International Journal of Computer Networks & Communications (IJCNC) Vol.2, No.

5, September 2010

Meta-learning in Grid-based Data Mining Systems

Moez Ben Haj Hmida and Yahya Slimani

Faculty of Sciences of Tunisia

Campus Universitaire. 2092 El Manar, Tunis, Tunisia
Phone : +216 71 872 600 - Fax : +216 71 871 666

{moez.benhajhmida, yahya.slimani}@fst.rnu.tn

Abstract

The Weka4GML framework has been designed to meet the requirements of distributed data
mining. In this paper, we present the Weka4GML architecture based on WSRF technology for
developing meta-learning methods to deal with datasets distributed among Data Grid. This
framework extends the Weka toolkit to support distributed execution of data mining methods, like
meta-learning. The architecture and the behaviour of the proposed framework are described in
this paper. We also detail the different steps needed to execute a meta-learning process on a
Globus environment. The framework has been discussed and compared to related works.

Keywords

Data mining; meta-learning; grid computing; distributed dataset; WSRF.

1 Introduction

Nowadays, we have a deluge of data from scientific, industry and commerce fields. Massive
amounts of data that are being collected are often heterogeneous, geographically distributed and
owned by different organisations. These data contain hidden knowledge that have to be extracted.
Data mining (also known as Knowledge Discovery) is the process of discovering useful
knowledge from data.

Data mining is a massive computing task that deals with memory resident data. With the huge
amount of stored data in a centralised or distributed system, traditional data mining techniques are
inefficient. The need of parallel and distributed computing becomes inevitable to deal with large-
scale data mining (Kargupta, 2000). Parallel data mining deals with tightly-coupled systems
based on fast networks, while distributed data mining deals with loosely-coupled systems with
lower interconnection. A wide range of algorithms have been proposed for these two fields and
demonstrate relatively good performances with large-scale datasets. But with the emerging
concept of Virtual Organizations that proposes to interact different administration domains or
organizations as a unified system, these techniques become obsolete. Obviously the design of
new solutions based on grid systems is needed.

The Grid is the computing architecture that provides capabilities for manipulating geographically
distributed resources as a single meta-system. The Grid provides protocols and services to share

DOI : 10.5121/ijcnc.2010.2514 215

International Journal of Computer Networks & Communications (IJCNC) Vol.2, No.5, September 2010

computing resources data and software managed by multi-institutional virtual organizations

(Reed, 2003). Grid services are provided by middleware systems such as Globus, Legion, gLite,
and Unicore (Reed, 2003). These middleware systems are meta-systems that run on top of
existing operating systems and networks. The development of grid systems supporting data
mining is needed to analyse and extract knowledge from data stored on grid systems. The design
of such systems has to take into account emerging Grid standards like Web Service Resource
Framework (WSRF).

WSRF is a set of proposed Web services specifications that describe the means by which a
stateful resource is defined and associated with a Web services description, forming the overall
type definition of a WS-Resource (Czajkowski, 2004).

This paper describes the development of a new framework based on WSRF standards and
discusses design aspects and execution mechanisms. This framework, called Weka4GML, aims to
port the Weka (Witten, 2005), a widely used data mining suite, to a Grid environment.

The remainder of the paper is organised as follows: section 2 gives a background about the
learning classifiers from distributed data sources, where section 3 introduces the evolution of
distributed data mining to Grid environments. Section 4 presents a WSRF-based implementation
of the Weka4GML services. Section 4 tries to compare our proposition to previous works.
Finally, Section 6 concludes the paper and opens future perspectives.

2 Distributed data mining

The objective of distributed data mining is to perform the data mining operations using the
peculiarities and the availability of distributed resources (Kargupta, 2000). It deals with loosely-
coupled systems: generally share-nothing machines interconnected by high-latency networks and
geographically distributed (Zaki, 2000).

Typically a distributed data mining algorithm partitions the database at several sites and mines a
local model for each partition, then merges them to build a global model. DDM algorithms try to
learn from the distributed data generally without exchanging the raw data.

Most of the distributed data mining classification approaches are based on the ensemble learning
method (Dietterich, 2000; Promidis, 2000). Ensemble methods are learning algorithms that
construct a set of classifiers (base-classifiers) and then classify new data points by taking a
(weighted or unweighted) vote of their predictions (Dietterich, 2000). Typically, an ensemble
learning method runs the learning algorithm several times, each time with a different subset of the
training examples. This method can be applied to a distributed environment, where data are stored
on different sites, thus producing different local models to be aggregated into a global model.
Based on this observation, several methods have been proposed such as: meta-learning,
knowledge probing, and collective learning (Kargupta, 2000).

The meta-learning method was proposed for homogeneous distributed databases (Prodromidis,
2000). A meta-learning process follows three steps, as shown in figure 1:
1 A base-classifier is trained from the local training-set at each site.
2 The produced classifiers are collected to a central site where their predictions are
generated on a separate validation-set to produce metalevel
data.
216
International Journal of Computer Networks & Communications (IJCNC) Vol.2, No.5, September 2010

3 The global classifier (meta-classifier) is trained from the meta-level data.

The meta-classifier algorithm can be an arbiter or a combiner of the different predictions

composing the meta-level data. Meta-learning reduces the communication cost when it exchanges
the learned models instead of training examples and shows good scalability on data size when it
learns from small subsets that fit in memory. These properties make the meta-learning method
suitable for large systems like Grids.

3 Grid-based data mining

Grid computing is a natural evolution of distributed computing that attempts to better utilize
unused compute capacity by exploiting the computing power of a large numbers of server
computers, desktop PCs, clusters, and other kind of hardware. The Grid is a distributed
infrastructure that coordinates resource sharing and problem solving in dynamic, multi-
institutional virtual organizations. A virtual organization consists on a set of individuals and
institutions sharing resources regarding a set of rules (Reed, 2003).

Traditional distributed computing technologies do not provide the flexibility and control on
sharing relationships needed to establish VOs. The Open Grid Services Architecture (OGSA)
(Reed, 2003) is proposed as a Service Oriented Architecture (SOA) for the development of Grid
systems. OGSA defines the mechanisms for creating, managing, and exchanging information
among entities, called Grid Services. The Globus Toolkit is a middleware providing a set of
OGSA capabilities based on WS-Resource Framework (WSRF) (Czajkowski, 2004). The WSRF
is a set of Web Service specifications that describe how to implement OGSA capabilities using
Web services.

Grid-based data mining has emerged for supporting complex data mining operations over
geographically distributed large data sets. Grid systems provide services for data access and
management for such distributed data. However, these services are not sufficient to analyze and
extract knowledge from huge amount of data stored within grid environments. Efforts for Grid-
enabling data mining applications produced a variety of systems, algorithms, and other tools (Ben
HajHmida, 2009).

Several projects like TeraGrid (Catlett, 2007) focused on the creation of Grid infrastructure
providing tools for the management and the manipulation of distributed data sources. Other
projects like the Knowledge Grid (Cannataro, 2003) have focused on the creation of systems for
distributed knowledge discovery over Grid infrastructures.

The TeraGrid project aims to provide a Grid infrastructure (Catlett, 2007) built over resources
available at four main sites (San Diego Supercomputing Center, National Center for
Supercomputing Applications, Caltech and Argonne National Lab). The architecture of the
TeraGrid adopts the existing Grid software technologies building a "virtual system" that describes
the capabilities and behavior of a resource through service specifications. TeraGrid offers a set of
application services, called TeraGrid Application Services, implemented upon basic software
components, called Grid Services.

The Knowledge Grid (Cannataro, 2003) is a high-level, service-oriented framework providing

grid-based data mining tools and services. The Knowledge Grid system uses the low-level
features of Globus allowing for the creation of geographically distributed data mining
217
International Journal of Computer Networks & Communications (IJCNC) Vol.2, No.5, September 2010

applications and related tasks, such as data management and knowledge representation. Recent
developments of the Knowledge Grid (Congiusta, 2007) have accomplished a re-design and re-
implementation of all the Knowledge Grid services as WSRF-compliant Grid Services. In such
latter version, the Knowledge Grid exposes two main set of services: Resource Management
Services and Execution Management Services. Through such services it is possible to design and
execute complex knowledge discovery applications through a visual environment.

TeraGrid and he Knowledge Grid are the most complete ongoing projects embracing the
emerging trends and providing an almost complete set of functionalities. But the deployment and
the use of such systems are quite consequent. The execution of a data mining task on distributed
datasets needs a design of a complex mining process by an experimented user. Contrariwise we
propose a simpler framework that hides to the user the complexity of distributed data mining
tasks execution.

4 Weka4GML framework

In this section we propose a novel framework, Weka4GML, enabling the support of meta-
learning on a Grid environment. Based on Web services technology, Weka4GML extends the
Weka toolkit (Witten, 2005) to Grid environments. Weka is a collection of sequential data mining
algorithms for knowledge discovery comprising standard data preprocessing, mining, and
visualization techniques. The Weka4GML framework is based on the Globus Toolkit, a widely
used Grid middleware supporting the WSRF standard. The proposed framework architecture is
WSRF compliant to be deployed on top of the Globus middleware.

4.1. Architecture description

The proposed architecture is designed according to the meta-learning execution scheme. As

shown in Figure 2, our framework is composed of four node types: storage nodes, base-
classification nodes, metaclassification nodes and user nodes. A storage node hosts one or several
fragments of the distributed dataset. It also publishes the stored data and their properties as a Web
service and shares local data with other Grid nodes via an FTP server. A base-classification node
builds a local model by executing a base-classification algorithm on a local dataset partition. The
produced model is then applied on the validation dataset to generate predictions and transfer them
to the meta-classification node. A metaclassification node integrates an FTP server to collect
meta-data on which a meta-level algorithm is executed to produce the final classifier. A user node
offers a graphical interface allowing users to choose the Weka supported algorithms, to explore
the datasets stored on the storage nodes, and to execute a meta-learning process on a Grid.

4.2. Behaviour description

As shown in Figure 2, each node comprises a Web service, allowing remote applications
invocation. Depending on its type, a node can also comprise a Weka module, an FTP server, or a
database server. In the following, we provide a behaviour description of the different nodes
composing the framework.

Storage node: a storage node is composed of a database server, an FTP server and a Web service
(Data Service). The database server stores the datasets to be mined. If the dataset is distributed on
several partitions, they can be stored on one storage node or among several ones. The FTP server

218
International Journal of Computer Networks & Communications (IJCNC) Vol.2, No.5, September 2010

contributes to the data distribution management, allowing the dataset partitions transfer to other
storage nodes (validation data) or a baseclassification node. The Data Service interacts with the
data management service of a user node by exposing properties concerning the stored datasets.
These properties, described in an XML file, comprise the dataset name, its type (centralised or
distributed), its size and the data type (discrete or continuous).

Base-classification node: a base-classification node is composed of a Weka module and a Web

service (Base-Classification Service). The Weka module comprises the set of algorithms provided
by Weka toolkit. These algorithms are exposed by the base-classification service allowing remote
invocation. The Base-Classification Service acts as an FTP client when it downloads from a
storage node the data to be analysed. It allows also meta-data upload towards meta-classification
node.

Meta-classification node: a meta-classification node differs from a base-classification node by

including an FTP server. This kind of node allocates base-classification tasks to the base-
classification nodes participating to the meta-learning process, and then collects the meta-data
produced by these nodes. Since a meta-classification node comprises a Weka module, it can also
act as a base-classifier. In addition of exporting the algorithms provided by Weka module, the
Meta-Classification Service notifies the user node of the final result.

User node: a user node is composed of a Weka module, a data manager and a meta-learning
module. The Weka module allows the user to choose the base-classification and meta-
classification algorithms composing a meta-learning process. The data manager allows gathering
the properties of the datasets stored on the Grid. These properties are used to identify the various
datasets and their distribution schema (number of partitions, addresses of the storage nodes, etc).
After choosing the algorithms and the dataset to be analysed, the user submits a meta-learning
process to the meta-learning module. This module interacts with the Meta-Classification Service,
to submit the needed requests to the execution of the metalearning process. Each request contains
the meta-classifier name, the baseclassifiers names, the base-classification node URIs and URLs
of the respective dataset partitions to be analysed.

4.3. Web Service Operations

As recommend by the WSRF specifications, the WS-ResourceLifetime supplies some basic

mechanisms to manage a resource lifecycle. This implies three basic operations: resource creation
(createResource), subscription to notification (subscribe), and resource destruction (destroy).

Meta-learning services

In addition of these WSRF specific operations, each Weka4GML service needs another operation
to execute the meta-learning process:
- Meta-Learning Service exposes metaLearn operation through which it
receives the arguments of the process to be executed.
- Base-Learning service exposes the learn operation which executes a
learning algorithm on target data.
The arguments needed by Meta-Learning Service to execute the metaLearn operation are
declared on the WSDL file as metaLearningParameters. Figure 3 shows an excerpt of the WSDL
type declarations associated with the Meta-Learning Service. The algorithm type defines data
mining algorithm identified by its name and parameters, where the algorithmList type is a
219
International Journal of Computer Networks & Communications (IJCNC) Vol.2, No.5, September 2010

complex type defining a list containing at least one element. In the same way, a type datasetList
is defined as a list containing at least one dataset. The dataset type contains a name describing the
dataset, a boolean telling if the dataset is distributed or not and an uri permitting the remote
access to the dataset partition. The datasetList describes all partitions forming the distributed
dataset.
The meta-learning process is performed through the metaLearn operation starting from the
parameters passed by the client. These parameters contain the list of dataset partition, their
relative validation set, the list of base-learning algorithms and the meta-learning algorithm.

Whenever the Meta-learning service is invoked, it creates a new WSResource which will store
the result of the meta-learning process on the FinalModel property. This property is declared on
the WSDL as shown by Figure 4.

Data services

In addition of these WSRF specific operations, a data service of a storage node needs other
operations to handle the stored data Data Service
exposes:
- createData operation to store datasets or dataset partitions. This operation is invoked by a user
node to add a new distributed dataset among storage nodes.
- searchData operation to search stored datasets on grid nodes. This operation is invoked by a
user node to search available datasets.
- createValidationData operation to create a validation dataset that will be used on the meta-
learning process. This operation is invoked by a user node after adding a new dataset or by the
meta-learning service during a meta-learning process execution.

The arguments needed by Data Service to execute the createValidationData operation are
declared on the WSDL file excerpt shown in Figure 5. The sampling algorithm is responsible of
extracting a number of representative rows of each data partition contained by the datsetList. In
each concerned storage node, local validation dataset is created. Finally, local datasets are
exchanged between storage nodes to be merged to build the validation dataset to be used on the
meta-learning process.

5 Meta-Learning process execution

This section describes the various steps of a meta-learning task execution in Weka4GML
framework. Figure 5 describes the interactions between the different framework nodes and their
respective Web services for the execution of a meta-learning task. This example does not take
into account the dataset discovery by the data management module (user node) as well as the
creation of validation data by the storage nodes. The example supposes that the data mining
process was already submitted to the meta-learning module. This process is executed by
respecting the following steps (see Figure 6):

1 Resource creation (meta-classification node): the meta-learning module (user node) invokes the
createResource operation of the Meta-Classification Service, which creates a new WS-Resource.
This resource will be in charge of the maintenance of the state of the submitted meta-learning
process. The state is stored as final model resource property. This resource is identified by a
unique reference, EndPoint Reference (EPR), that distinguishes it from the remainder of

220
International Journal of Computer Networks & Communications (IJCNC) Vol.2, No.5, September 2010

the resources of the same service. This reference will be used, thereafter, by the meta-learning
module for future invocations of this resource.

2 Notification subscription (meta-classification node): the meta-learning module invokes the

subscribe operation, which subscribes to the notification of the changes of final model resource
property. With each change of this property, one notification containing the new value will be
sent to the meta-learning module.

3 Task submission (meta-classification node): the meta-learning module invokes the metaLearn
operation, which executes a meta-learning algorithm. This operation receives arguments such as
the metaclassifier name, the names of the base-classifiers and the baseclassification node URIs
and URLs of the partitions to be analysed.

4 Resource creation (base-classification node): the Meta-Classification Service invokes in turn

the createResource operation of the Base- Classification Services passed as arguments in the
metaLearn operation. In each node, a new WS-Resource is created and identified by an EPR.

5 Notification subscription (base-classification node): the Meta- Classification Service node

subscribes to the notification of the changes of the predictions property of each new resource of a
Base- Classification Service.

6 Task submission (base-classification): the Meta-Classification Service invokes the learn

operation of each Base-Classification Service to execute base- classification algorithm in parallel.
The learn operation receives arguments such as the base-classification algorithm name and the
URL of the partition to be analysed.

7 Data transfer: each Base-Classification Service downloads the dataset partition to be analysed
and the corresponding validation data, by sending a request to the FTP server of the target storage
node.

8 Base-classification: when data to be analysed are downloaded, each node of the Base-
Classification Service executes the required algorithm. A classification algorithm is executed to
create a basemodel, tested on the validation data, to generate predictions which will form the
meta-data. The result is recorded in the predictions property of the resource created on step 4.

9 Results notification (base-classifiers): whenever the state of the property predictions has been
changed, the Base-Classification Service notifies its new value to the Meta-Classification Service
by invoking the deliver operation. A request is sent to the FTP server of the meta-classification
node to transfer the validation data. Since the validation data is the same on all nodes, the FTP
server handles the first transfer request and ignores the others.

10 Resource destruction (base-classification node): the Meta- Classification Service invokes the
destroy operation whenever it is notified by a Base- Classification Service, which destroys each
resource created on step 4.

11 Meta-classification: upon predictions of the various base-classifiers and the gathered

validation data, the Meta-Classification Service executes a meta-level algorithm provided by the
Weka module to create the meta-data and to generate the final model. The result is stored in final
model property of the resource created on step 1.
221
International Journal of Computer Networks & Communications (IJCNC) Vol.2, No.5, September 2010

12 Final result notification: when the state of final model property has been changed, the Meta-
Classification Service notifies its new value to the user node by invoking the deliver operation.

13 Resource destruction (meta-classification node): the user node invokes the destroy operation,
which destroys the resource created on step 1.

6 Related works

Many projects like GridWeka (Khoussainov, 2004), WekaG (Pérez, 2005), Weka4WS (Talia,
2005) and FAEHIM (Ali, 2005) aimed to adapt the toolkit Weka to a Grid environment.

GridWeka (Khoussainov, 2004) is an ongoing work at the University of Dublin, which distributes
classification algorithms using the crossvalidation method over computers in an ad-hoc Grid. The
system is composed of two main components: Weka Server and Weka Client. Each participating
machine runs the original Weka as server component. The Weka client allows users to specify the
target algorithms and datasets, and also to specify resources and tasks constraints.

Another system using the client server architecture is WekaG (Pérez, 2005). The WekaG server
implements data mining algorithms while the client side instantiates grid services. WekaG is an
implementation of a more general architecture called Data Mining Grid Architecture (DMGA)
based on Weka and Globus Toolkit 3. DMGA provides the composition of services on
workflows. WekaG provides the same user interface as Weka and can support parallel algorithms.
The first prototype implemented uses only the Apriori algorithm as a grid service. Unfortunately,
this algorithm cannot be used by meta-learning methods, whereas Weka4GML allows the use of
any algorithm supported by Weka by exposing the entire Weka module as a Web service.

Weka4WS (Talia, 2005) extends Weka allowing the execution of all its data mining algorithms on
remote Grid nodes. To enable remote invocation, the data mining algorithms provided by the
Weka library are exposed as a Web Service, which can be easily deployed on the available Grid
nodes. The architecture of Weka4WS includes three kinds of nodes: storage nodes, which contain
the datasets to be mined; compute nodes, on which remote data mining algorithms are run; user
nodes, which are the local machines of users. Remote execution is managed using basic WSRF
mechanisms (state management, notifications, etc.), while the Globus Toolkit 4 services are used
for standard Grid functionalities, such as security and file transfer. Weka4WS can only handle a
dataset contained by a single storage node. This dataset is then transferred to computing nodes to
be mined. If data are considerably large this transfer will cause high communication overhead. On
the contrary, Weka4GML handles naturally distributed datasets that are contained by several
storage nodes. In its last version Weka4WS supports the design of complex discovery processes,
such as meta-learning, by porting the Weka KnowledgeFlow to Web services. This means that the
user has to compose by itself all the meta-learning process, whereas Weka4GML takes in charge
the process composition making it transparent to the user.

Very similar is the Federated Analysis Environment for Heterogeneous Intelligent Mining
(FAEHIM)) project (Ali, 2005), aimed to support data mining based on a grid services approach.
The data mining algorithms implementations are derived from the Weka library and converted
into a set of grid services. Complex mining processes are designed using the Triana (Churches,
222
International Journal of Computer Networks & Communications (IJCNC) Vol.2, No.5, September 2010

2005) workflow engine. Like Weka4WS, FAEHIM do not support the data distribution and
leaves to the user the charge of complex discovery processes design, whereas Weka4GML
supports data distribution and automates the composition and the execution of a metalearning
process.

GridWeka and WekaG tried to adapt Weka to a Grid infrastructure using the client server
architecture. But this architecture is not so flexible and powerful to support large scale
requirements such as security, resource management etc. By contrast, Weka4GML, Weka4WS
and FAEHIM are based on a grid services approach.

Compared to the existing systems, our framework deals with distributed data and offers an easier
execution of the meta-learning process. Indeed, Weka4WS and FAEHIM provide workflow
engines by which users can compose a complex workflow schema such as the distribution of a
dataset on different nodes and the execution of a meta-learning process. However, Weka4GML
offers to users to discover distributed datasets, to choose the base classifiers to use, and then the
framework holds the execution and the coordination of the distributed tasks on the target nodes.

Table 1 compares the reviewed works and the Weka4GML framework according to the system
architecture, the supported standards, the used middleware, and the support of workflow. In this
table we note Middleware by MW, Web Services by WS and Extended Weka workflow engine
by EWWE.

Table 1 Comparing Weka4GML to the reviewed frameworks

System Architecture Standard MW Workflow

GridWeka Client/ WS JVM BPE
server
WekaG Client/ WSRF Globus3 Ad-hoc
server
Weka4WS Grid WSRF Globus4 EWWE
services
FAEHIM Grid WS Globus4 Triana
services
Weka4GML Grid WSRF Globus4 Framework
services handled

7 Conclusion

In this paper, we described the evolution of data mining from distributed systems to Grid systems.
Grid computing is a promising architecture that opens new horizons and faces new challenges.
We also proposed a new framework based on WSRF technology for the distributed execution of
the meta-learning approach. This framework is deployed on the top of the Globus middleware.
Comparatively to similar works, Weka4GML is a framework which allows the discovery of
knowledge in a completely distributed manner, starting from distributed data.

223
International Journal of Computer Networks & Communications (IJCNC) Vol.2, No.5, September 2010

The proposed framework can not concurrent other systems like TeraGrid or the Knowledge Grid.
It is considered as an initiative to port parallel and distributed mining algorithms to Grids
regarding the peculiarities of such infrastructure.

As perspective of our work, we propose to extend this framework to support other distributed data
mining techniques and to manage faulttolerance.

References

Ali, A. S., and Taylor, I. J. (2005) ‘Web services composition for

distributed data mining’, In the 2005 IEEE International Conference
on Parallel Processing Workshops, pp. 11-18.
Ben HajHmida, M., Congiusta, A. (2009) ‘Parallel, Distributed, and Grid-
Based Data Mining: Algorithms, Systems, and Applications’, in
Mario Cannataro (Ed.), Handbook of Research on Computational Grid Technologies for Life
Sciences, Biomedicine and Healthcare, IGI Global.
Cannataro, M. and Talia, D. (2003) ‘The knowledge grid: An architecture
or distributed knowledge discovery’, Commun. ACM, 46(1), pp. 89-
93.
Catlett, C. et al (2007), ‘TeraGrid: Analysis of Organization, System
Architecture, and Middleware Enabling New Types of Applications’,
In L. Grandinetti (Ed.), Advances in Parallel Computing. IOS Press.
Churches, D., Gombas, G., Harrison, A., Maassen, J., Robinson, C.,
Shields, M., Taylor, I., and Wang, I. (2005) ‘Programming scientific
and distributed workflow with triana services’, Concurrency and
Computation: Practice and Experience, 18(10), 1021-1037.
Congiusta, A., Talia, D., and Trunfio, P. (2007) ‘Distributed data mining
services leveraging wsrf’, Future Generation Computing Systems,
23(1), 34-41.
Czajkowski, K. (2004) ‘The WS-Resource Framework Version 1.0’,
http://www-106.ibm.com/developerworks/library/ws-resource/wswsrf.
pdf.
Dietterich, T. G. (2000) ‘An experimental comparison of three methods
for constructing ensembles of decision trees: Bagging, boosting, and
randomization’, Machine Learning, 40, 139-157.
Kargupta, H., and Chan, P. (Eds.), (2000). Advances in Distributed and
Parallel Knowledge Discovery. MIT Press.
Khoussainov, R., Zuo, X., and Kushmerick, N. (2004) ‘Grid-enabledweka:
A toolkit for machine learning on the grid’, ERCIM News, 59.
Pérez, M. S., Sànchez, A., Herrero, P., Robles, V., and Pena, J. M. (2005)
‘Adapting the weka data mining toolkit to a grid based environment’,
In 3rd Atlantic Web Intelligence Conference, pp. 492-497.
Prodromidis, A. and Chan, P. (2000) ‘Meta-learning in Distributed Data
Mining Systems: Issues and Approaches’, In Hillol Kargupta and
224
International Journal of Computer Networks & Communications (IJCNC) Vol.2, No.5, September 2010

Philip Chan (Eds.), Advances of Distributed Data Mining. MIT/AAAI

Press, pp. 81-114.
Reed, D. A., Mendes, C. L., Lu, C., Foster, I., and Kesselman, C. (Eds.),
(2003) The Grid 2: Blueprint for a New Computing Infrastructure -
Application Tuning and Adaptation. Morgan Kaufman.
Talia, D., Trunfio, P., and Verta, O. (2005) ‘Weka4WS: A WSRF-Enabled
Weka Toolkit for Distributed Data Mining on Grids’, In the 9th
European Conference on Principles and Practice of Knowledge
Discovery in Databases, 3721, 309-320.
Witten, I. H., and Frank, E. (Eds.), (2005) Data mining: Practical machine
learning tools and techniques. Morgan Kaufmann.

Authors

Moez Ben Haj Hmida received the Master degree in computer science from the University of
Tunis El Manar, Tunisa, in 2005. He is currently pursuing the PhD in computer sciences at the
same university. He is currently lecturer at the department of Computer Science of Faculty of
Sciences of Tunis, Tunisia. His current research interests include parallel, distributed and Grid-
based data mining applications.

Yahya Slimani received the B.Sc.(Eng.), Dr Eng and PhD degrees from the Computer Science
Institute of Alger’s (Algeria), University of Lille (French) and University of Oran (Algeria), in
1973, 1986 and 1993, respectively. He is currently Professor at the Department of Computer
Science of Faculty of Sciences of Tunis. These research activities concern data mining,
parallelism, distributed systems and grid computing.
Dr. Slimani has published more than 80 papers from 1986 to 2005. He contributed to Parallel and
Distributed Computing Handbook, Mc Graw-Hill, 1996. He joined the Editorial Boards of the
Information International Journal in 2000.

225
International Journal of Computer Networks & Communications (IJCNC) Vol.2, No.5, September 2010

Figure 1 Meta-learning from distributed data sources

226
International Journal of Computer Networks & Communications (IJCNC) Vol.2, No.5, September 2010

Figure 2 Meta-learning process on Weka4GML framework

Figure 3 Input/Output parameters of the metaLearn operation

<types>
...
<xsd:element name="algorithm">
<xsd:complexType>
<xsd:sequence>
<xsd:element name="name" type="xsd:string"/>
<xsd:element name="parameters" type="xsd:string"/>
</xsd:sequence>
</xsd:complexType>
</xsd:element>

<xsd:element name=" algorithmList ">

227
International Journal of Computer Networks & Communications (IJCNC) Vol.2, No.5, September 2010

<xsd:complexType>
<xsd:sequence>
<xsd:element name="algorithm" type="xsd: algorithm" minOccurs="1"
maxOccurs="unbounded"/>
</xsd:sequence>
</xsd:complexType>
</xsd:element>

<xsd:element name="dataset">
<xsd:complexType>
<xsd:sequence>
<xsd:element name="name" type="xsd:string"/>
<xsd:element name="distributed" type="xsd:boolean"/>
<xsd:element name="uri" type="xsd:string"/>
</xsd:sequence>
</xsd:complexType>
</xsd:element>

<xsd:element name=" datasetList ">

<xsd:complexType>
<xsd:sequence>
<xsd:element name=" dataset " type="xsd: dataset " minOccurs="1"
maxOccurs="unbounded"/>
</xsd:sequence>
</xsd:complexType>
</xsd:element>

<xsd:element name="metaLearningParameters">
<xsd:complexType>
<xsd:sequence>
<xsd:element name="taskID" type="xsd:long"/>
<xsd:element name="metaLearningAlgorithm" ref="tns:algorithm"/>
<xsd:element name="baseLearnersList" ref="tns:algorithmList"/>
<xsd:element name="datasetList" ref="tns:datasetList"/>
<xsd:element name="validationData" ref="tns:dataset"/>
</xsd:sequence>
</xsd:complexType>
</xsd:element>
</types>

228
International Journal of Computer Networks & Communications (IJCNC) Vol.2, No.5, September 2010

Figure 4 Meta-Learning Service Messages

<!-- RESOURCE PROPERTIES -->

<xsd:element name="FinalModel" type="xsd:string"/>

<xsd:element name="MetaLearningResourceProperties">
<xsd:complexType>
<xsd:sequence>
<xsd:element ref="tns: FinalModel " minOccurs="1"
maxOccurs="1"/>
</xsd:sequence>
</xsd:complexType>
</xsd:element>

Figure 5 Input/Output parameters of the createValidationData operation

<xsd:element name=" createValidationDataParameters">

<xsd:complexType>
<xsd:sequence>
<xsd:element name="taskID" type="xsd:long"/>
<xsd:element name="samplingAlgorithm" ref="tns:algorithm"/>
<xsd:element name="datasetList" ref="tns:datasetList"/>
<xsd:element name="validationData" ref="tns:dataset"/>
</xsd:sequence>
</xsd:complexType>
</xsd:element>

229
International Journal of Computer Networks & Communications (IJCNC) Vol.2, No.5, September 2010

Figure 6 Execution steps of meta-learning process

230