WEB SERVICE SELECTION BASED ON RANKING OF QOS USING ASSOCIATIVE CLASSIFICATION

International Journal on Web Service Computing (IJWSC), Vol.3, No.1, March 2012
DOI : 10.5121/ijwsc.2012.3101 1
WEB SERVICE SELECTION BASED ON RANKING OF
QOS USING ASSOCIATIVE CLASSIFICATION
Molood Makhlughian1
, Seyyed Mohsen Hashemi2
, Yousef Rastegari3
and Emad
Pejman1
1
Department of Computer Engineering, Islamic Azad University-South Tehran Branch,
Tehran, Iran
{molood.makhlughian, emad.pejman}@gmail.com
2
Dean of Software Engineering and Artificial Intelligence Department, Islamic Azad
University-Science and Research Branch, Tehran, Iran
hashemi@isrup.com
3
Department of Electrical and Computer Engineering, Shahid Beheshti University,
Tehran, Iran
y_rastegari@sbu.ac.ir
ABSTRACT
With the explosive growth of the number of services published over the Internet, it is difficult to select
satisfactory web services among the candidate web services which provide similar functionalities. Quality
of Service (QoS) is considered as the most important non-functional criterion for service selection. But this
criterion is no longer considered as the only criterion to rank web services, satisfying user’s preferences.
The similarity measure (outputs–inputs similarity) between concepts based on ontology in an inter-
connected network of semantic Web services involved in a composition can be used as a distinguishing
criterion to estimate the semantic quality of selected services for the composite service. Coupling the
semantic similarity as the functional aspect and quality of services allows us to further constrain and select
services for the valid composite services.
In this paper, we present an overall service selection and ranking framework which firstly classify
candidate web services to different QoS levels respect to user’s QoS requirements and preferences with an
Associative Classification algorithm and then rank the most qualified candidate services based on their
functional quality through semantic matching. The experimental results show that proposed framework can
satisfy service requesters’ non-functional requirements.
KEYWORDS
Web Service Selection, Quality of Service (QoS), Classification Data Mining & Semantic Web Services
1. INTRODUCTION
Service Oriented Computing (SOC) and its predominant incarnation as Web Services have
emerged as a powerful concept for building software systems [1].

2
Web service technology offers a potential solution for developing distributed business processes
and applications, which can be accessible via the Internet. Considering the rapid increase of Web
users and the growing complexity of their demands, simple atomic services are inadequate. When
individual Web services are not able to meet complex requirements, SOC provides a flexible
framework for reusing and composing existing web services in order to build value-added
composite services. At first the functionalities required for these complex requirements (namely
the tasks) and their interactions, the control and data flow, are identified. Secondly, an appropriate
implementation must be selected and bounded to each task. However with the rapidly growing
number of available services, a large number of services can be found for realizing every task
which can provide expected functionality for each of them. So it leads to the issue of selecting the
best Web services among a list of “Candidate Web services”, with the same functionalities. These
services are, of course, different from one another in non-functional properties such as response
time, availability, throughput, security, reliability, and execution cost [2], and are therefore
different in terms of efficiency.
A specific issue emerges to this regard is selecting the best set of services based on their Quality
of services (QoS) to participate in the composition, meeting QoS constraints set by users. QoS is
a measure for how well a service serves the customer.
Dynamic service environments cause some difficulties in service selection. As the services’
availability cannot be known a priori, or QoS conditions fluctuate in such environments, service
selection and composition must be performed at runtime. Therefore the execution time of service
selection algorithms is heavily constrained, whereas the computational complexity of the problem
is NP-hard [3]. Hence, finding an optimal composite service may not be practical. Due to
changing conditions in such environments there is no guarantee that the selected services for
composite service will be available at runtime or its QoS will not be fixed concerning the
advertised one in WSDL.
In this paper, we present a service selection algorithm that copes with above issues. This
algorithm consists of two phases. In the first phase, we use a classification data mining algorithm
to classify web service candidates into different QoS levels respect to the defined QoS constraints
form the user and using the result of this classification to define a utility value for each of the
service candidates. In the second phase we focus on composing the best services of each task and
more specifically on their functional level that aims to selecting and inter-connecting web
services by means of their semantic connections.
The remainder of this paper is structured as follows. In the next Section we give an overview of
related works. In Section 3, we present our service selection approach and give the details of it,
and we conduct a set of experiments to evaluate its timeliness and optimality in Section 4.
Finally, in Section 5, we conclude with a summary of our contributions and the future
perspectives of this work.
2. RELATED WORKS
The approaches aim at determining the optimal service composition using brute-force-like
algorithms (i.e., the execution time and costs are exponential even if simplifications are used),
new approaches are satisfied with finding a nearly-optimal solution. These approaches are using
(meta-) heuristics. (Meta-) heuristics are general search methods that exclude a huge number of
solutions, because they do not consider optimal solution in their population. Hence (meta-)
heuristics are more efficient than exact algorithms. The main disadvantage however is that they
do not find the optimal solution in most cases, because, the exclusion of solutions is based on

3
assumptions, thus there is no guarantee, that the excluded solutions do not contain the optimal
one.
All existing approaches can be divided into exact approaches finding the optimal service
composition and (meta-) heuristics selecting a nearly service composition. Instead of exact
algorithms, (meta-) heuristics seem to be the better choice for QoS-aware service selection and
composition.
Beside many proprietary heuristics [4,5,6], currently the Genetic Algorithm is the most promising
heuristic in QoS-aware service selection and composition. C. Jaeger et al. [7] present a heuristic
based on the genetic algorithm. A big challenge related to the Genetic Algorithms is the choice of
the genetic operators: selection, mutation and recombination, which have a big influence of the
efficiency and correctness of the algorithm. As these parameters are chosen randomly, the order
in which service candidates are checked is chosen randomly. On the other hand, as the genetic
algorithm can run endlessly, the users have to define a constant number of iterations fixed a
priori, and fixing a high number of iterations does not give any guarantee about the quality of the
result. Therefore, the genetic algorithm is deemed not useful for the purpose of selecting near-
optimal compositions.
In recent years, some approaches (e.g., [8,9]) use the power of Data mining algorithms in
knowledge extraction and pattern discovery among the huge amounts of data in web service
selection.
Wu et al. [8] present a Bayesian network based Qos assessment model for web services. That
could predict the service capability in various combinations of users’ Qos requirements. This
approach is used to evaluate the capability of each service, and the one with best capability is
selected as the binding service. Though it uses Bayesian network classification algorithm for each
provider/service to predict the level of QoS, it is computationally complex and is based on
probabilities, moreover it just considers local constraints in web service selection and doesn’t
mention the global constraints.
Ben Mabrouk et al. [9] present a heuristic approach for service composition in dynamic
environments. This solution uses the K-means algorithm to classify the web services to QoS
levels, then it uses the result of this clustering in an utility function in order to rank web service
candidates for each task as a local selection part, then it uses a search tree to select the best
services to form the composition plans in an ordered way. The proposed solution is
computationally expensive in both of the clustering algorithm and the structure of the search tree
in a composition plans with the high number of activities, also it suffers from the deficiency of
the clustering techniques, because clustering is an example of unsupervised learning. Furthermore
it does not mention the semantic matching between output and input of the services.
3. QOS-AWARE SERVICE SELECTION ALGORITHM
Our proposed approach starts form the assumption that the user uses a graphical user interface to
define his/her request. With the help of this interface user can express his request in terms of QoS
requirements and his relative preferences for each of them. Then the interface formulates it as a
machine-understandable specification. The composition plan is given from the expert that its
functional tasks and all the controls and data flow between them are characterized. For every
activity in the composition plan, a service discovery phase gives a set of service candidates able
to fulfill the functional aspect.

4
The proposed service selection method consists of a heuristic algorithm based on classification
data mining algorithm, CBA [10] algorithm. This classification allows for classifying candidate
web services with respect to the QoS requirements and preferences defined by the user into a set
of different QoS levels. Further with considering the functional aspect we use these levels to
determine the utility of each candidate service to do a near-optimal selection.
As the outlined process of the proposed method is depicted in figure 1, our heuristic approach
deals with the service selection problem through the following phases:
1. Scaling phase, which is a pre-processing phase to normalizing QoS values.
2. Service selection by local classification, which classifies candidate services according to
different QoS levels and determines the utility of each candidate service.
3. Ranking based on functional aspect, which uses the results of the last phase to selecting
the best services according to the functional aspect.
Figure 1. The outlined process of the proposed method

5
3.1. Scaling Phase
QoS attributes could be either positive or negative, thus some QoS values need to be maximized,
(i.e., the higher the value, the higher the quality) for example availability and reliability, whereas
other values have to be minimized, (i.e., the higher the value, the lower the quality). This includes
criteria such as execution time and execution price. To cope with this issue, the scaling phase
normalizes every QoS attribute value according to the following formulas [11]. The values of
negative attributes are normalized by expression (1) and the values of positive attributes are
normalized by (2).
ܳ௜,௝
ᇱ
= ൝
ொೕ
೘ೌೣ
ିொ೔,ೕ
ொೕ
೘ೌೣ
ିொೕ
೘೔೙ , ݂݅ ܳ௝
௠௔௫
− ܳ௝
௠௜௡
≠ 0
1, ݈݁‫݁ݏ‬
(1)
ܳ௜,௝
ᇱ
= ቐ
ொ೔,ೕିொೕ
೘೔೙
ொೕ
೘ೌೣ
ିொೕ
೘೔೙ , ݂݅ ܳ௝
௠௔௫
− ܳ௝
௠௜௡
≠ 0
1, ݈݁‫݁ݏ‬
(2)
Where ܳ௜,௝
ᇱ
signifies the normalized value of QoS attribute j associated with candidate service ܵ௝.
It is computed using the current value ܳ௜,௝ and ܳ௝
௠௔௫
and ܳ௝
௠௜௡
which respectively denote the
maximum and minimum values of QoS attribute j among all the candidate services.
After the data normalization is completed, all the QoS attributes values are lying in the [0, 1]
interval and also have the same utility increase direction.
3.2. Service Selection by Local Classification
The classification is performed on candidate web services for all the activities in the abstract
service composition. The purpose of this phase is to classify candidate web services to the
multiple QoS levels with respect to the user’s preferences on QoS attributes. In this classification
each level consists of a set of candidate services that have approximately the same QoS value.
These levels determine the relative importance of candidate services for selecting near-optimal
compositions. To perform this classification we use an Associative Classification technique, CBA
algorithm.
3.2.1. Classification Overview
CBA algorithm is based on Associative Classification which is the integration of two important
data mining techniques, Classification rule mining and association rule mining. Associative
Classification is a novel strategy performing classification where model of classification is based
on a set of association rules in which consequent of each rule is restricted to contain only class
attribute values.
Some definitions of associative classification are introduced here:
Definition 1 (Training Data Set ‫.)ܦ‬ If an object can be described by features ‫ܣ‬ଵ, ‫ܣ‬ଶ, ‫ܣ‬ଷ, . , ‫ܣ‬௠ and
each object belongs to some class in a finite set of classes ‫,ܥ‬ then training data set ‫ܦ‬ is a
collection of instances < ‫ݒ‬ଵ, ‫ݒ‬ଶ, ‫ݒ‬ଷ, . , ‫ݒ‬௠, ܿ > where ‫ݒ‬௜ ranges over the domain of feature vector
‫ܣ‬௜ and ܿ represents the class of the object. As a preprocessing step any continuous valued
attribute need be discretized.

6
Definition 2 (Future Object ܱ). An instance < ‫ݒ‬ଵ, ‫ݒ‬ଶ, ‫ݒ‬ଷ, . , ‫ݒ‬௠ > whose class label is not known
will be called future object and classifier will be used to predict its class.
Definition 3 (Class Association Rules: CARs). CARs will be mined from Training data set ‫.ܦ‬ A
special subset of association rules in which the precedent of rule contains the form of attribute
value pairs and the consequent is restricted to take only class attribute value. Like association
rules, CARs are also subjected to a minimum support and confidence threshold values. For
instance a rule ‫,ݎ‬ ‫ܣ‬௜ ‫ݒ‬ଵ, ‫ܣ‬௝ ‫ݒ‬ଶ, … ! ܿ௞ is a CAR and supp(‫,)ݎ‬ conf(‫)ݎ‬ respectively denote support
and confidence of rule.
CBA algorithm is an ordered rule algorithm based on coverage analysis. It consists of two parts
Rule Generator and Classifier Builder. The goal of rule generator part is to mine all CARs from ‫ܦ‬
that their support and confidence are above minimum threshold values. Rule generation approach
is based on Apriori [12]. Classifier builder uses class association rules (CARs). The goal is to
short list high confidence rules from CARs to form a classifier. CARs are sorted in the
descending order of their precedence. We call this set of CARs ܴ. To build the classifier we select
high precedence rules ‫ݎ‬௜ in ܴ to cover ‫.ܦ‬ Finally classifier takes the form
< ‫ݎ‬ଵ, ‫ݎ‬ଶ, ‫ݎ‬ଷ, . , ‫ݎ‬௠, ݂݀݁ܽ‫ݏݏ݈ܽܿ_ݐ݈ݑ‬ >. ݂݀݁ܽ‫ݏݏ݈ܽܿ_ݐ݈ݑ‬ is the most probable class for all items. If all
rules in the classifier are unable to classify a future object then it will be classified according to
default class.
3.2.2. Defining QoS Levels
In order to classify candidate services, first we need to determine the QoS levels. To this regard
we define ݊ QoS levels, where the value of ݊ is defined with respect to the service density and the
expected accuracy for producing service classes.
Further the training data set is defined by the expert and the number of classes is the number of
QoS levels and the attributes are requested QoS attributes defined by the user.
Each class defines the attributes which their values have a defined distance from the range of the
values that specified from the user’s request. (i.e., the first class shows the services which the
value of their QoS attributes are in the same range with the user’s demand and the second class
shows the services which the value of their QoS attributes are one level lower than the user’s
demand and so on).
3.2.3. Computing the Service Utilities
The objective of our proposed algorithm is selecting the best services for the near-optimal
compositions. Indeed, having a large number of choices for services during dynamic binding
increases the number of alternative service compositions and subsequently a large number of
compositions prevent the starvation in dynamic service environments.
In this regard, we define a utility function ܷ௜ which shows the relative importance of a candidate
service ܵ௜. This function is computed based on two parameters: the first one is the QoS level
which the service is classified in it through the classification, it represents the importance of the
service (i.e., if it belongs to a class that its attributes are closer to the user demands, it will be
better choice for selecting). The second one is the overall quality of the service. Concerning to
this parameter, the service with the higher quality has the higher ability to be selected.
This function is computed as follows:

7
ܷ௜ = ‫ݏݏ݈ܽܥ‬௜ ∗ ‫݁ݒܣ‬௜ Where ‫݁ݒܣ‬௜ =
∑ ொ೔,ೕ
ᇲ೙
ೕసభ
௡
(3)
Where ‫ݏݏ݈ܽܥ‬௜ is the coefficient of the class to which service ܵ௜ belongs. ‫݁ݒܣ‬௜ is the total quality
of service ܵ௜, which computed as the average of the normalized QoS attributes values.
3.3. Ranking Based on Functional Aspect
The aim of this phase is selecting the best services for the near-optimal compositions such that 1)
consider the best quality regard to the user’s demand and 2) the composition of these services has
high functionality.
3.3.1. Considering the Most Eligible Services
In order to select services for the near-optimal compositions we use the utility function that we
have computed from the result of the classification part. So we do not consider all the possible
combinations of services. In this regard, we just mention the candidate services that have a utility
value over a defined threshold. This threshold allows us to focus on the most qualified services to
our regard.
For defining threshold we consider two aspects: the number of required compositions and the
execution time of the algorithm. Indeed, if we decrease the value of this threshold, the number of
considered candidate services decreases. Consequently, the number of compositions and the
execution time of the algorithm decreases respectively. Tuning the trade-off between these two
aspects will make our algorithm adaptable; hence it could be applied to multiple dynamic service
environments according to their constraints.
3.3.2. Functional Aspect of the Composition
Unlike most of approaches which just focus on the quality of composition by means of non
functional parameters (i.e. QoSs), the quality of semantic links can be considered as a
distinguishing functional criterion for semantic web service compositions.
Here we focus on the functional level of composing the candidate web services. The functional
criteria of semantic link, was introduced for the first time in [13] which defined as a semantic
connection between an output of a service and an input parameter of another service. Since the
qualities of these connections are valued by a semantic matching between their parameters,
semantic link composition could be estimated and ranked as well. Through the results of these
estimations some compositions are inappropriate. Indeed a composite service which does not
provide acceptable quality of semantic links might be useless as a service that does not provide
the desired functionality. Indeed the semantic connection between Web services is considered as
essential to form new value-added Web services.
Here we address the problem of optimizing in service selection with respect to this functional
criterion. In other words, we focus on the aspects of selecting a set of appropriate service
candidates for each task. We define an objective function In order to consider this aspect,
preferences and constraints which are defined by end-user.
3.3.2.1. Semantic Links
In semantic web, input and output parameters of web services are concepts referred to an
ontology ܶ, where the OWL-S profile [14], SA-WSDL [15] or WSMO capability [16] can be
used to describe them (through semantic annotations).

8
At functional level, web service composition consists of retrieving some semantic links between
output parameters ܱ‫ݏ_ݐݑ‬௜ ∈ ܶ of service ‫ݏ‬௜ and input parameters ‫ݏ_݊ܫ‬௝ ∈ ܶ of other service ‫ݏ‬௝.
Such a link i.e., semantic link, ‫݈ݏ‬௜,௝ (Figure 2) between two functional parameters of ‫ݏ‬௜ and ‫ݏ‬௝ is
formalized as 〈‫ݏ‬௜, ்ܵ݅݉൫ܱ‫ݏ_ݐݑ‬௜, ‫ݏ_݊ܫ‬௝൯, ‫ݏ‬௝〉. Thereby ‫ݏ‬௜ and ‫ݏ‬௝ are partially linked according to a
matching function ்ܵ݅݉. This function expresses which matching type is employed to chain
services.
The range of ்ܵ݅݉ is reduced to the four well known matching types:
• Exact If the output parameter ܱ‫ݏ_ݐݑ‬௜ of ‫ݏ‬௜ and the input parameter ‫ݏ_݊ܫ‬௝ of ‫ݏ‬௝ are
equivalent; formally, ܶ ⊨ ܱ‫ݏ_ݐݑ‬௜ ≡ ‫ݏ_݊ܫ‬௝.
• PlugIn If ܱ‫ݏ_ݐݑ‬௜ is sub-concept of ‫ݏ_݊ܫ‬௝; formally, ܶ ⊨ ܱ‫ݏ_ݐݑ‬௜ ⊆ ‫ݏ_݊ܫ‬௝.
• Subsume If ܱ‫ݏ_ݐݑ‬௜ is super-concept of ‫ݏ_݊ܫ‬௝; formally, ܶ ⊨ ‫ݏ_݊ܫ‬௝ ⊆ ܱ‫ݏ_ݐݑ‬௜.
• Intersection If the intersection of ܱ‫ݏ_ݐݑ‬௜ and ‫ݏ_݊ܫ‬௝ is satisfiable; formally, ܶ ⊭
ܱ‫ݏ_ݐݑ‬௜ ⊓ ‫ݏ_݊ܫ‬௝ ⊆ ⏊.
• Disjoint Otherwise ܱ‫ݏ_ݐݑ‬௜ and ‫ݏ_݊ܫ‬௝ are incompatible i.e., ܶ ⊨ ܱ‫ݏ_ݐݑ‬௜ ⊓ ‫ݏ_݊ܫ‬௝ ⊆ ⏊.
Figure 2. A Semantic Link ‫݈ݏ‬௜,௝ between ‫ݏ‬௜, ‫ݏ‬௝.
3.3.2.2. Semantic Link Quality
Several candidate services are grouped together in every task of an abstract composition. A
method to differentiate their semantic links is to consider their different functional quality criteria.
In this way, we use the semantic link quality model introduced in [17].
Here We consider the matching quality as the quality criteria for the semantic links ‫݈ݏ‬௜,௝ defined
by 〈‫ݏ‬௜, ்ܵ݅݉൫ܱ‫ݏ_ݐݑ‬௜, ‫ݏ_݊ܫ‬௝൯, ‫ݏ‬௝〉. The Matching Quality ‫ݍ‬௠ of a link ‫݈ݏ‬௜,௝ is a value in (0, 1]
defined by ்ܵ݅݉൫ܱ‫ݏ_ݐݑ‬௜, ‫ݏ_݊ܫ‬௝൯ i.e., 1 for Exact matching type,
ଷ
ସ
for PlugIn,
ଵ
ଶ
for Subsume and
ଵ
ସ
for Intersection.
The Disjoint match type is not considered since ܱ‫ݏ_ݐݑ‬௜ ⊓ ‫ݏ_݊ܫ‬௝ is satisfied.
Given the above quality criteria, the quality vector of a semantic link ‫݈ݏ‬௜,௝ is defined as follows:
‫ݍ‬൫‫݈ݏ‬௜,௝൯ = ቀ‫ݍ‬௠൫‫݈ݏ‬௜,௝൯ቁ (4)
In case of services ‫ݏ‬௜ and ‫ݏ‬௝ related by more than one semantic link, the value of each criterion is
retrieved by computing their average.

9
3.3.2.3. Generation of the Search Graph
Our algorithm uses the structure of a graph for selecting the most appropriate candidate services
for the composition. Moreover in this graph we use a priority queue to exploring the candidate
services in a defined order corresponding to their importance of QoSs and semantic similarities
with other services. These priority queues reduce the time spend for building and traversing the
search graph by inclusion the unexplored candidate services in themselves. In addition they can
increase the speed of search to acquiring the optimal composite services according to their
importance, by visiting the first services of each queue. Hence this structure will decrease the
computational complexity of traversing the graph and memory usage in the cases which the
abstract service compositions comprises a large number of tasks or there are many candidate
services for each task.
This graph is built from the candidate services according to the following rules:
• Each node represent a task in the abstract service composition;
• The order of these nodes is the corresponding order in the composition plan (i.e., if there
is a link from task ܶ௫ to ܶ௬ in the abstract service composition, then the corresponding
node of ܶ௬ will be after the corresponding node of ܶ௫);
• These nodes are made of a priority queue which consists of all the service candidates for
the corresponding task which its ܷ௜ is above the defined threshold as we have introduced
before;
Candidate services for each task go to the priority queue in this order:
• For the tasks which do not have any incoming task in the composition plan, web services
stay in the queue based on their utility values, i.e. the service with the highest utility value
(ܷ௜) stays at the top of the queue;
• For the other tasks, web services stay in the queue based on the following formula:
‫ܨ‬௜ = ܷ௜ ∗ ‫ݍ‬൫‫݈ݏ‬௜,௝൯ (5)
Where ‫ܨ‬௜ is the final utility value for each candidate service based on their QoS values
that respects the user needs and the matching quality of their semantic links between
them and their preceding service;
To build the search graph, first of all we add the nodes correspond to the tasks which do not have
any incoming task. Then we select a web service which is at the top of its queue. To select the
services for the other nodes, we compute the matching quality of semantic links between the
selected services and all the candidate services for the subsequent nodes then based on these
values and the utility values (which was computed in the local selection part), stay the candidate
services for these nodes in the priority queue based on their ‫ܨ‬௜ values. As the previous nodes, the
service which is top of the queue is selected as the chosen service. And so on the other services
for each node will have selected which the candidate services of each node stay in their
corresponding queue based on the value of their ‫ܨ‬௜.
The selected services from top of the queues form the near-optimal composite service. This
composite service is the first service that have the highest importance and unexplored services in
each queue are the candidates for alternative composite services which when the services in the
first composite service are no longer available or their QoS decreases (e.g., due to network
disconnection or weak network connectivity) during the execution of the composition, could be
replace with it.

10
The introduced structure of the search graph, allow for producing all the alternative composite
services by selecting the unexplored candidate services in each queue. But on the other side,
producing all the alternative composite services needs more computation so it is a time
consuming process (for calculating semantic similarities between services), also saving these
composite services need high memory, hence for tuning the trade-off between memory usage and
the execution time of the algorithm, we produce only the first alternative composite service which
is the best alternative.
We also propose a method for the case when a service which takes part in the composite service
is no longer available, we could propose another composite service by replacing the unavailable
service with another candidate service in its corresponding node of the search graph. In this case,
for switching to another composite service we must rearrange the priority queue in that
corresponding node, based on new utility values. This utility value is as the same as the ‫ܨ‬௜
function but the quality of the semantic links will be calculated as the follow:
At first we calculate the semantic quality of the connection between the selected service in the
previous node with the services for this node and the semantic quality of the connection between
the selected service in the following node with the services for this node, then the new semantic
quality value for the new ‫ܨ‬௜ is computed based on the average of these two computed quality for
each service in this node.
4. EXPERIMENTAL EVALUATION
4.1. Experimental Settings
We conducted experiments on an Intel(R) Core(TM) Duo CPU, 2.53GHz, 4GB RAM and a
windows 7 operating system. In this experiment we focus on the execution time of our method.
This metric measures the response time of our algorithm with respect to the size of the problem in
terms of the number of activities and the number of services per activity. In these experiments, we
measure the execution time of the ranking phase.
For the quantity allocation of the service qualitative parameters we use the QWS real dataset [18,
19]. This dataset includes measurements of 9 QoS attributes for 2500 real web services. The
dataset was measured using commercial benchmark tools for web services, which were located
using public sources on the Web, including UDDI registries, search engines and service portals.
In these studies, the authors provide a set of QoS metrics (i.e., response time, throughput,
availability, validation accuracy, reliability). We use these metrics as a sample input data for our
algorithm.
To accomplish the classification, we used the CBA1
tool for implementing the classification
algorithm. This tool has a graphical user interface and is designed particularly for implementing
the CBA algorithm. Hence for this part we do not measure the response time.
For classifying the candidate services we have defined 3 different quality levels. These levels are
specified to the distance of the QoS attributes to the user demands.
After the classification carried out, the classes’ coefficient were defined as follow: 1 for services
in the first class,
ଷ
ସ
for services in the second class and
ଵ
ସ
for services in the third class.
In general, on Contrary to QoS given by providers, the quality of semantic links are estimated
according to DL reasoning. Matching quality of each semantic link has been inferred in a pre-
1
http://www.comp.nus.edu.sg/~dm2/

11
processing step of semantic reasoning. Standard DL reasoning inference is achieved by means of
a DL reasoner Fact++2
[20].
For the purpose of these experiments, we vary the number of activities and the number of services
per activity between 10 and 50. The number of QoS constraints is fixed to 4 constraints (i.e.,
Response time, Availability, Throughput and Reliability) and for the sake of precision each
experiment is executed 20 times and finally we calculate the mean value of the obtained results
for our evaluation.
4.2. Performance Results
In figure 3 we demonstrate the execution time of ranking phase for the calculated near-optimal
composite service. These measurements are obtained by fixing the number of QoS constraints to
4 and varying the number of activities and the number of service candidates per activity between
10 and 50. The obtained measurements show that the execution time of our algorithm increases
along with the number of activities and the number of services per activity, which is an expected
result.
Figure 3. Execution time of the ranking phase for near-optimal composite service
Also figure 4 demonstrates the execution time of ranking phase for the first alternative composite
service. The increment of the execution time compared with the figure 3 is due to the increment
of the number of tasks and candidate services so the spending time for calculating semantic
similarity of output-input concepts increases subsequently. DL reasoning is the most time
consuming process in large-scale problem of quality-driven semantic web service composition
(i.e., number of tasks and candidate services greater than 100 and 350).This is caused by the large
number of potential semantic links between tasks.
2
http://owl.man.ac.uk/factplusplus/

12
Figure 4. Execution time of the ranking phase for first alternative composite service
As the figure 3 and 4 show, our method has a good execution time compared with the
evolutionary algorithms which the later algorithms (e.g., genetic algorithms, PSO) take a long
time to execute composite service (e.g., for a number of activities more than 25, it takes 2847 and
2236 ms, respectively for the PSO and GA algorithms [21]).
Moreover unlike our proposed method which aims to produce the composite services in an
ordered way, the evolutionary algorithms produce the composite services in an unordered way
and there is no evidence that the first composite service which produced by these algorithms be
the optimal one, in other words the next produced composite services may be better than the first
one.
5. CONCLUSION
As the number of the services published over the internet is growing at a very fast pace, selecting
satisfactory web services among the candidate web services which provide the same
functionalities is difficult. QoS is considered as the most important non-functional criterion for
further filtering services. Besides this criterion, the quality of semantic links can be considered as
a distinguishing functional criterion for semantic web service selection and composition.
We address the web service selection problem by defining constraints within a quality model to
balance QoS metric with functional quality. The functional quality evaluates the quality of
semantic links between the semantic description of output-input parameters of web services,
while QoS focuses on the non-functional criterion to retrieve the satisfactory services regard to
the user’s requirements and preferences.
Our objective has been to address service selection in the context of a QoS-aware middleware for
dynamic service environments. To do so we have proposed an approach which at first by using
the CBA algorithm, aims to classify the candidate web services to different QoS levels,
differentiate the services within each class, respect to their distances from the user’s demand for
the QoS criterion. By this classification a utility value is defined for each service that shows its
relative importance. In the next part, by the use of the structure of a graph, we have ranked the
candidate services by defining the final utility of each service considering the results of the
classification phase and their functional quality through measuring the semantic similarity

13
between their output-input concepts. Finally by producing alternative composite services in an
ordered way, we could cope with changing conditions of dynamic service environments.
Our proposed approach has four advantages: First, it uses the classification data mining algorithm
to specify the most eligible services respect to the user demand. Applying data mining algorithms
to this field brings new ideas. Second, by producing alternative composite services satisfying QoS
constraints, our algorithm could cope with changing conditions in dynamic service environments.
Third it shows a satisfying capability in terms of execution time, which it is an important point in
dynamic service environments and finally with applying the semantic similarity between services
by semantic links it increases the accuracy of selection.
Our proposed method makes part of our ongoing research by using strong data mining algorithm
in order to decrease the execution time and improve the optimality of our heuristic algorithm, and
besides the local optimization, by combining the global constraint of the user to the local
constraints, considering the QoS constraints through the whole composition to ensure meeting
global QoS constraints.
REFERENCES
[1] M. P. Papazoglou, P. Traverso, S. Dustdar, and F. Leymann, “Service-Oriented Computing: State of
the Art and Research Challenges,” Computer, vol.40(11), November. 2007, pp. 38-45.
[2] A. Benveniste, “Composing Web Services in an Open World: QoS Issues,” Proc. 5th Int. Conf. on
Quantitative Evaluation of Systems, 2008, p.121, doi:10.1109/QEST.2008. 49.
[3] P. M. Garey and D. Johnson, Computers and Intractability: a Guide to the Theory of NP-
Completeness, W.H. Freeman, 1979.
[4] B. T. Yu and K.-J. Lin: “Service Selection Algorithms for Composing complex Services with
Multiple QoS Constraints”. Proc. 3rd International Conference on Service-oriented Computing
(ICSOC’05), Springer, Amsterdam, Netherland, Dezember 12-15, 2005, pp. 130-143.
[5] M. C. Jaeger, G. Muhl and S. Golze: “QoS-aware composition of Web services: a look at selection
algorithms.” Proc. IEEE International Conference on Web Services (ICWS’05), IEEE Computer
Society, Orlando, Florida, USA, July 11- 15, 2005, pp. 800-808.
[6] T. Yu, Y. Zhang and K-J. Lin. “Efficient Algorithms for Web Services Selection with End-to-End
QoS Constraints,” ACM Trans. Web, 1(1):6, 2007.
[7] M. C. Jaeger and G. Muሷhl, “QoS-based Selection of Services: The Implementation of a Genetic
Algorithm,” In Torsten Braun, Georg Carle, and Burkhard Stiller, editors, Kommunikation in
Verteilten Systemen (KiVS 2007) Industriebetrage, Kurzbeitrage und Workshops, March 2007, pp.
359-350, Bern, Switzerland, VDE Verlag, Berlin und Offenbach.
[8] G. Wu, J. Wei, X. Qiao, L. Li, “A Bayesian network based Qos assessment model for web services,”
Proc. IEEE International Conference on Services Computing (SCC 2007), 2007, pp. 498-505.
[9] N. B. Mabrouk, S. Beauche, E. Kuznetsova, N. Georgantas and V. Issarny, “QoS-aware Service
Composition in Dynamic Service Oriented Environments,” Proc. Middleware, 2009, pp.123-142.
[10] B. Liu, W. Hsu, and Y. Ma, “Integrating classification and association rule mining,” Proc. Knowledge
Discovery and Data Mining, 1998, pp. 80-86.
[11] Z. Liangzhao, B. Boualem, H.H. Anne, Ngu, D. Marlon, K. Jayant and C. Henry, “QoS-Aware
Middleware for Web Services Composition,” IEEE Trans, Softw. Eng, 2004, pp. 311-327.
[12] R. Agrawal and R. Srikant, “Fast algorithms for mining association rules,” Proc. Int. Conf. on Very
Large Databases, VLDB-94, SanDiago, Chile, 1994, pp. 487-499.
[13] L. Freddy and L. Alain, “A formal model for semantic web service composition,” Proc. ISWC, 2006,
pp. 385–398.
[14] A. Ankolenkar, M. Paolucci, N. Srinivasan and K. Sycara, The OWL-S coalition, OWL-S 1.1.
Technical report, 2004.
[15] J. Kopecký, T. Vitvar, C. Bournez and J. Farrell, “SAWSDL: Semantic annotations for WSDL and
XML schema,” Proc. IEEE Internet Computing, 2007, pp. 60–67.
[16] D. Fensel, M. Kifer, J. de Bruijn and J. Domingue. Web Service Modeling Ontology, W3C
submission. 2005.

14
[17] L. Freddy and M. Nikolay, “Towards Scalability of Quality Driven SemanticWeb Service
Composition,” Proc. IEEE International Conference on Web Services, 2009, pp. 469-476.
[18] E. Al-Masri, and Q. H. Mahmoud, “Discovering the best web service,” (poster), Proc. 16th
International Conference on World Wide Web (WWW), 2007, pp. 1257-1258.
[19] E. Al-Masri and Q. H., Mahmoud, “QoS-based Discovery and Ranking of Web Services,” Proc. IEEE
16th International Conference on Computer Communications and Networks (ICCCN), 2007, pp. 529-
534.
[20] I. Horrocks, “Using an expressive description logic: Fact or fiction?,” In KR, 1998, pp. 636–649.
[21] C. Ming, W. Zhen-wu, “An Approach for Web Services Composition Based on QoS and Discrete
Particle Swarm Optimization,” Proc. IEEE Eighth ACIS International Conference on Software
Engineering, Artificial Intelligence, Networking and Parallel/Distributed Computing, 2007, pp. 37-41
Authors
M. Makhlughian received her B.Sc. degree in Software Engineering, from the Azad
University-Parand Branch, Tehran, Iran in 2009 and her M.Sc. degree in Software
Engineering from the Azad University-South Tehran Branch, Tehran, Iran in 2012. Her
research interests include Web Service technology, Service-Oriented Computing, Service
Annotation, Selection and Composition, Ontology Engineering and Data Mining.
Dr. S. M. Hashemi received his M.S. degree in Computer Science from Amirkabir
University of Technology (Tehran Polytechnic University) in 2003, and his PhD degree in
Computer Science from the Azad University in 2009. Moreover, he is currently a faculty
member at Science and Research Branch, Azad University, Tehran. His current research
interests include Software Intensive Systems, E-X systems (E-Commerce, E-Government,
E-Business, and so on), Global Village Services, Grid Computing, IBM SSME, Business
Modelling, Agile Enterprise Architecting through ISRUP, and Globalization Governance
through IT/IS Services.
Y. Rastegari is PhD. Candidate of Computer Engineering Department of Shahid
Beheshti University (SBU) in Tehran, Iran. He is also member of Automated Software
Engineering Research Group (ASER) and Information System Architecture Research
Center at SBU. His research interests include Service Oriented Architecture, Software
Architecture, Software Product Line and Ontology Engineering. He is working on
Service Choreography Adaptation challenge for Collaborative Business Processes.
E. Pejman received his B.Sc. degree in Computer Science and Software Engineering
from the Yazd University, Yazd, Iran in 2009 and his M.Sc. degree in Computer Science
and Software Engineering from the Azad University-South Tehran Branch, Tehran, Iran
in 2012. His research interests include web services, QoS-based service composition,
robotics and image processing.

WEB SERVICE SELECTION BASED ON RANKING OF QOS USING ASSOCIATIVE CLASSIFICATION

More Related Content

What's hot (12)

Similar to WEB SERVICE SELECTION BASED ON RANKING OF QOS USING ASSOCIATIVE CLASSIFICATION (20)

Recently uploaded (20)

WEB SERVICE SELECTION BASED ON RANKING OF QOS USING ASSOCIATIVE CLASSIFICATION