal solu-
tions for computationally hard problems. The integration point where metaheuristic
algorithms were exploited for the sake of clustering was mostly observed in clus-
ter center initialization, for example, for K-means and Fuzzy C-means algorithms.
Other studies were found to use metaheuristic algorithms to mimic the clustering
mechanism itself.
A Review of Evolutionary Data Clustering Algorithms for Image Segmentation 211

References

1. Abdel-Basset, Mohamed, Laila Abdel-Fatah, and Arun Kumar Sangaiah. 2018. Metaheuristic
algorithms: A comprehensive review. In Computational intelligence for multimedia big data
on the cloud with engineering applications, 185–231. Elsevier.
2. Al-Qudah, Dana A., Al-Zoubi, Ala’ M., Pedro A. Castillo-Valdivieso, and Hossam Faris. 2020.
Sentiment Analysis for e-Payment Service Providers Using Evolutionary eXtreme Gradient
Boosting. IEEE Access, 8:189930–189944.
3. Al-Zoubi, Ala’ M., Hossam Faris, Ja’far Alqatawna, and Mohammad A. Hassonah. 2018.
Evolving support vector machines using whale optimization algorithm for spam profiles detec-
tion on online social networks in different lingual contexts. Knowledge-Based Systems 153:
91–104.
4. Al-Zoubi, Ala’ M., Ali Rodan, and Azmi Alazzam. 2018. Classification model for credit data.
In 2018 fifth HCT information technology trends (ITT), 132–137. IEEE.
5. Ali, Ahmed Fouad, Abdalla Mostafa, Gehad Ismail Sayed, Mohamed Abd Elfattah, and
Aboul Ella Hassanien. 2016. Nature inspired optimization algorithms for CT liver segmen-
tation. In Medical imaging in clinical applications, 431–460. Springer.
6. Bong, Chin-Wei, and Mandava Rajeswari. 2011. Multi-objective nature-inspired clustering and
classification techniques for image segmentation. Applied Soft Computing 11 (4): 3271–3282.
7. Bong, C.W., and M. Rajeswari. 2012. Multiobjective clustering with metaheuristic: Current
trends and methods in image segmentation. IET Image Processing 6 (1): 1–10.
8. Braik, Malik, Alaa Sheta, and Sultan Aljahdali. 2019. Diagnosis of brain tumors in mr images
using metaheuristic optimization algorithms. In International conference Europe Middle East
& North Africa information systems and technologies to support learning, 603–614. Springer.
9. Chakraborty, Rupak, Rama Sushil, and M.L. Garg. 2019. An improved PSO-based multilevel
image segmentation technique using minimum cross-entropy thresholding. Arabian Journal
for Science and Engineering 44 (4): 3005–3020.
10. Chouhan, Siddharth Singh, Ajay Kaul, and Uday Pratap Singh. 2018. Soft computing
approaches for image segmentation: A survey. Multimedia Tools and Applications 77 (21):
28483–28537.
11. Das, Goutam. 2013. Bat algorithm based softcomputing approach to perceive hairline bone
fracture in medical x-ray images. International Journal of Computer Science & Engineering
Technology (IJCSET) 4 (04).
12. Di Barba, Paolo, Ivo Dolezel, Maria Evelina Mognaschi, Antonio Savini, and Pavel Karban.
2014. Non-linear multi-physics analysis and multi-objective optimization in electroheating
applications. IEEE Transactions on Magnetics 50 (2): 673–676.
13. Faris, Hossam, Mohammad A. Hassonah, Ala’ M. Al-Zoubi, Seyedali Mirjalili, and Ibrahim
Aljarah. 2018. A multi-verse optimizer approach for feature selection and optimizing SVM
parameters based on a robust system architecture. Neural Computing and Applications 30 (8):
2355–2369.
14. Faris, Hossam, Ali Asghar Heidari, Ala’ M. Al-Zoubi, Majdi Mafarja, Ibrahim Aljarah,
Mohammed Eshtay, and Seyedali Mirjalili. 2020. Time-varying hierarchical chains of salps
with random weight networks for feature selection. Expert Systems with Applications 140:
112898.
15. Faris, Hossam, Majdi M. Mafarja, Ali Asghar Heidari, Ibrahim Aljarah, Ala’ M. Al-Zoubi,
Seyedali Mirjalili, and Hamido Fujita. 2018. An efficient binary salp swarm algorithm with
crossover scheme for feature selection problems. Knowledge-Based Systems 154: 43–67.
16. Gahukar, Sayali D., and S.S. Salankar. 2014. Segmentation of MRI brain image using fuzzy C
means for brain tumor diagnosis. International Journal of Engineering Research and Applica-
tions 4 (4): 107–111.
17. Gharavi, H., M.M. Ardehali, and S. Ghanbari-Tichi. 2015. Imperial competitive algorithm opti-
mization of fuzzy multi-objective design of a hybrid green power system with considerations
for economics, reliability, and environmental emissions. Renewable Energy 78: 427–437.
212 L. Al-Qaisi et al.

18. Gopal, N. Nandha, and M. Karnan. 2010. Diagnose brain tumor through MRI using image
processing clustering algorithms such as fuzzy C means along with intelligent optimization
techniques. In 2010 IEEE international conference on computational intelligence and comput-
ing research, 1–4. IEEE.
19. Hassonah, Mohammad A., Rizik Al-Sayyed, Ali Rodan, Ala’ M. Al-Zoubi, Ibrahim Aljarah,
and Hossam Faris. 2019. An efficient hybrid filter and evolutionary wrapper approach for
sentiment analysis of various topics on Twitter. Knowledge-Based Systems 105353
20. Horng, Ming-Huwi, and Ren-Jean Liou. 2011. Multilevel minimum cross entropy threshold
selection based on the firefly algorithm. Expert Systems with Applications 38 (12): 14805–
14811.
21. Jansi, S., and P. Subashini. 2015. A novel fuzzy clustering based modified firefly algorithm
with chaotic map for MRI brain tissue segmentation. MAGNT Research Report 3 (1): 52–58.
22. Jia, Heming, Chunbo Lang, Diego Oliva, Wenlong Song, and Xiaoxu Peng. 2019. Hybrid
grasshopper optimization algorithm and differential evolution for multilevel satellite image
segmentation. Remote Sensing 11 (9): 1134.
23. Kalaiselvi, T., P. Nagaraja, and Z. Abdul Basith. Fuzzy C-means clustering with GSO based
centroid initialization for brain tissue segmentation in MRI head scans. In Computational
methods, communication techniques and informatics, 53.
24. Kandhway, Pankaj, and Ashish Kumar Bhandari. 2019. A water cycle algorithm-based multi-
level thresholding system for color image segmentation using Masi entropy. Circuits, Systems,
and Signal Processing 38 (7): 3058–3106.
25. Karnan, M., and T. Logheshwari. 2010. Improved implementation of brain MRI image seg-
mentation using ant colony system. In 2010 IEEE international conference on computational
intelligence and computing research, 1–4. IEEE.
26. Khairuzzaman, Abdul Kayom Md, and Saurabh Chaudhury. 2017. Moth-flame optimization
algorithm based multilevel thresholding for image segmentation. International Journal of
Applied Metaheuristic Computing (IJAMC) 8 (4): 58–83.
27. Khairuzzaman, Abdul Kayom Md, and Saurabh Chaudhury. 2017. Multilevel thresholding
using grey wolf optimizer for image segmentation. Expert Systems with Applications 86: 64–
76.
28. Klawonn, Frank, and Frank Höppner. 2003. What is fuzzy about fuzzy clustering? Understand-
ing and improving the concept of the fuzzifier. In International symposium on intelligent data
analysis 254–264. Springer.
29. Kumar, Dinesh, Vijay Kumar, and Rajani Kumari. 2019. Automatic clustering using quantum-
based multi-objective emperor penguin optimizer and its applications to image segmentation.
Modern Physics Letters A 1950193.
30. Mirjalili, Seyedali, Pradeep Jangir, and Shahrzad Saremi. 2017. Multi-objective ant lion opti-
mizer: A multi-objective optimization algorithm for solving engineering problems. Applied
Intelligence 46 (1): 79–95.
31. Mittal, Himanshu, and Mukesh Saraswat. 2018. An optimum multi-level image thresholding
segmentation using non-local means 2D histogram and exponential Kbest gravitational search
algorithm. Engineering Applications of Artificial Intelligence 71: 226–235.
32. Monsef, H., M. Naghashzadegan, A. Jamali, and R. Farmani. 2019. Comparison of evolutionary
multi objective optimization algorithms in optimum design of water distribution network. Ain
Shams Engineering Journal 10 (1): 103–111.
33. Mousavi-Avval, Seyed Hashem, Shahin Rafiee, Mohammad Sharifi, Soleiman Hosseinpour,
Bruno Notarnicola, Giuseppe Tassielli, and Pietro A. Renzulli. 2017. Application of multi-
objective genetic algorithms for optimization of energy, economics and environmental life
cycle assessment in oilseed production. Journal of Cleaner Production 140: 804–815.
34. Nanda, Satyasai Jagannath, Ishank Gulati, Rajat Chauhan, Rahul Modi, and Uttam Dhaked.
2019. A K-means-galactic swarm optimization-based clustering algorithm with Otsu’s entropy
for brain tumor detection. Applied Artificial Intelligence 33 (2): 152–170.
35. Nerurkar, Pranav, Archana Shirke, Madhav Chandane, and Sunil Bhirud. 2018. Empirical
analysis of data clustering algorithms. Procedia Computer Science 125: 770–779.
A Review of Evolutionary Data Clustering Algorithms for Image Segmentation 213

36. Pant, Anurag, Sai Srujan Chinta, and Balakrushna Tripathy. 2018. Kernelised clustering algo-
rithms fused with firefly and fuzzy firefly algorithms for image segmentation. In International
conference on soft computing systems, 125–132. Springer.
37. Parveen, S. Shaik, and C. Kavitha. 2015. Segmentation of CT lung nodules using FCM with fire-
fly search algorithm. In 2015 international conference on innovations in information, embedded
and communication systems (ICIIECS), 1–6. IEEE.
38. Pasquier, Philippe, Angelo Zarrella, and Denis Marcotte. 2019. A multi-objective optimization
strategy to reduce correlation and uncertainty for thermal response test analysis. Geothermics
79: 176–187.
39. Pei, Jialun, Long Zhao, Xiangjun Dong, and Xue Dong. 2017. Effective algorithm for deter-
mining the number of clusters and its application in image segmentation. Cluster Computing
20 (4): 2845–2854.
40. Pham, Thuy Xuan, Patrick Siarry, and Hamouche Oulhadj. 2018. Integrating fuzzy entropy
clustering with an improved PSO for MRI brain image segmentation. Applied Soft Computing
65: 230–242.
41. Ramadas, Meera, and Ajith Abraham. 2019. Metaheuristics for data clustering and image
segmentation. Springer.
42. Rodrigues, Paulo S., Guilherme A. Wachs-Lopes, Horst R. Erdmann, Monael P. Ribeiro, and
Gilson A. Giraldi. 2017. Improving a firefly meta-heuristic for multilevel image segmentation
using Tsallis entropy. Pattern Analysis and Applications 20 (1): 1–20.
43. Sadiq, Ali Safa, Hossam Faris, Ala’ M. Al-Zoubi, Seyedali Mirjalili, and Kayhan Zrar Ghafoor.
2019. Fraud detection model based on multi-verse features extraction approach for smart city
applications. In Smart cities cybersecurity and privacy, 241–251. Elsevier.
44. Saha, Indrajit, Ujjwal Maulik, and Sanghamitra Bandyopadhyay. 2009. An improved multi-
objective technique for fuzzy clustering with application to IRS image segmentation. In Work-
shops on applications of evolutionary computation, 426–431. Springer.
45. Samundeeswari, E.S., P.K. Saranya, and R. Manavalan. 2016. Segmentation of breast ultra-
sound image using regularized K-means (RekM) clustering. In 2016 international conference
on wireless communications, signal processing and networking (WiSPNET), 1379–1383. IEEE.
46. Saxena, Varun, Deeksha Goel, and Tarun Kumar Rawat. 2018. Image segmentation using
meta-heuristic algorithms. In 2018 international conference on computing, power and com-
munication technologies (GUCON), 661–666. IEEE.
47. Shakibapour, Elham, António Cunha, Guilherme Aresta, Ana Maria Mendonça, and Aurélio
Campilho. 2019. An unsupervised metaheuristic search approach for segmentation and volume
measurement of pulmonary nodules in lung CT scans. Expert Systems with Applications 119:
415–428.
48. Sheta, Alaa, Hossam Faris, Ali Rodan, Elvira Kovač-Andrič, and Al-Zoubi, Ala’ M. 2018.
Cycle reservoir with regular jumps for forecasting ozone concentrations: two real cases from
the east of Croatia. Air Quality, Atmosphere & Health, 11 (5):559–569.
49. Shirakawa, Shinichi, and Tomoharu Nagao. 2009. Evolutionary image segmentation based on
multiobjective clustering. In 2009 IEEE congress on evolutionary computation, 2466–2473.
IEEE.
50. Stupar, Andrija, Timothy McRae, Nenad Vukadinovic, Aleksandar Prodic, and Josh A. Taylor.
2019. Multi-objective optimization of multi-level DC-DC converters using geometric program-
ming. IEEE Transactions on Power Electronics.
51. Xu, Rui, and Don Wunsch. 2008. Clustering, vol. 10. Wiley.
52. Yang, Dongdong, Licheng Jiao, Maoguo Gong, and Fang Liu. 2011. Artificial immune multi-
objective SAR image segmentation with fused complementary features. Information Sciences
181 (13): 2797–2812.
53. Zhang, Mengxuan, Licheng Jiao, Wenping Ma, Jingjing Ma, and Maoguo Gong. 2016. Multi-
objective evolutionary fuzzy clustering for image segmentation with MOEA/D. Applied Soft
Computing 48: 621–637.
54. Zhao, Feng, Hanqiang Liu, Jiulun Fan, Chang Wen Chen, Rong Lan, and Na Li. 2018. Intu-
itionistic fuzzy set approach to multi-objective evolutionary clustering with multiple spatial
information for image segmentation. Neurocomputing 312: 296–309.
214 L. Al-Qaisi et al.

55. Zhao, Feng, Zhe Zeng, Han Qiang Liu, and Jiu Lun Fan. 2019. A Kriging-assisted reference
vector guided multi-objective evolutionary fuzzy clustering algorithm for image segmentation.
IEEE Access 7: 21465–21481.
56. Zhu, Hong, Hanzhi He, Jinhui Xu, Qianhao Fang, and Wei Wang. 2018. Medical image seg-
mentation using fruit fly optimization and density peaks clustering. Computational and Math-
ematical Methods in Medicine 2018.
Pavement Infrastructure Asset
Management Using Clustering-Based
Ant Colony Optimization

Saqib Gulzar and Hasnain Ali

Abstract Pavement asset management comprises of all the activities involved in the
planning stage and design stage followed by execution involving inventory, resource
allocation, and construction activities of a pavement structure. It also comprises of
maintenance activities of the pavement infrastructural system during the design life
supported by continuous rehabilitation of pavement sections whenever and wherever
necessary. To achieve the above goals efficiently and effectively, management strate-
gies and tools are required which help the planners and decision-makers choose the
optimum combination of activities and resources, based on defined multi-objective
setups. Ant colony optimization (ACO) technique offers one such option to improve
the efficiency of decision-making and clustering analysis in pavement asset man-
agement. It is inspired by the natural behavioral system of ants and its probabilistic
formulations which lead to the selection of optimum strategies. It extends the scope
of planning through a feedback mechanism in the management of pavement systems.
In this book chapter, the need for the management of pavement assets is deliberated,
followed by the detailed theoretical background of ant colony optimization and its
possible application in the pavement asset management is deliberated. A framework
is proposed for the application of ACO in pavement management system. Further,
an ACO based clustering approach is deliberated to choose optimum bids for asphalt
pavement construction projects based on time of completion, total cost, and quality
of construction.

Keywords Ant colony · Data clustering · Optimization · Evolutionary

computation · Nature-inspired algorithms · Metaheuristics · Unsupervised
learning · Machine learning · Data mining

S. Gulzar (B) · H. Ali

Formerly, Indian Institute of Technology Delhi, New Delhi, India
e-mail: [email protected]
H. Ali
e-mail: [email protected]

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer 215
Nature Singapore Pte Ltd. 2021
I. Aljarah et al. (eds.), Evolutionary Data Clustering: Algorithms and Applications, Algo-
rithms for Intelligent Systems,
https://doi.org/10.1007/978-981-33-4191-3_10
216 S. Gulzar and H. Ali

1 Introduction

The importance of pavement management system can be realized from the fact that
the current road system has been deficient in meeting the traffic demand over the years
owing to the improper and staggered planning coupled with the inherent deficiencies
in the geometric design of roadways, construction methodologies employed which
are very inefficiently executed and ill-planned maintenance activities add to the over-
all chaos [1]. Further, the need for proper use of resources subject to the constraints of
feasibility, geographic and climatic conditions, have made it imperative to establish
a code of practice for managing the pavement assets efficiently. Even though every
year, road maintenance activities are carried out, there are multiple premature pave-
ment failures suggesting that the pavements are not being designed, constructed, and
maintained properly; the drainage system is not actively synchronized with pavement
construction leading to poor performance of pavement structures and the frequent
crushing, cutting, and filling of pavements for laying service lines indicate the mis-
management, lack of coordination in the implementation of developmental activities,
and improper planning in the overall constructional development setup. The diffi-
culties of not being able to predict the deterioration of pavements and design correct
maintenance strategies have led to poor pavement performance, both structurally
and functionally, leading to the increase in vehicle operating costs, accidents, and
competing demands [2, 3]. Many a times the use of advanced materials helps in
improving pavement performance [4, 5]. A pavement management system indicates
the structure needed to prevent acceptable deterioration level during its design life
and permits the prediction of structural and functional assessment of the deterio-
ration condition. It outlines what potential projects are essential and where these
should be done, and the type of improvements needed to prevent premature failure
of pavements and also suggests means and alternate ways of extending the pavement
life by operational and maintenance synchronization across multiple project levels.
The pavement management system (PMS) focusses on optimizing (minimizing)
the overall cost considering the road user cost, pavement life cycle cost including
the road condition and road administrator costs to reach a minimum cost which
satisfies all the basic tenets of the stakeholders and efficient use of resources as shown
in Fig. 1.
The overview of the components of a pavement management system is shown
in Fig. 2. The data about the condition of the road and its inventory along with the
pavement information which consists of surface profile, structural design, frictional
and surface texture characteristics, etc., taken together are fed into the database. This
database contains the present, as well as historical data, which is used for evalu-
ating the pavement performance, optimal resources required, road user demands,
road surface requirements, structural requirements, pavement layer requirements,
material requirements, maintenance projects with schedules and also enables the
selection of approaches for pavement monitoring, decisions for resurfacing, repair
or rehabilitation, best management strategy for pavement construction, maintenance,
and rehabilitation. Thus, a complete pavement management system comprises of a
Pavement Infrastructure Asset Management Using Clustering … 217

Fig. 1 Optimizing overall cost

Fig. 2 Overview of PMS components

coordinated system of activities, all focused towards accomplishing the best value
possible for the available public funds in constructing and maintaining safe, smooth,
and economical road pavement networks.
The operation levels within a PMS are shown in Fig. 3, which can be divided under
network level and project level activities, all linked to the database which forms the
core for any management system. It should be noted that engineering analysis and
research to come up with better evaluation models [6], optimization techniques [7]
and management strategies [8] should also be funded to enable efficient utilization
of resources which keep on diminishing with each passing year [9].
218 S. Gulzar and H. Ali

Fig. 3 Overview of PMS at network and project levels and their inter-relationships

2 Theoretical Background

2.1 Optimization Techniques

Optimization techniques form as a component of the subject of operations research

which deals with the implementation of mathematical techniques to decision-making
scenarios and problems in order to establish the best or optimal solutions based on the
given objective function subject to the existing or future constraints. The optimization
techniques vary based on the problems which are classified on the basis of constraint
existence, typical nature of variables of design, structure of the overall problem,
type of equations involved, allowable values of design variables, deterministic or
probabilistic nature of all variables, decoupling nature, and total number of objective
functions. These optimization methods may be purely mathematical, stochastic or
statistical in nature. With the advancement of understanding of biological systems,
many unconventional optimization techniques have been developed such as algo-
rithms based on genetics (GA), simulated annealing (SA), particle swarm optimiza-
tion (PSO), ant colony optimization (ACO), neural networks, fuzzy optimization,
etc. In this paper, the ant colony optimization technique is evaluated for possible use
in pavement management systems and discussed in detail in subsequent sections.
Pavement Infrastructure Asset Management Using Clustering … 219

2.2 Ant Colony Optimization (ACO)

Ant colony optimization is based on the real coordinative behavioral traits of ants
which are able to converge and locate the shortest path to the source of food from
their homes. The preceding ants who successfully find food leave a volatile substance
called pheromone behind them on their path which guides the following ants to reach
the food source through the successful shortest path. This forms the communication
basis among ants and enables them to solve complex problems, moving around
a series of obstacles each time finding an optimum pathway leading them to the
food source. Various experimental studies have verified this fact that an ant group
perform unanimously and find the shortest path from their home to the food source
spatially through their communication methodology involving pheromone deposition
and evaporation, allowing for path update as when the pheromone evaporates [10,
11]. The optimization algorithms taking inspiration from ant’s behavior in finding
the shortest paths by using only information based on the chemical substance called
pheromone, initially developed in 1990s, have been used for various discrete type
optimization problems and have performed successfully [12–15].

2.3 ACO Behavioral Process

The ACO process can be understood by qualitatively representing the complete dis-
crete optimization problem like a set of multilayered, multistep series of alternatives
as shown in Fig. 4. The total number of design variables involved in the optimization
problems is set equal to the number of layers or steps in the graphical representation
while the count of nodes in each step represents the permissible value of each vari-
able of design subject to the constraints and equal the total count of discrete values
allowed for the respective design variable. Thus, each individual node is associated
with an allowable discrete value of the design variable of the optimization problem.
Assuming that an ant colony consists of N ants starting from their home node
and travel towards finding a food destination. The ants travel through a number of
layered framework from the first layer to the last layer, and in each iteration cycle
move towards their food but end up at a destination node in the way. At each node,
each ant has a choice of selecting multiple paths but selects only one based on the
pheromone deposition. Each ant makes a choice at each node and that forms a local
solution, however, this continues at each node and with each iteration, the global
path selected based on the pheromone concentration represents the optimal path as
ant progress towards their food source [16]. This optimal path is shown as a bold line
shown in Fig. 4. The probability of choosing the next node is called the ant searching
transition rule and the trail followed by the following ants is governed by the local
retracing and updating rule as explained in the subsequent sections.
At the start of the optimization process in the first iteration, all paths are allotted an
equal concentration of pheromone, thereby, in the first iteration, all ants begin from
220 S. Gulzar and H. Ali

Fig. 4 ACO behavioral process

Fig. 5 a Ants exploring all paths with equal probability; b As time progresses, the shorter path
witnesses more ants; c All ants converging on the shorter path

home or nest node and reach food or destination nodes by randomly or haphazardly
selecting any node in each step. However, with the next iteration, the paths are
updated based on the updating rule and finally the optimization process ends up if the
maximum number of iterations specified is reached or the single solution emerges
out within the specified number of iterations. The optimal solution consists of all
those nodes representing each design variable, which lie on the path containing the
highest amount of pheromone concentration. Thus, at the optimum, all ants choose
the same best path converged after a number of iterations, which is the shortest and
most efficient way of traveling from home node to food destination mode.
The ant colony optimization process has been illustrated in Fig. 5, wherein ants
travel from home to food initially through all routes but with each iteration converge
to an optimum path represented by the shortest path between the home node and the
food destination node [17].
Pavement Infrastructure Asset Management Using Clustering … 221

2.4 Ant Searching Transition Rule

An ant r , positioned at an ith node needs to choose the next jth node based on
the pheromone deposition between the two nodes equal to ψiδj . This probability of
choosing jth node from the preceding ith node based on the pheromone deposition
ψiδj is called an ant searching transition rule and is given below

ψiδj
Pi(rj ) =  δ , ∀ j ∈ Ni(r )
ψi j (1)
j∈Ni(r )

Pi(rj ) = 0, ∀ j ∈
/ Ni(r ) (2)

where δ denotes the degree of pheromone importance and Ni(r ) denotes all the nodes
around node i when r th ant is position at an ith node. It should be noted that Ni(r )
contains all nodes surrounding ith node except the preceding nodes which has been
already visited before. This helps in preventing an ant from going back to the same
node which has already been visited before node i. This formulation is applied at
each node until an ant traveling from home node reaches food destination nodes and
subsequently applied to each ant each iteration coupled with updating rule as given
in the next section.

2.5 Local Retracing and Updating Rule

Before coming back to the home node, the r th ant deposits ψ (r ) of pheromone on
the path segments it has gone through. The pheromone value ψi j on the path segment
(i, j) traveled is updated as per the below local retracing and updating rule

ψi j ← ψi j + ψ (r ) (3)

This updating causes an increase in the pheromone deposit on the path segment
(i, j), thus the probability of this path segment being selected from the following acts
will increase, as they’ll be choosing the subsequent nodes as per the ant searching
transition rules which is based on the amount of pheromone deposition along each
segment connecting the nodes across different layers.
222 S. Gulzar and H. Ali

2.6 Evaporation of Pheromone Deposits

One of the important aspects of ant colony optimization is the evaporation of

pheromone deposits. The pheromone evaporation helps in the exploration of alternate
paths during the ant searching transition process and avoids pre-timed convergence to
the local sub-optimal solutions, thereby, eliminating the poor selection of segments
in the process of path selection. It also helps in limiting the maximum value of the
pheromone deposit attainable on any path segment.
When an r th ant travels to the next node, the pheromone evaporates from all
the path segments (i, j) in accordance to the pheromone evaporation rules as given
below

ψi j ← (1 − ρ)ψi j , ∀(i, j) ∈ S (4)

where ρ ∈ (0, 1] is a parameter representing the rate of evaporation and is called as

pheromone decay factor and S denotes all the segments traveled by an ant r in its
path from home nest to food destination.
One complete iteration consists of a complete cycle of ant’s searching transi-
tion or movement, evaporation of pheromone, and deposition of pheromone through
updating. Once all ants have returned from the food source to the home node, the
pheromone deposit information is updated according to the updating relation, but
considering the contribution of the best ant to the path segments which is inversely
proportional to the length of the path traveled by the r th ant. The global updating
principle at the end of each iteration is given by the below relation.


n
ψi j = (1 − ρ)ψi j + ψi(rj ) ;
r =1
where, ψi(rj ) = LBr (5)
B = An arbitrary constant
L r = Path length traveled by r th ant

The best ant pheromone deposition equation ψi(rj ) = LBr is implemented as per
the below scheme

αξbest
(r ) , if (i, j) ∈ global optimum path
ψi j = ξwor st (6)
0, otherwise

where ξbest is the best value while ξwor st is the worst value of the governing objective
function among all the paths traveled by a total of N ants and α is a parameter to
adjust the scale of the updating of pheromone at a global level. A higher value of
α indicates that higher amount of pheromone has been deposited on the global best
Pavement Infrastructure Asset Management Using Clustering … 223

path, where more ants will exploit this path in the forthcoming iterations. This helps
in assigning greater pheromone levels to the paths which results in better objective
function values.

2.7 Global Implementation Algorithm

The algorithm for implementing ant colony optimization technique for a minimiza-
tion problem such as pavement management system which requires to achieve best
possible use of resources and services at the minimum cost is given in Table 1. It
involves four steps consisting series of sub-schemes starting with basic assumptions,
followed by computation of probability of path segment selection and evaluation of
objective function for each set of design variables for each ant and determination of
best and worst values, following which test for convergence is done and subsequently
pheromone evaporation and updating is done for the next iteration until we get the
optimum path or the maximum iterations number is achieved.

3 ACO in Pavement Management System

3.1 Prospects

The construction of a roadway involves numerous resources, labor, and machinery

amounting to crores of rupees and once construction is completed, the next most
important task is the management of pavement infrastructure, maintenance activities
which have to be periodic as these help in extending the service life of pavements,
improve their functional, as well as structural performance, and ensure sufficient
drainage systems for the pavement structure. The road infrastructure contributes a
lot to the economy of a nation, thus with increasing awareness about the importance
of an efficient road infrastructure at the policy level, more funds are being allocated
to the national highway development and maintenance activities across all countries,
especially in the developing world. Therefore, it is very important that all the con-
struction, as well as maintenance work, should be well planned, ensuring the best
possible utilization of all the resources subject to the constraint of resources and with
the climate change happening, more focus has been on the efficient use of natural
resources involved in the construction of highways, as well as reduction in the emis-
sion during asphalt laying and vehicle operations, with well-maintained highways
reducing the vehicle operating costs and emissions. Further, it is very important to
select the optimum strategy for repair, rehabilitation, and maintenance works, as
well as to cater the demands of these works, depend on how well the planning has
been done. Therefore, it becomes inevitably essential to carry out the optimization of
various alternatives and select the best to avoid any losses and reduce the complexity
224 S. Gulzar and H. Ali

Table 1 Global ant colony optimization algorithm

Step 1: Select Hyper-parameters
Assume the following
• N = Number of ants in the colony
• Permissible discreet values of n design variables
• Add equal amounts of pheromone ψi1j initially along all segments
• Set iteration number l = 1
• Pheromone decay factor, ρ = 0.5 to 0.8
Step 2: Compute path selection probability
2.1 Compute the probability of selecting segment between (i, j)
(1)
ψi j
Pi j = p (1) ; (7)
m=1 ψim
i = 1, 2, . . . , m;
j = 1, 2, . . . , p; α = 1 or large
2.2 Find cumulative probability ranges associated with different paths based on probabilities
given in step 2.1
Step 3: Evaluation of different paths
3.1 Find the specific path chosen by ant r by generating N random numbers in the range of (0,
1), one for each ant and determining the cumulative range previously found in step 2.2 for any
ant r for a particular variable i
3.2 Repeat step 3.1 for all design variable i = 1, 2, 3, . . . , n
3.3 Evaluate the objective function corresponding to complete paths ∀ ants
(r = 1, 2, 3, . . . , N ) and design variables (i = 1, 2, 3, . . . , n); ξr = ξ(r, n)
3.4 Determine the best and worst paths among N paths chosen by different ants
ξbest = Min.ξ(r, n) ∀k = 1, 2, 3, . . . , N (8)
ξwor st = Max.ξ(r, n) ∀k = 1, 2, 3, . . . , N
Step 4: Obtain best path upon convergence of algorithm
4.1 Test for convergence of the whole ant colony optimization process. It is assumed that the
process has converged if N ants take the best path
4.2 If convergence is not achieved, assume all ants return home and start again in search of
food. The new iteration number is set l = l + 1 and updating rule is applied to the pheromones
on different path segments as
N
ψil j = ψiold
j + ψi(rj ) (9)
r =1
j = (1 − ρ)ψi j
where ψiold l

4.3 With new value of ψi j , go to step 2 and repeat the whole algorithm, until the process
l

converges

of the problem. Ant colony optimization offers one such technique to help resolve
engineering-cum-managerial optimization problems in pavement asset management
system.
Pavement Infrastructure Asset Management Using Clustering … 225

3.2 Application in Pavement Management Systems

Serdel and Serin proposed a metaheuristic method based on an ant colony opti-
mization algorithm for planning the economic evaluation and maintenance works on
pavements considering the rehabilitation, routine, and major maintenance options for
flexible pavements [18] governed by the availability of funds, historical accounts, and
network or project level considerations based on organizational and political factors.
In the proposed model, six constraints such as requirement of production, availabil-
ity of labor, budget, equipment, material, and rehabilitation have been considered
using the methodological framework of ant colony optimization algorithm imple-
mented through visual basic software. The authors were able to compute optimum
maximum work load for maintenance work priority wise and found that it decreased
with increasing the number of ant in the colony. Their work has been based on the
formulations proposed by Fwa et al. [19], which provide an integration optimiza-
tion programming model tool to prioritize routine maintenance works in highway
pavements. Further, Fwa et al. [20] suggested that pavement maintenance is a multi-
objective optimization problem as such agencies need to employ advanced network
level optimization techniques comprising of dynamic route searching features and
inherent capability to handle multiple solutions must be adopted. They suggested
the use of a genetic algorithm-based optimization technique and illustrated its use
through a numerical example performed for two- and three-objective optimization,
respectively. Ant colony optimization is also a similar optimization technique which
can serve as a tool to handle multi-objective optimization problems encountered in
pavement management system. Chikezie et al. [21] also suggested the use of multi-
objective optimization in pavement maintenance and rehabilitation works. Gao et
al. [22] has illustrated the network level selection feature of ant colony optimiza-
tion technique and proposed a model based on the methodological framework of
ant colony optimization. Sivagaminathan and Ramakrishnan [23] have proposed a
hybrid approach for optimum network level selection using neural networks and ant
colony optimization. It helps in the selection of input variables which help in the
prediction of an optimum route with an acceptable level of accuracy. Christodoulou
[24] has presented a methodology for scheduling resource-constrained construction
projects using ant colony optimization technique. The author has used the ACO
process methodology and efficiently computed optimum solution in less than 50
iterations. Further, the author has compared the ACO technique with other tradi-
tional and modern algorithms and concluded that ACO is relatively simple than a
genetic algorithm and particle swarm optimization, and representatively similar to
traditional CPM but more efficient and robust. Fileccia Scimemi [25] have used ant
colony optimization technique coupled with Levy perturbation for back-calculating
the elastic moduli and thickness of multiple layers of airport pavements using surface
deflection data obtained from nondestructive testing. They showed that the ant colony
optimization technique can offer a viable technique to obtain the above parameters
correctly while considering the inherent variation of inputs and calculated parame-
ters. The ACO algorithm removes the anomalies of local optima through its global
226 S. Gulzar and H. Ali

searching rule and gives accurate results over other optimization methods. Zhitao et
al. [26] have given an algorithmic framework for scheduling of construction works
based on ant colony optimization. These construction works involve building, as
well as pavement infrastructure, and may be optimized well using the ACO process.
In road construction, there is an additional delay caused by the ongoing traffic due
to the construction activities leading to a loss in speed and increase in travel time
which result in huge economic losses, as such the optimum scheduling is even more
essential in such case and ACO offers a viable method to achieve this objective. The
authors have integrated the inputs based on expert knowledge into the metaheuristic
ACO algorithm called as CoANT to achieve an optimum scheduling scheme and
have implemented it on realistic transportation networks. CoANT is seen to work
faster, better, and is able to produce competitive schedules which aid the agencies in
decision-making. In our previous work [2], the applicability of ant colony optimiza-
tion in pavement management system was showcased and the ACO algorithm was
used to explore the resource allocation and time scheduling of pavement construction
at a project level. Further, the use of ACO algorithm in clustering analysis for project
allocation and efficient bidding process in pavement construction projects was also
deliberated.
Julia and Meyer [27] reviewed the two network level optimization clustering tech-
niques based on ant colony optimization and concluded that the algorithms based
on ant colony optimization are very effective for multi-objective, decentralized, and
self-organizing problems and can offer a robust solution alternative to compute the
optimum route, whether based on priority, budgetary constraints or resources con-
straint. It implicitly implies that, optimization problems in pavement management
systems can be effectively handled using ant colony algorithms and may prove ben-
eficial than any other conventional optimization technique. Based on the use of ant
colony optimization in pavement management system, a general framework as shown
in is proposed. This framework uses the existing ant colony optimization algorithms
and provides a way to use it in handling pavement construction management prob-
lems at a project and network level.

4 Ant Colony Optimization-Based Clustering Analysis in

PMS

As described in Sect. 2, the pavement management system incorporates multiple

processes and activities. All such activities revolve around the construction of a road
infrastructure. A road construction project involves a range of activities and many a
times, there are different subcontractors for different activities. So, a project given
out by an agency is let by a major contracting company, which in turn, sublets it
to different subcontractors to handle a different set of activities. As stated before,
both the agency and the contractor want to get the project done at minimum cost
and duration without compromising the quality of the project. The whole process
Pavement Infrastructure Asset Management Using Clustering … 227

Fig. 6 Ant colony optimization framework in pavement management system

can be treated as a multi-objective optimization problem and we will now proceed

to formulate the objective functions and constraints of such a problem. In order to
reach an optimal or feasible solution, an ant colony optimization-based clustering
algorithm is used, and it is shown that such a protocol proves to be very effective in
such a scenario [28] (Fig. 6).
In an asphalt pavement construction project, the consideration of Indirect cost (IC )
per time period over the entire duration of the project (D p ) plays an important role
in letting the project to contractors or subcontractors. Every road project is assigned
a due date (Dd ). Depending on the actual time of completion, the contractor may get
incentive, say α cost for the time saved as a reward or may be charged a penalty, say δ
for the overtime taken to finish the project. For any activity l, let us suppose there are a
total of Sl available subcontractors, such that each subcontractor m bids a cost clm with
a proposed duration of dlm with an expected quality of qlm . In order to complete all
the project activities, we need to select a near-optimal group of subcontractors based
so as to minimize the total duration (D p ) and cost (C p ) of the project, maximize the
quality of the project (Q pτ ). The objective function for minimizing the total duration
of the project can be written as below


n
Objective 1 : min D p = max dlm klm (10)
l∈W p m=1
228 S. Gulzar and H. Ali

where klm is a binary variable (0, 1) for each activity l, and W p is the activity sequence
of pth path.
Similarly, the objective function for minimizing the total cost of the project for
all the activities of the project (N ), can be written as below


N 
Sl
Objective 2 : min C P = clm klm + IC D p + δ(D p − Dd ) + α(Dd − D p )
l=1 m=1
(11)
The final objective function is to maximize the quality of the project. It should
be noted that there is a minimum threshold, Q min for all the possible alternatives of
an activity which is measured against the average quality, Q avg for all the possible
subcontracting alternatives eligible for the project. The deviation between the two,
say τ sets the desirability of each parameter by putting the relative weights on each
and determines the overall quality of the project. So, the objective function may be
written as below

Objective 3 : max Q pτ = τ Q min + (1 − τ )Q avg (12)

The constraints are set in the quality and also on the way the contracting process is
done. The constraints on the quality of the project are as below

Q min = min{qlm : klm = 1}, 1 ≤ l ≤ N , 1 ≤ m ≤ Sl (13)

Q min ≤ {qlm × klm + H (1 − klm )}, 1 ≤ l ≤ N , 1 ≤ m ≤ Sl (14)


N 
Sl
qlm klm (15)
Q avg = l=1 m=1
N
, 1 ≤ l ≤ N , 1 ≤ m ≤ Sl

It should be noted that H is an arbitrarily large number and Q min can never exceed
H . This is done to keep the constraint nonbinding. The indicators that are used to
check the quality of an asphalt pavement include total air voids in the mix, asphalt
binder content, in-situ air void content, ride quality, etc. Similarly, the constraint on
the time to ensure that project duration is within the due date can be written as

D p ≤ Dd (16)

Also, to ensure that each activity is done successively in the project network, a
constraint on the starting time of a scheduled activity, ti is applied as given below
Pavement Infrastructure Asset Management Using Clustering … 229

Table 2 An illustrative alternative matrix showing different combinations of cost, duration, and
quality of different project activities proposed by different project contractors
Project Cost (C p )—Duration (D p )—Quality (Q pτ ) alternatives
activity
Alternative 1 Alternative 2 … Alternative 10
Cp Dp Q pτ Cp Dp Q pτ Cp Dp Q pτ
1 2000 10 100 2500 12 99 2000 13 98
2 4000 12 98 4500 14 100 35000 13 97
3 6000 12 95 5000 13 97 4500 10 100
4 4500 9 97 4200 11 98 6000 15 100
5 3000 8 99 6000 9 95 6000 12 100
6 5000 12 95 3500 8 98 3750 11 99
7 8000 13 100 9000 15 96 8500 14 98
8 800 5 95 550 6 96 750 8 96
9 500 3 97 950 7 98 900 9 96
10 600 7 98 320 8 100 450 5 97
11 25000 25 99 29000 23 100 35000 29 97
12 30000 29 100 35000 35 95 29500 32 98
13 30000 24 100 20000 20 98 25000 28 99
14 2000 11 98 2500 9 99 2410 10 97
15 9000 9 97 7000 11 99 6000 15 96
16 6000 10 95 3000 12 98 5550 13 95
17 3000 9 96 4000 8 97 3500 9 98
18 4000 10 95 5000 7 96 3590 11 99
19 3000 13 98 3000 10 95 2500 15 99
20 5000 8 99 4580 5 100 7000 3 100


Sl
t P ≥ tl + qlm klm , 1 ≤ l ≤ N (17)
m=1

To ensure that each activity is assigned only one subcontracting alternative, the
constraint can be written as


Sl
klm = 1, 1 ≤ l ≤ N (18)
m=1

Now, in order to solve this multi-objective optimization problem, the steps high-
lighted in Table 3 are systematically executed and the flow of the algorithm is as per
Fig. 7. It should be noted that clustering analysis is done along with ACO by consider-
ing the groups of subcontractors based on the bid time, cost and quality of the project
activity. A list of activities grouped by the available subcontractors is evaluated and
230 S. Gulzar and H. Ali

a near-optimal cluster is selected. It should be noted that there is always a trade-off

involved in multi optimization problems such as this and achieving an exact solution
is not always feasible. A brief illustration on synthetic data for a flexible pavement
construction project based on the data from Feng et al. [29] is given in Table 2.
Consider a flexible pavement construction project involving a set of 20 (say)
activities. Let us suppose the bidding process receives quotations from 10 contractors
proposing a different combination of time, cost, and quality for each activity. The
options matrix will look similar to Table 2. It shows the different combinations of
cost, duration, and quality of different project activities proposed by different project
contractors. The cost is in dollars, the duration is in days, and the quality is rated on
a scale of 1 to 100. Other parameters such as the proposed date of completion is 220
days, reward and penalty costs are chosen to be $250 and $150 per day, etc. With
20 activities in the project and 10 contractors, there are over 4 billion combinations
for scheduling and executing the project. In order to obtain the optimal combination
of contracting alternatives such that the project time and cost are minimized, while
the quality is maximized, the ant colony optimization algorithm searches the sub-
space and with each successive iteration leaves out the nonoptimal solutions. In
order to initialize the implementation algorithm, several trials were conducted and
finally a suitable value of initial parameters is obtained. After running the algorithm
for several thousand iterations, a trade-off matrix is obtained and a near-optimal
solution is chosen. An example solutions trade-off table is shown in Table 3.
It can be seen that there are different solutions with different optimal conditions.
For example, solution 26 offers lowest cost but project time and quality are not
highest. Similarly, solution 24 offers lowest project duration, but is not necessarily
cheapest or of best quality. Also, solution 25 offers the best quality but is quite costly
and takes a longer time. So, it is essential to select a solution by trading-off on the
selection criteria and selecting a near-optimal solution. The performance for all the
seven sample solutions is compared and contrasts are drawn as shown in Figs. 7,
8, and 9 for cost, time or duration, and quality of the project. Finally, the trade-off
plot for all three selection criteria is shown in Fig. 10. The trade-off analysis is done
for each pair of selection criteria and the set of project alternatives that give the
near-optimal solution is selected. Further, in order to evaluate the efficiency of the
ant colony-based algorithm, a performance analysis is conducted using the metrics
such as degree and speed of convergence, sparsity, and number of local optima, etc.
The variance of solutions, the computation time, and the root mean square distance
are often used as well.
The activities that result in a similar selection criterion are clustered together to
achieve higher efficiency and quicker convergence at sufficiently low computation
time. Daniel [30] reports that the solution obtained using ant colony optimization
algorithm is highly dependent on the correct implementation of the pheromone met-
ric. For small changes in the pheromone metric, there were huge changes in the
optimal solution, highlighting the sensitivity of ACO to pheromone trail. As dis-
cussed by Christodoulou [24], the algorithm provides an efficient methodology in
constructing solutions in network level optimization problems. There is a lot of scope
for the improvement in ACO algorithms by incorporating forward coupling, as well
Pavement Infrastructure Asset Management Using Clustering … 231

Table 3 Solution trade-off table

Solution Project activities Selection criteria
Cost ($) Time (days) Quality (%)
21 [1, 1, 2, 1, 3, 1, 5, 187900 275 83.526
4, 1, 8, 6, 1, 2, 1,
6, 5, 4, 3, 8, 9]
22 [2, 1, 5, 4, 2, 6, 5, 190065 265 85.996
1, 2, 3, 9, 8, 2, 4,
4, 5, 1, 2, 4, 5]
23 [3, 3, 3, 1, 1, 1, 1, 191565 289 84.153
5, 1, 6, 4, 1, 8, 1,
6, 2, 9, 1, 5, 2]
24 [2, 1, 5, 4, 1, 8, 3, 220513 255 87.225
2, 9, 1, 5, 2, 8, 4,
1, 5, 2, 1, 5, 1]
25 [2, 3, 2, 2, 2, 3, 1, 251200 295 89.551
2, 5, 1, 4, 5, 1, 6,
5, 2, 8, 7, 4, 5]
26 [5, 1, 5, 4, 2, 5, 4, 151559 288 81.322
5, 6, 2, 3, 1, 8, 1,
1, 1, 2, 6, 5, 1]
27 [8, 1, 9, 7, 9, 1, 9, 187900 277 79.854
8, 1, 6, 5, 9, 1, 1,
5, 6, 1, 1, 1, 5]

Fig. 7 Cost performance of

the sample solutions
232 S. Gulzar and H. Ali

Fig. 8 Time performance of

the sample solutions

Fig. 9 Quality performance

of the sample solutions
Pavement Infrastructure Asset Management Using Clustering … 233

Fig. 10 Trade-off plot of selection criteria

as backward indentation mechanism at the project, as well as network levels. There

is also prospective scope in design optimization as well as rehabilitation prioriti-
zation problems and scheduling of maintenance activities. The overall computation
efficiency and varied applicability of the ACO algorithms for handling complex prob-
lems comprising of huge number of activities, resources, and manpower needs to be
evaluated using numerical simulations and programming tools.

5 Conclusion

Pavement management system involves multiple optimization and clustering prob-

lems both at project level, as well as at network level. In order to improve efficiency
in pavement construction, management, and maintenance, it is important to use evo-
lutionary algorithms such as ant colony optimization to reduce losses as showcased
in this chapter. An introduction to pavement infrastructure was given, followed by
the description of ant colony optimization technique. The steps involved in the imple-
mentation of ACO in pavement management system were highlighted. Ant colony
optimization proves to offer the advantage of having a robust search algorithm at
a global level with good accuracy. The objective functions for an asphalt pavement
construction project along with the constraints were defined in a multi-objective
optimization problem and the steps to use of ACO to obtain near-optimal solution
based on clustering analysis involving project duration, cost, and quality of each
activity involved were presented. The implementation algorithm presented in this
chapter can be used for different types of pavement management scenarios and it
234 S. Gulzar and H. Ali

is hoped that a full case study showing the complete application of the ant colony
optimization-based clustering analysis in PMS will be presented in future.

References

1. Vajjarapu, Harsha, Ashish Verma, and Saqib Gulzar. 2019. Adaptation policy framework for
climate change impacts on transportation sector in developing countries. Transportation in
Developing Economies 5 (1): 3.
2. Gulzar, Saqib, Hasnain Ali, and Chetan Doddamani. 2017. Ant colony optimization in pave-
ment asset management. In Proceedings of ASCE India conference 2017. Indian Institute of
Technology Delhi.
3. Gulzar, Saqib, Hasnain Ali, and Chetan Doddamani. 2017. A conceptual framework for intro-
ducing ‘mobility as a service’ in India: Opportunities & challenges. In Proceedings of ASCE
India conference 2017. Indian Institute of Technology Delhi.
4. Gulzar, Saqib, and Shane Underwood. 2019. Use of polymer nanocomposites in asphalt binder
modification. In Advanced functional textiles and polymers: Fabrication, processing and appli-
cations, 405.
5. Gulzar, Saqib, and Hazratullah Paktin. 2017. Scope of using nanomaterials in pavement engi-
neeering. In Proceedings of nano India 2017. Indian Institute of Technology Delhi.
6. Ali, Hasnain, Yash Guleria, Sameer Alam, and Michael Schultz. 2019. A passenger-centric
model for reducing missed connections at low cost airports with gates reassignment. IEEE
Access 7: 179429–179444.
7. Ali, Hasnain, Yash Guleria, Sameer Alam, Vu N. Duong, and Michael Schultz. 2004. Impact
of stochastic delays, turnaround time and connection time on missed connections at low cost
airports. In Proceedings of the 13th USA/Europe air traffic management. R&D Seminar.
8. Ali, Hasnain, Raphael Delair, Duc-Thinh Pham, Sameer Alam, and Michael Schultz. Dynamic
hot spot prediction by learning spatial-temporal utilization of taxiway intersections. In Interna-
tional conference on artificial intelligence and data analytics in air transportation (AIDA-AT
2020) (in press).
9. Haas, Ralph, and W. Ronald Hudson. 2015. Pavement asset management. Wiley.
10. Dorigo, Marco. 1992. Optimization, learning and natural algorithms. PhD thesis, Politecnico
di Milano.
11. Heidari, Ali Asghar, Hossam Faris, Seyedali Mirjalili, Ibrahim Aljarah, and Majdi Mafarja.
2020. Ant lion optimizer: Theory, literature review, and application in multi-layer perceptron
neural networks. In Nature-inspired optimizers, 23–46. Springer.
12. Mirjalili, Seyedali, Jin Song Dong, and Andrew Lewis. 2020. Ant colony optimizer: Theory,
literature review, and application in auv path planning. In Nature-inspired optimizers, 7–21.
Springer.
13. Mirjalili, Seyedali. 2015. The ant lion optimizer. Advances in Engineering Software 83: 80–98.
14. Mirjalili, Seyedali, Pradeep Jangir, and Shahrzad Saremi. 2017. Multi-objective ant lion opti-
mizer: A multi-objective optimization algorithm for solving engineering problems. Applied
Intelligence 46 (1): 79–95.
15. Mafarja, Majdi, Derar Eleyan, Salwani Abdullah, and Seyedali Mirjalili. 2017. S-shaped vs.
v-shaped transfer functions for ant lion optimization algorithm in feature selection problem.
In Proceedings of the international conference on future networks and distributed systems, 21.
ACM.
16. Rao, Singiresu S. 2019. Engineering optimization: Theory and practice. Wiley.
17. Ali, Hasnain, and Arpan Kumar Kar. 2018. Discriminant analysis using ant colony optimization-
an intra-algorithm exploration. Procedia Computer Science 132: 880–889.
18. Terzi, Serdal, and Sercan Serin. 2014. Planning maintenance works on pavements through ant
colony optimization. Neural Computing and Applications 25 (1): 143–153.
Pavement Infrastructure Asset Management Using Clustering … 235

19. Fwa, Tien Fang, Kumares C. Sinha, and John D.N. Riverson. 1988. Highway routine mainte-
nance programming at network level. Journal of Transportation Engineering 114 (5): 539–554.
20. Fwa, T.F., W.T. Chan, and K.Z. Hoque. 2000. Multiobjective optimization for pavement main-
tenance programming. Journal of Transportation Engineering 126 (5): 367–374.
21. Chikezie, Clarkson Uka, Adekunle Taiwo Olowosulu, and Olugbenga Samuel Abejide. 2013.
Multiobjective optimization for pavement maintenance and rehabilitation programming using
genetic algorithms. Archives of Applied Science Research 5 (4): 76–83.
22. Gao, Hai-Hua, Hui-Hua Yang, and Xing-Yu Wang. 2005. Ant colony optimization based net-
work intrusion feature selection and detection. In 2005 international conference on machine
learning and cybernetics, vol. 6, 3871–3875. IEEE.
23. Sivagaminathan, Rahul Karthik, and Sreeram Ramakrishnan. 2007. A hybrid approach for
feature subset selection using neural networks and ant colony optimization. Expert Systems
with Applications 33 (1): 49–60.
24. Christodoulou, Symeon. 2009. Scheduling resource-constrained projects with ant colony opti-
mization artificial agents. Journal of Computing in Civil Engineering 24 (1): 45–55.
25. Fileccia Scimemi, G., T. Turetta, and C. Celauro. 2016. Backcalculation of airport pavement
moduli and thickness using the lévy ant colony optimization algorithm. Construction and
Building Materials 119: 288–295.
26. Xiong, Zhitao, David Rey, Vinayak V. Dixit, and S. Travis Waller. 2014. An algorithmic frame-
work for the scheduling of construction projects based on ant colony optimization and expert
knowledge. In 17th international IEEE conference on intelligent transportation systems (ITSC),
2446–2452. IEEE.
27. Handl, Julia, and Bernd Meyer. 2007. Ant-based and swarm-based clustering. Swarm Intelli-
gence 1 (2): 95–113.
28. Mungle, Santosh, Lyes Benyoucef, Young-Jun Son, and M.K. Tiwari. 2013. A fuzzy clustering-
based genetic algorithm approach for time-cost-quality trade-off problems: A case study of
highway construction project. Engineering Applications of Artificial Intelligence 26 (8): 1953–
1966.
29. Feng, Chung-Wei, Liang Liu, and Scott A. Burns. 1997. Using genetic algorithms to solve
construction time-cost trade-off problems. Journal of Computing in Civil Engineering 11 (3):
184–189.
30. Angus, D. 2005, January. Solving a unique shortest path problem using ant colony optimisation.
Communicated by T. Baeck, 1–26.
A Classification Approach Based on
Evolutionary Clustering and Its
Application for Ransomware Detection

Raneem Qaddoura, Ibrahim Aljarah, Hossam Faris, and Iman Almomani

Abstract Ransomware analysis and detection has been recently applied using super-
vised and unsupervised machine learning approaches. Combining both approaches
simplifies the complexity of the data and builds an expert classifier for each pre-
dicted cluster. In this paper, a hybrid approach for detecting ransomware is proposed.
It combines evolutionary clustering approach using Grey Wolf Optimizer (GW O)
with an ensemble of Support Vector Machine (SV M) classification on a dataset col-
lected from Android benign and ransomware applications. Experiments are applied
to identify the best number of clusters for the proposed approach on the selected
dataset and to compare the results with those obtained from applying the SV M sin-
gle classification approach. Results show that applying SV M-GW Ok=2 approach
outperforms the corresponding SV M classifier in terms of accuracy.

Keywords Clustering-based classification · Grey Wolf Optimizer · Support

Vector Machine classifier · Ransomware detection · Evolutionary computation ·
EvoCluster

R. Qaddoura
Information Technology, Philadelphia University, Amman, Jordan
e-mail: [email protected]
I. Aljarah (B) · H. Faris · I. Almomani
King Abdullah II School for Information Technology, The University of Jordan, Amman, Jordan
e-mail: [email protected]
H. Faris
e-mail: [email protected]
I. Almomani
e-mail: [email protected]; [email protected]
I. Almomani
Security Engineering Lab, Prince Sultan University, Riyadh, Saudi Arabia

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer 237
Nature Singapore Pte Ltd. 2021
I. Aljarah et al. (eds.), Evolutionary Data Clustering: Algorithms and Applications, Algo-
rithms for Intelligent Systems,
https://doi.org/10.1007/978-981-33-4191-3_11
238 R. Qaddoura et al.

1 Introduction

Data has become an important asset with the emergence of cloud and IoT services
and their consequences of big data acquisition and storage [5, 31]. This draws every-
one’s attention to the importance of data protection and security. Thus, ransomware
detection approaches have been investigated to protect data from external threats.
Machine learning approaches have recently been applied on ransomware detec-
tion. These approaches fall under two main types: supervised learning and unsuper-
vised learning[10, 23, 24]. Supervised learning is a classification task of constructing
a training model by learning from previously labeled observations for future data pre-
dictions [2]. Unsupervised learning is a clustering task which explores the structure
of unlabeled data and explores useful patterns [4, 8, 9, 11–13, 45, 49] by group-
ing data according to their similarities and dissimilarities [26, 47, 48]. Evolutionary
clustering is another form of unsupervised learning which started to gain attention by
many researchers [25, 44, 47] due to the improved performance over the traditional
unsupervised clustering algorithms.
In the area of ransomware detection, both supervised and unsupervised machine
learning approaches are used to generate high performance detection of threats. Com-
bining both approaches while considering evolutionary clustering as an unsupervised
learning approach is an attractive direction to investigate.
In this work, a hybrid classification approach for detecting ransomware in Android
applications is proposed. The hybrid approach consists mainly of two stages: in the
first stage, an evolutionary clustering approach using Grey Wolf Optimizer (GWO) is
applied to segment the training dataset into a number of sub training datasets. Then
in the second stage, an ensemble of support vector machine classifiers (SV M) is
trained, where each SV M in the ensemble is trained based on a different subtraining
dataset (i.e., cluster). Basically, the idea is to simplify the complexity of the data and
build an expert classifier for each cluster of it. Therefore, the aim of the approach is to
boost the detection accuracy in comparison with the conventional single classification
approach.
The remainder of the paper is organized as follows: The next section shows the
related work in ransomware detection. Section 3 presents GW O optimizer. Section 4
describes in detail the proposed algorithm. Section 5 discusses the data preparation
and description. Section 6 presents the experiments conducted to evaluate and verify
the efficiency of the algorithm. Section 7 concludes the main findings of this work.

2 Related Work

Ransomware is a malware which forbids users from accessing their files unless
they pay attackers to gain access. Automatic analysis and detection of ransomware
has been recently applied using supervised and unsupervised machine learning
approaches.
A Classification Approach Based on Evolutionary … 239

Supervised classification approaches can be found extensively for ransomware

detection. The work of [28] considers creating machine learning models using
Random Forest, Support Vector Machine, and K-Nearest Neighbors approaches,
and evaluating their performance in detecting ransomware. Two frameworks com-
bining Reverse engineering, Natural Language Processing, and Machine Learning
approaches are proposed in the work of [42, 43]. Authors of [51] applied various
experiments related to network parameters and structures of the multi-layer percep-
tron to detect and classify ransomware threats. The work of [31] includes classifying
infected files based on file entropy analysis for ransomware detection in backup sys-
tems using K-nearest neighbor, linear model, decision tree, decision tree ensemble,
kernel trick, and neural network approaches. The paper [6] introduces NetConverse,
which is a machine learning evaluation study to detect Windows ransomware network
traffic. Other studies for detecting ransomware can be found in [18, 20, 38].
On the other hand, fewer works can be found for ransomware detection using
unsupervised clustering machine learning approaches compared to those found for
supervised approaches. Authors of [1] proposed a framework considering a modified
version of sequential K-means clustering algorithm for detecting cloud insiders or
malware perform malicious tasks on cloud customers’ virtual machines. Another
study [16] analyzed ransomware threats using the Cyber-genome clustering method,
which are then combined with anti-malware product logs to identify distinct ran-
somware families. Other works [35–37] considers fuzzy hashing and a clustering
method to detect ransomware threats.
However, another approach of clustering-based classification has proven in many
applications that it is efficient and performs higher quality of results compared to
those obtained by single approach of applying either classification or clustering.
These applications include injury severity prediction of road traffic accidents [50],
Superpixel Segmentation for PolSAR Image Classification [29], color histopatholog-
ical image segmentation [19], hyperspectral image classification [33], indoor local-
ization [3], breast tumor detection [52], Medical Relation detection[30], and text
classification [39].
Thus, the clustering-based classification approach can be examined to detect ran-
somware threats, which to the best of our knowledge, could not be found in the
literature. In this paper, we propose a clustering-based classification approach for
ransomware detection using GW O clustering algorithm and SV M classification
algorithm.

3 Grey Wolf Optimizer (GW O)

GW O is a nature-inspired optimizer inspired by the social activities performed by

grey wolfs, which are the social hierarchy of leadership and the group hunting. The
former activity considers a leader wolf (α), supporting wolves (β), and remaining
wolves (δ and ω) [22, 34], and the later activity considers tracking, chasing, pursu-
240 R. Qaddoura et al.

ing, encircling, harassing, and attacking the prey [34]. These activities are reflected
mathematically in GW O algorithm.
The first activity of leadership maps the best solution to the α wolf, the other best
solutions are mapped to the β and δ wolves. The remaining solutions are mapped to
the ω wolves.
The second activity of hunting is reflected by predicting the prey position when
converging to a better solution. This follows four processes: encircling, hunting,
attacking, and searching. The encircling process is performed by two vectors calcu-
lated by Eqs. 1 and 2, and two coefficient values calculated by Eqs. 3 and 4. The
hunting process is performed by a vector calculated by Eq. 5 and three values cal-
culated by Eqs. 6, 7, and 8. The attacking process reflects exploiting the space by
moving towards the prey and attacking it. The searching process reflects exploring
the space by diverging from the prey to avoid local optima solutions.

D = |C · X p (t) − X (t)| (1)

X (t + 1) = X (t) − A.D (2)

C = 2 · r2 (3)

A = 2a · r1 − a (4)

X (t + 1) = (X 1 + X 2 + X 3 )/3 (5)

X 1 = X α − A1 · (|C1 · X α − X |) (6)

X 2 = X β − A2 · (|C2 · X β − X |) (7)

X 3 = X δ − A3 · (|C3 · X δ − X |) (8)

where t is the iteration number, r1 and r2 are random vectors of values between 0
and 1, a is a vector which linearly decreases from 2 to 0, X is the position of the
wolf, X p is the position of the prey, X α is the position vector of the alpha wolf, X β
is the position vector of the beta wolf, X δ is the position vectors of the delta wolves.
A1 & C1 , A2 & C2 , and A3 & C3 are coefficient vectors.

4 Proposed Approach

This section discusses the proposed approach, which is followed when running the
experiments. The proposed approach is presented in Figure 1. First, the dataset is
split into training and testing instances with a ratio of 80%:20%, respectively. The
clustering process is then applied to the training instances using GW O algorithm.
This generates k clusters in which each cluster has similar instances. For each cluster,
A Classification Approach Based on Evolutionary … 241

Data Set

80% 20%

Training
Testing

Measure
distances Allocate instances
GWO to clusters

Centroid 1
Centroid 2
Centroid k

Testing
Testing
Cluster 1
Testing
Cluster 2
Cluster k
Clustering

Predict

Training Generate Models Model 1

Training
Cluster 1 Model 2
using SVM Model k
Training
Cluster 2
Cluster k

Generate

Classification
Results

Fig. 1 Proposed approach

classification is performed by fitting a model to the training instances of the cluster

using SV M algorithm. Then, the testing instances are allocated to clusters accord-
ing to the euclidean distance between the instances and the centroids, by which an
instance is allocated to a cluster having the minimal distance between the instance
and the corresponding centroid. Finally, the labels for the instances of each cluster
are predicted using the corresponding model, which forms the classification results.
The remainder of this section discusses the two main processes of the proposed
approach: the clustering process of GW O algorithm using TWCV fitness function,
which tests the convergence of GW O, and the classification process of the clustered
instances using SV M.
242 R. Qaddoura et al.

4.1 Clustering with GW O

Clustering with GW O is performed by enhancing the centroids’ selection of the

clustering task [46]. Random selection of centroids for each individual is performed at
the first iteration of the algorithm, which is enhanced in later iterations by minimizing
the value of the Total Within Cluster Variance (TWCV) [41] between the centroid
and the cluster instances. The best enhanced individual at the last iteration is returned
as a solution.
Total Within Cluster Variance (TWCV) [21, 32, 41] evaluates the convergence of
evolutionary algorithms. As the name indicates, it calculates the total within cluster
variance. In the experiments, GW O tries to minimize the value of TWCV to enhance
the quality of the clustering results. TWCV is calculated by the following equation:

 1   2
N 
F K F
T WCV = pn2 f − pk f (9)
n=1 f =1 k=1
| pk | f =1

where F is the number of features, pn f is feature f of the data point n, pk f is feature

f of the data point k, and | pk | is the number of data points in cluster k.
The solution obtained consists of the training instances grouped to k different
training clusters. The centroids for each cluster are used to map each testing instance
to the closest cluster, forming k testing clusters.

4.2 Classification with SV M

For the training instances of each cluster generated from the previous process, a
linear kernel SV M classification is performed, which creates a fitted model. SV M
works by creating a hyperplane in space [40]. The algorithm separates the instances
of each cluster from others by finding the hyperplane with the largest distance from
the training instances [27, 40].
The fitted model of the training instances of each cluster is used to predict the labels
of the corresponding testing instances of the cluster. Recall that the testing instances
of the same cluster are recognized by the previous clustering process. Performing
this step for each cluster results in predicting the labels for all testing instances.
The power of performing the clustering process as a preparatory process to the
classification process is to find a specialized model for the instances in a training
cluster, which takes into consideration similar characteristics of the instances. This
helps the classification algorithm generate enhanced classification results.
A Classification Approach Based on Evolutionary … 243

5 Dataset Description

Selection of Android benign and ransomware apps: The group of Android appli-
cations that were collected in [14, 15] has been used to build the dataset under
study. The dataset was constructed based on 1000 Android application samples. This
dataset is balanced and consists of 500 benign apps in addition to 500 ransomware
apps. The benign apps were downloaded from Google play store.1 These apps were
chosen from different categories which are free and highly downloaded by users.
Ransomware samples were collected from different sources including Koodous,2
RansomProper project[17] and VirusTotal3 to ensure gathering a unique set of ran-
somware that belong to different families with locking, encryption, and threatening
messages features.
Reverse Engineering Process: Android apps are usually available and distributed
in the format of APK files. APK is a compressed file that holds the overall app’s
source code that was originally written in java. To analyze the app, this APK needs
to be decompiled to retrieve important files like “.xml” and “.dex” files. There are
many tools available and can be used to decompile the APK through a process called
“Reverse Engineering” APKtool4 was used by this research to decompile the APK
files into “.smali” files.
Code parsing and features extraction: The small files were then parsed to extract
a finite set of features. The selected features were the API packages and permissions
belong to Oreo Android release 27 (API 27). From this release, 199 API packages
and 115 permissions were considered; with an overall 314 features in addition to the
identifier of the app such as the hash code and also the category of the app whether
benign or ransomware. The features extraction process counted the frequencies of
API calls and permissions in each Android app to generate the following joint vector
V for each one of them:

V = {a1 , a2 , . . . an } ∪ { p1 , p2 , . . . pm } (10)

where ai is the number of occurrences of the ith API call in the application, p j is the
number of occurrences of the jth permission in the application, n is the total number
of API packages and m is the total number of permissions.
Data Cleansing: As part of data cleansing, duplicate vectors were removed to ensure
unique records and produce a quality dataset. The duplicate vectors could have
resulted from identical Android apps using different identities (different hash codes).

1 https://play.google.com/store.
2 https://koodous.com/apks.
3 https://www.virustotal.com.
4 https://ibotpeaches.github.io/Apktool.
244 R. Qaddoura et al.

6 Experimental Results

The experiments are conducted using two tools: EvoCluster framework [46] for the
clustering process with GW O, which is a recent framework for clustering datasets
using nature-inspired algorithms, by which GW O is one of the algorithms of the
framework. The popular scikit learn tool [40] is used for the classification process
with linear SV M.
The experiments are evaluated by measuring the accuracy value, which reflects the
quality of the classification process. It is the ratio of the number of the true predicted
values (TP + TN) to the number of all predicted values (TP + TN + FP + FN), which
is observed by Eq. 11.

Accuracy = (T P + T N )/(T P + T N + F P + F N ) (11)

where TP is the true positive, FP is the false positive, TN is the true negative, and
FN is the False negative.
The proposed approach is evaluated by running the experiments 30 times for
each value of k, which are 2, 3, and 4. Then, comparing the results of the average
value of the accuracy measure for each value of k against the average accuracy
value of applying a single SV M classification. Table 1 shows the performance of
SV M-GW O for different values of k against single SV M classification approach.
As observed from Table 1, SV M-GW O with k value of 2 outperforms itself with
other values of k, and outperforms the corresponding SV M algorithm. It has the
largest value of the average accuracy compared to the others on the selected dataset.
It also has a recognizable highest value for the best and worst results.
The reason behind the high value of accuracy for the algorithm having k = 2
compared with the other values of k is due to the nature of the dataset instances where
an instance is a ransomware or benign attack. This confirms that the classification
of the dataset should be considered binary. In addition, the high value of accuracy
for the SV M-GW O with k = 2 compared to SV M is due to the powerful effect of
clustering as a preparatory step for the classification process.
Figure 2 shows the interquartile range, average accuracy value, best accuracy
values, and worst accuracy value [7] for 30 runs of SV M, SV M-GW Ok=2 , SV M-
GW Ok=3 , and SV M-GW Ok=4 . It is observed from the figure that the SV M-GW O

Table 1 Results for using SV M with GW O with different values of k using the accuracy evaluation
measure (Highest values of the average, standard deviation, best, and worst are marked with bold)
Average Std Best Worst
SV M 0.9517 0.0024 0.955 0.95
SV M-GW Ok=2 0.9803 0.0066 0.99 0.96
SV M-GW Ok=3 0.8608 0.0831 0.94 0.605
SV M-GW Ok=4 0.7063 0.0836 0.89 0.6
A Classification Approach Based on Evolutionary … 245

Fig. 2 Box plot for the accuracy values of SV M, SV M-GW Ok=2 , SV M-GW Ok=3 , and SV M-
GW Ok=4

approach with k = 2 gives better results for all runs. The worst accuracy value of
it is better than the best one for the others. In addition, the standard deviation for
the proposed approach is recognized to be very minimal, which indicates a stable
approach.

7 Conclusion

In this paper, ransomware detection has been applied by proposing a hybrid approach
of evolutionary clustering approach using GW O with an ensemble of SV M clas-
sifier on a dataset of an Android benign and ransomware applications, to boost the
detection accuracy of this approach in comparison with the conventional single clas-
sification approach. Experiments are made to detect the best number of clusters for
GW O, and to compare the results achieved by the proposed approach and the corre-
sponding single classification approach. Results show that SV M-GW O with k = 2
outperforms the single classification approach in terms of accuracy.
For future work, this paper can be extended by enhancing the proposed approach
and experimenting it with different evolutionary clustering algorithms and different
classification algorithms, in which a different classifier can be identified for differ-
ent clusters obtained from clustering. Further, the proposed approach could also be
experimented on unbalanced ransomware dataset or any other applications.
246 R. Qaddoura et al.

References

1. Abdelsalam, Mahmoud, Ram Krishnan, and Ravi Sandhu. 2017. Clustering-based IaaS cloud
monitoring. In 2017 IEEE 10th international conference on cloud computing (CLOUD), 672–
679. IEEE.
2. Aggarwal, Charu C. 2015. Data classification. In Data mining, 285–344. Springer.
3. Akram, Beenish Ayesha, Ali Hammad Akbar, and Ki-Hyung Kim. 2018. CEnsLoc:
Infrastructure-less indoor localization methodology using GMM clustering-based classification
ensembles. Mobile Information Systems 2018.
4. Al-Madi, Nailah, Ibrahim Aljarah, and Simone A. Ludwig. 2014. Parallel glowworm swarm
optimization clustering algorithm based on mapreduce. In 2014 IEEE symposium on swarm
intelligence, 1–8. IEEE.
5. Al Shorman, Amaal, Hossam Faris, and Ibrahim Aljarah. 2020. Unsupervised intelligent system
based on one class support vector machine and grey wolf optimization for IoT botnet detection.
Journal of Ambient Intelligence and Humanized Computing 11 (7): 2809–2825.
6. Alhawi, Omar M.K., James Baldwin, and Ali Dehghantanha. 2018. Leveraging machine learn-
ing techniques for windows ransomware network traffic detection. In Cyber threat intelligence,
93–106. Springer.
7. Aljarah, Ibrahim, Al-Zoubi Ala’M, Hossam Faris, Mohammad A. Hassonah, Seyedali Mir-
jalili, and Heba Saadeh. 2018. Simultaneous feature selection and support vector machine
optimization using the grasshopper optimization algorithm. Cognitive Computation 1–18.
8. Aljarah, Ibrahim, and Simone A. Ludwig. 2012. Parallel particle swarm optimization clustering
algorithm based on mapreduce methodology. In 2012 fourth world congress on nature and
biologically inspired computing (NaBIC), 104–111. IEEE.
9. Aljarah, Ibrahim, and Simone A. Ludwig. 2013. Mapreduce intrusion detection system based
on a particle swarm optimization clustering algorithm. In 2013 IEEE congress on evolutionary
computation, 955–962. IEEE.
10. Aljarah, Ibrahim, and Simone A. Ludwig. 2013. A new clustering approach based on glowworm
swarm optimization. In 2013 IEEE congress on evolutionary computation, 2642–2649. IEEE.
11. Aljarah, Ibrahim, and Simone A. Ludwig. 2013. Towards a scalable intrusion detection sys-
tem based on parallel PSO clustering using mapreduce. In Proceedings of the 15th annual
conference companion on Genetic and evolutionary computation, 169–170.
12. Aljarah, Ibrahim, Majdi Mafarja, Ali Asghar Heidari, Hossam Faris, and Seyedali Mir-
jalili. 2020. Clustering analysis using a novel locality-informed grey wolf-inspired clustering
approach. Knowledge and Information Systems 62 (2): 507–539.
13. Aljarah, Ibrahim, Majdi Mafarja, Ali Asghar Heidari, Hossam Faris, and Seyedali Mirjalili.
2020. Multi-verse optimizer: Theory, literature review, and application in data clustering. In
Nature-inspired optimizers, 123–141. Springer.
14. Alsoghyer, S., and I. Almomani. 2020. On the effectiveness of application permissions for
android ransomware detection. In 2020 6th conference on data science and machine learning
applications (CDMA), 94–99.
15. Alsoghyer, Samah, and Iman Almomani. 2019. Ransomware detection system for android
applications. Electronics 8 (8): 868.
16. Arrott, Anthony, Arun Lakhotia, Ferenc Leitold, and Charles LeDoux. 2018. Cluster analy-
sis for deobfuscation of malware variants during ransomware attacks. In 2018 international
conference on cyber situational awareness, data analytics and assessment (Cyber SA), 1–9.
IEEE.
17. Chen, Jing, Chiheng Wang, Ziming Zhao, Kai Chen, Ruiying Du, and Gail-Joon Ahn. 2017.
Uncovering the face of android ransomware: Characterization and real-time detection. IEEE
Transactions on Information Forensics and Security 13 (5): 1286–1300.
18. Chen, Li, Chih-Yuan Yang, Anindya Paul, and Ravi Sahita. 2018. Towards resilient machine
learning for ransomware detection. arXiv:1812.09400.
A Classification Approach Based on Evolutionary … 247

19. Dhal, Krishna Gopal, Arunita Das, Swarnajit Ray, and Sanjoy Das. 2019. A clustering based
classification approach based on modified cuckoo search algorithm. Pattern Recognition and
Image Analysis 29 (3): 344–359.
20. Dion, Y., and Sarfraz N. Brohi. 2020. An experimental study to evaluate the performance of
machine learning alogrithms in ransomware detection. Journal of Engineering Science and
Technology 15 (2): 967–981.
21. Du, Jun, Emin Erkan Korkmaz, Reda Alhajj, and Ken Barker. 2005. Alternative clustering
by utilizing multi-objective genetic algorithm with linked-list based chromosome encoding. In
International workshop on machine learning and data mining in pattern recognition, 346–355.
Springer.
22. Faris, Hossam, Ibrahim Aljarah, Mohammed Azmi Al-Betar, and Seyedali Mirjalili. 2018.
Grey wolf optimizer: A review of recent variants and applications. Neural Computing and
Applications 30 (2): 413–435.
23. Faris, Hossam, Ibrahim Aljarah, and Ja’far Alqatawna. 2015. Optimizing feedforward neural
networks using krill herd algorithm for e-mail spam detection. In 2015 IEEE Jordan conference
on applied electrical engineering and computing technologies (AEECT), 1–5. IEEE.
24. Faris, Hossam, Ibrahim Aljarah, Seyedali Mirjalili, Pedro A. Castillo, and Juan Julián Merelo
Guervós. 2016. EvoloPy: An open-source nature-inspired optimization framework in Python.
In IJCCI (ECTA), 171–177.
25. Hajeer, Mustafa, and Dipankar Dasgupta. 2017. Handling big data using a data-aware HDFS
and evolutionary clustering technique. IEEE Transactions on Big Data 5 (2): 134–147.
26. Han, Jiawei, Jian Pei, and Micheline Kamber 2011. Data mining: Concepts and techniques.
Elsevier.
27. Hastie, Trevor, Robert Tibshirani, and Jerome Friedman. 2009. The elements of statistical
learning: Data mining, inference, and prediction. Springer Science & Business Media.
28. Hirano, Manabu, and Ryotaro Kobayashi. 2019. Machine learning based ransomware detection
using storage access patterns obtained from live-forensic hypervisor. In 2019 sixth international
conference on internet of things: Systems, Management and security (IOTSMS), 1–6. IEEE.
29. Hou, Biao, Chen Yang, Bo Ren, and Licheng Jiao. 2018. Decomposition-feature-iterative-
clustering-based superpixel segmentation for PolSAR image classification. IEEE Geoscience
and Remote Sensing Letters 15 (8): 1239–1243.
30. Kim, Youngjun, Ellen Riloff, and Stéphane M. Meystre. 2017. Exploiting unlabeled texts
with clustering-based instance selection for medical relation classification. In AMIA annual
symposium proceedings, vol. 2017, 1060. American Medical Informatics Association.
31. Lee, Kyungroul, Sun-Young Lee, and Kangbin Yim. 2019. Machine learning based file entropy
analysis for ransomware detection in backup systems. IEEE Access 7: 110205–110215.
32. Liu, Yimin, Tansel Özyer, Reda Alhajj, and Ken Barker. 2005. Integrating multi-objective
genetic algorithm and validity analysis for locating and ranking alternative clustering. Infor-
matica 29 (1).
33. Lu, Ting, Shutao Li, and Jón Atli Benediktsson. 2017. Iterative clustering based active learning
for hyperspectral image classification. In 2017 IEEE international geoscience and remote
sensing symposium (IGARSS), 3664–3667. IEEE.
34. Mirjalili, Seyedali, Seyed Mohammad Mirjalili, and Andrew Lewis. 2014. Grey wolf optimizer.
Advances in Engineering Software 69: 46–61.
35. Naik, Nitin, Paul Jenkins, Jonathan Gillett, Haralambos Mouratidis, Kshirasagar Naik, and
Jingping Song. 2019. Lockout-tagout ransomware: A detection method for ransomware using
fuzzy hashing and clustering. In 2019 IEEE symposium series on computational intelligence
(SSCI), 641–648. IEEE.
36. Naik, Nitin, Paul Jenkins, and Nick Savage. 2019. A ransomware detection method using fuzzy
hashing for mitigating the risk of occlusion of information systems. In 2019 international
symposium on systems engineering (ISSE), 1–6. IEEE.
37. Naik, Nitin, Paul Jenkins, Nick Savage, and Longzhi Yang. 2019. Cyberthreat hunting-part 2:
Tracking ransomware threat actors using fuzzy hashing and fuzzy c-means clustering. In 2019
IEEE international conference on fuzzy systems (FUZZ-IEEE), 1–6. IEEE.
248 R. Qaddoura et al.

38. Niture, Nandkumar A. 2020. Machine learning and cryptographic algorithms—analysis and
design in ransomware and vulnerabilities detection. Machine Learning 1: 19.
39. Onan, Aytug. 2017. Hybrid supervised clustering based ensemble scheme for text classification.
Kybernetes.
40. Pedregosa, F., G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P.
Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M.
Perrot, and E. Duchesnay. 2011. Scikit-learn: Machine learning in Python. Journal of Machine
Learning Research 12: 2825–2830.
41. Peng, Peter, Omer Addam, Mohamad Elzohbi, Sibel T. Özyer, Ahmad Elhajj, Shang Gao,
Yimin Liu, Tansel Özyer, Mehmet Kaya, Mick Ridley, et al. 2014. Reporting and analyzing
alternative clustering solutions by employing multi-objective genetic algorithm and conducting
experiments on cancer data. Knowledge-Based Systems 56: 108–122.
42. Poudyal, Subash, Dipankar Dasgupta, Zahid Akhtar, and K. Gupta. 2019. A multi-level ran-
somware detection framework using natural language processing and machine learning. In 14th
international conference on malicious and unwanted software, MALCON.
43. Poudyal, Subash, Kul Prasad Subedi, and Dipankar Dasgupta. 2018. A framework for analyz-
ing ransomware using machine learning. In 2018 IEEE symposium series on computational
intelligence (SSCI), 1692–1699. IEEE.
44. Qaddoura, Raneem, Waref Al Manaseer, Mohammad A.M. Abushariah, and Mohammad Aref
Alshraideh. 2020. Dental radiography segmentation using expectation-maximization clustering
and grasshopper optimizer. Multimedia Tools and Applications.
45. Qaddoura, Raneem, Hossam Faris, and Ibrahim Aljarah. 2020. An efficient clustering algorithm
based on the k-nearest neighbors with an indexing ratio. International Journal of Machine
Learning and Cybernetics 11 (3): 675–714.
46. Qaddoura, Raneem, Hossam Faris, Ibrahim Aljarah, and Pedro A Castillo. 2020. EvoCluster:
An open-source nature-inspired optimization clustering framework in Python. In International
conference on the applications of evolutionary computation (part of EvoStar), 20–36. Springer.
47. Qaddoura, R., H. Faris, and I. Aljarah. 2020. An efficient evolutionary algorithm with a near-
est neighbor search technique for clustering analysis. Journal of Ambient Intelligence and
Humanized Computing 1–26.
48. Qaddoura, R., H. Faris, I. Aljarah, J. Merelo, and P. Castillo. 2020. Empirical evaluation of
distance measures for nearest point with indexing ratio clustering algorithm. In Proceedings of
the 12th international joint conference on computational intelligence, vol. 1, 430–438. NCTA.
ISBN 978-989-758-475-6, https://doi.org/10.5220/0010121504300438.
49. Shukri, Sarah, Hossam Faris, Ibrahim Aljarah, Seyedali Mirjalili, and Ajith Abraham. 2018.
Evolutionary static and dynamic clustering algorithms based on multi-verse optimizer. Engi-
neering Applications of Artificial Intelligence 72: 54–66.
50. Taamneh, Madhar, Salah Taamneh, and Sharaf Alkheder. 2017. Clustering-based classification
of road traffic accidents using hierarchical clustering and artificial neural networks. Interna-
tional Journal of Injury Control and Safety Promotion 24 (3): 388–395.
51. Vinayakumar, R., K.P. Soman, K.K. Senthil Velan, and Shaunak Ganorkar. 2017. Evaluating
shallow and deep networks for ransomware detection and classification. In 2017 international
conference on advances in computing, communications and informatics (ICACCI), 259–265.
IEEE.
52. Zhang, Qi, Yang Xiao, Jingfeng Suo, Yu. Jun Shi, Yi Guo Jinhua, Yuanyuan Wang, and
Hairong Zheng. 2017. Sonoelastomics for breast tumor classification: A radiomics approach
with clustering-based feature selection on sonoelastography. Ultrasound in Medicine & Biology
43 (5): 1058–1069.