QoS based Web Service Selection and Multi-Criteria Decision Making Methods

With the continuing proliferation of web services offering similar efficacies, around the globe, it has become a challenge for a user to select the best web service. In literature, this challenge is exhibited as a 0-1 knapsack problem of multiple dimensions and multiple choices, known as an NP-hard problem. Multi-Criteria Decision Making (MCDM) method is one of the ways which suits this problem and helps the users to select the best service based on his/her preferences. In this regard, this paper assists the researchers in two conducts: Firstly, to witness the performance of different MCDM methods for large number of alternatives and attributes. Secondly, to perceive the possible deviation in the ranking obtained from these methods. For carrying out the experimental evaluation, in this paper, five different well-known MCDM methods have been implemented and compared over two different scenarios of 50 as well as 100 web services, where their ranking is defined on an account of several Quality of Service (QoS) parameters. Additionally, a Spearman’s Rank Correlation Coefficient has been calculated for different pairs of MCDM methods in order to provide a clear depiction of MCDM methods showing the least deviation in their ranking. The experimental results comfort web service users in conquering an appropriate decision on the selection of suitable service.


I. Introduction and Background
A web service is a self-describing software application which can be advertised, located and used over the Internet [5]. To initiate the provisioning of web service, a user must first identify the service that is desired. "How will I effectively select the web service that will meet my performance requirements?" [1]; the answer to this question, however, still remains a challenge faced by the user because of numerous services sharing similar functionalities in the web environment. The rise in the number of Web Services has been caused due to growing demands of increasing the flexibility of IT infrastructure in order to support rapidly evolving business needs [2]. The selection of web services is not only limited to meeting the users' needs, but also, non-functional information, including, reliability, response time, etc. [3].

A. Motivation
A Web Service can also be viewed as one of the encouraging technologies that could help business units to systematize their web operations on a large scale by automatic discovery and consumption of services [5]. In the Web Service Architecture [6], the Service Requester (client or user) may receive a pool of web services from the Service Provider (server) as per the initial query in "Service Discovery" stage. Subsequently, in "Service Selection" stage, the "best" web service, which satisfies all the constraints set by the original requester, is selected from the pool. This process of service filtration is carried out based on the degree of satisfaction to the users' non-functional requirements known as Quality of Service (QoS) parameters. For example, while booking flights or downloading music, there exists a number of available services sharing identical functionalities, however, they exhibit different QoS. A web service with remarkable QoS can deliver big competitive influence to service providers while bringing the social prosperity to service consumers. It has been acknowledged from the literature that due to intensive global competition, the experts recognize the decision on web service selection an important activity.

B. Quality of Service (QoS) Parameters for Web Service Selection
The QoS based Web Service selection has gained the attention of many researchers in recent years, since maintaining the quality of their web services has become the topmost priority of each web service provider. In this paper, therefore, the activity of web service selection is carried out based on both functional as well as non-functional QoS parameters [4,11]. QoS attributes are measured on a scale of 0 to 9. To facilitate the description, the set of QoS attributes is divided into two subsets: Benefit (Positive) attributes and Cost (Negative) attributes, as shown in Table I. The values of positive attributes need to be maximized, whereas the values of negative attributes need to be minimized. For the sake of simplicity, the values closer to 9 for benefit criteria and closer to 1 for cost criteria are considered good. methods have proven their effectiveness in addressing different complex real-world decision making problems. MCDM methods use knowledge from many fields, including economics, mathematics, behavioral decision theory, software engineering, computer technology and information systems [9]. In the context of this paper, the goal of MCDM methods is to find one web service from a pool of several web services such that the QoS is optimized and users' end-to-end QoS requirements are satisfied. In MCDM methods, each problem is stated in matrix design as: where, A 1 , A 2 , … A m are possible 'm' alternatives or choices among which decision makers have to choose, C 1 , C 2 , ..C n are possible 'n' criteria or attributes the basis of which the alternatives are ordered or selected, x ij is the rating of alternative A i with respect to criterion C j .
The general flowchart of MCDM method is shown in Fig 1. The sensitivity analysis is done to address the problems of uncertainty, imprecision, and inaccurate determination [10]. Various MCDM methods used for web service selection in the literature are briefly summarized in Table II, however Tables III and IV conveys

D. Objectives of Present Research
The objectives of present study can be stated as follows: • To address the problem of web service selection based on QoS.
• To study how service consumers are benefited by selecting the appropriate Web Service based on QoS using MCDM in literature.
• To implement and compare five different MCDM methods (AHP, TOPSIS, SAW, VIKOR and COPRAS) with two different large sets of web services (50 and 100) against 9 QoS attributes.
To evaluate the deviation in the rankings of different MCDM methods using Spearman's Rank Correlation Coefficient and to conclude which MCDM methods produce similar ranking.

E. Organization of Paper
The remainder of this paper is organized as follows: Section 2 throws light upon the dataset and several methods used for present study. Section 3 discusses the results from experimental evaluation. Finally, the conclusion and highlights on possible continuations of this work are addressed in Section 4.

A. Dataset Used
The Quality of web service (QWS) dataset version 2.0 [12,43] is chosen for the present study. This dataset includes a set of 2,507 web services and their QWS measurement which were conducted in 2007, using Web Service Broker (WSB) framework. Each row in the dataset represents a web service and its corresponding QWS measurements on nine different QoS parameters. In this paper, two different sets of 50 and 100 web services have been constructed by random sampling for the experimental study of various MCDM methods.

B. Various MCDM Methods
MCDM methods help in selecting the optimal one from a set of alternatives with respect to the predefined set of attributes. MCDM methods are continuously growing in the application areas of Business, Mathematics, Decision Sciences, Management and Accounting, Social Sciences, Medicine, Environmental Science, Economics, Econometrics and Finance, etc. [19]. In this paper, for QoS based web service selection, five different existing MCDM methods are taken into account: Simple Additive Weighting (SAW) [15], Analytic Hierarchy Process (AHP) [14], Technique for Order Preference by Similarity to Ideal Solution (TOPSIS) [9], VIšekriterijumsko KOmpromisno Rangiranje (VIKOR) [17] and COmplex PRoportional ASsessment (COPRAS) [18]. These methods are summarized in Tables III and IV. The general formal definition of MCDM is represented in (1):

C. AHP (Analytical Hierarchy Process)
Suppose J represents the number of criterion and a jk represents the importance of j th criterion relative to k th criterion in PC matrix. The AHP consists of following steps [14]: i) Breaking down of a problem into a hierarchy of decision criteria and alternatives.
ii) Relative importance of each criterion is measured with respect to other criterion, which is known as Pairwise Comparison PC process. The consistency checks are also performed while the evaluations are made by the decision makers.
iii) Normalization of pairwise comparison matrix by using (2.1).
iv) The Criteria Weight Vector is built with the help of (2.2).    Suppose I represents the number of alternatives and J represents the number of criterion; x ij is the value assigned to i th alternative with respect to j th criterion. The TOPSIS consists of following steps [9]: i) Normalization of decision matrix using (3.1).
ii) Calculation of the weighted normalized decision matrix using (3.2).
v ij = w j . r ij (3.2) where W={w 1 ,w 2 ,…w J }=relative weight about the criterion iii) Determination of the positive ideal and negative ideal solutions which maximizes and minimizes the benefit criteria and cost criteria respectively using (3.3) and (3.4).
iv) Calculation of each alternative from positive ideal solution and negative ideal solution with the help of (3.5) and (3.6) respectively (Euclidean distance). vi) Ranking of alternatives is done as per the increasing order of relative closeness. Higher the relative closeness to ideal solution is, better is the alternative.

E. COPRAS (COmplex PRoportional ASsessment)
Suppose I represents the number of alternatives and J represents the number of criterion; x ij is the value assigned to i th alternative with respect to j th criterion; q j represents the significance of j th criterion. The steps of COPRAS are summarized as follows [18]: i) Normalization of decision matrix using (4.1).
ii) Calculation of Maximizing Criteria S + and Minimizing Criteria Susing (4.2) and (4.3) respectively. S +i = J' ∑ j=1 r ij . q j (4.2) where i=1,2,…I; j' is the number of benefit criterion Si = J'' ∑ j=1 r ij . q j (4.3) where i=1,2,…I; j" is the number of cost criterion iii) Relative weight of each alternative is obtained using 4.4.
iv) Ranking of alternatives is done according to the ascending order of relative weight. Higher the relative weight of alternative is, higher is the priority of alternative.

F. VIKOR (VIšekriterijumsko KOmpromisno Rangiranje)
Suppose I represents the number of alternatives and J represents the number of criterion; f ji is the value of j th criterion for the i th alternative The VIKOR involves the aforementioned steps [17]: where j =1,2,….J (5.4) iii) Computation of Q j using (5.5). -C2-Acceptable stability in decision making: the alternative a 1 must also be best ranked by S and R.
vi) If any of the conditions is not true, then the set of compromised solutions are proposed consisting of -If only C2 is not satisfied, then alternatives a 1 and a 2 are proposed.
-If only C1 is not satisfied, then alternatives a 1 , a 2 , …a m are proposed; a m is determined by the relation Q(a m )-Q(a 1 )˂DQ for maximum 'm' (the positions of these alternatives are "in closeness").

G. SAW (Simple Additive Weighting)
Suppose I represents the number of alternatives and J represents the number of criterion; x ij is the value assigned to i th alternative. The SAW method consists of the following steps [15]: i) Normalization of decision matrix using (6.1) and (6.2) respectively for cost and benefit criterion.
ii) Calculation of weighted normalized values is done using (6.3).

H. Spearman's Rank Correlation Coefficient
The Spearman's rank correlation coefficient method assists in finding the similarity between two sets of ranking obtained from two different 'k th ' and 'i th ' MCDM methods, using (7): where, n is the number of web services, and d i is the difference between the ranks of two MCDM methods. A larger absolute value indicates a good agreement between one MCDM method and other MCDM method [13].

III. Results and Discussions
For carrying out the experiments, two sets of web services are constructed out of present dataset: one set of 50 web services (Scenario 1) and other set of 100 web services (Scenario 2). Five different MCDM methods such as AHP, COPRAS, SAW, TOPSIS, and VIKOR

AHP
• Hierarchical structure of criteria.
• Pairwise comparison gives better comparisons of criteria.
• Gives option to evaluate quantitative and qualitative criteria and alternatives.
• If Hierarchical structure of criteria is not made properly user may get worst ranking. • In special cases (currencies exchange), it may not work. • Absolute zero doesn't exist.

TOPSIS
• Scalar value that accounts for both the best and worst alternatives simultaneously. • Sound logic that represents the thesaurus of human choice.
• Possess rank reversal problem. • Its use of Euclidian distance does not consider the correlation of attributes, difficult to weight and keep consistency of judgment.

COPRAS
• Degree of utility is the bases of ranking.
• It doesn't require transformation of cost and benefit criteria. It has complex aggregation procedure.

SAW
Simple to understand and implement.
Result not always real to situation.

VIKOR
It gives Ideal and compromised solution.
Complex to understand and implement.
are implemented in which the weights for different QoS attributes are calculated using AHP method in each scenario. The outline of experiments can be illustrated from Fig 2. The top five ranking of each method for each scenario is shown in Fig 3. In order to check the consistency of user inputs to QoS attributes, the matrix analysis is usually done for AHP [14,30], however, this effective technique of checking the evaluations made by the decision maker is used for each method while constructing the pairwise comparison matrices. The formula used for obtaining the Consistency Index is shown in (8). CI = ((λ max -n) / (n-1) (8) where, n is the matrix size and λ max is the eigen value.
Generally, the value CI = 0 is obtained by a perfectly consistent decision-maker, however, the smaller inconsistency may be tolerated.
In particular, the inconsistencies are tolerable and reliable results have been expected from each method, if (9) holds true.
CI / RI < 0.1 (9) where, RI is Random Index, means the Consistency Index when matrix has entries which are completely random [30], and the ratio CI/ RI is known as Consistency Ratio.  (Table V) AHP method shows 29 th , 50 th and 18 th numbered are the best three web services to be selected, whereas 21 st , 6 th and 38 th numbered are the worst three to be selected. Moreover, COPRAS method produces 29 th , 23 rd and 50 th as the best three web services and 38 th , 7 th and 27 th as worst three web services. On the other hand, using SAW method, 18 th , 40 th and 30 th numbered three best web services are obtained, whereas 21 st , 38 th and 6 th are worst. Furthermore, TOPSIS method gives best three ranks to 29 th , 23 rd and 50 th numbered web services, while worst three ranks to 7 th , 6 th and 38 th . Also, 30 th , 19 th and 24 th numbered are best three web services, while 32 nd , 36 th and 22 nd are worst three web services chosen by VIKOR method. (Table VI) AHP method shows 79 th , 29 th and 50 th numbered are the best three web services to be selected, whereas 38 th , 96 th and 50 th numbered are the worst three web services to be selected. While, COPRAS method produces 29 th , 23 rd and 79 th as best three web services and 64 th , 77 th and 27 th as worst three web services. Similarly, using SAW method, 18 th , 95 th and 40 th numbered three best web services are obtained, whereas 6 th , 96 th and 53 rd are worst. On one hand, TOPSIS method gives best three ranks to 29 th , 23 rd and 79 th numbered web services, on the flip side, worst three ranks to 96 th , 38 th and 53 rd . Also, 30 th , 56 th and 87 th numbered are best three web services, while 78 th , 22 nd and 73 rd are worst three web services chosen by VIKOR method.

Scenario 2: 100 Web Services and 5 MCDM methods
Clearly, in both the scenarios (1 and 2), rankings obtained from five different MCDM methods in the QoS based Web Service Selection problem yields divergent results. This difference in ranking can be seen due to either the use of different normalization techniques on decision matrix or the use of different aggregating methods in each MCDM method. Further, in order to evaluate the closeness of ranking, for both the scenarios, Spearman's rank correlation coefficients are calculated using (7).
The closeness of the correlation coefficient value (example, 0.9526), ranging between -1 to 1, in approximation to unity indicates complete dependency and reliability of either of the methods used. The dependency reduces with each unit reduction in the coefficient value. The negative sign indicates reverse trend existing between the two methods i.e. the rank value increasing in one method shall be declining under the second method in comparison. The inter-relationship between the MCDM methods is analyzed through correlation matrix using (7) as shown in Tables VII and VIII. It is found that AHP and TOPSIS show maximum Correlation value i.e. 0.9535 in Scenario 1 (Table VII) and 0.9526 in Scenario 2 (Table VIII), indicating the strongest correlation, as the values are generously high. Thus, it can be concluded that in this web service selection problem, AHP and TOPSIS can be used effectively for making similar types of decisions. All other combinations show positive correlation except COPRAS and VIKOR. These findings hold true for both scenarios of 50 as well as 100 web services.

IV. Conclusions and Future Scope
Nowadays, there is a need to distinguish increasing number of web services with similar functionalities, being made accessible across the Internet, using a set of QoS parameters. The QoS level displays abundant influence on degree of the web service usability as well as effectiveness, both of which further influences the service popularity. In this regard, the problem of web service selection based on QoS using MCDM method is addressed in this paper. Firstly, the rankings of web services are calculated using five different MCDM methods, including AHP, TOPSIS, VIKOR, SAW, and COPRAS. Secondly, since these rankings show divergent results, Spearman's rank correlation coefficient is used to compute the degree of similarity in the rankings of one MCDM method with other MCDM methods. The whole process is done for two larger sets of web services: 50 (Scenario 1) and 100 (Scenario 2). Maximum co-efficient correlation value is deduced for the combination of AHP and TOPSIS in both the scenarios. The experimental outcomes on different sets of web services using different MCDM methods reveal that AHP and TOPSIS methods show good agreement with each other. In future, the work can be extended using different correlation methods such as Pearson, Kendall or any other correlation method. The more focus can also be put on Rank Reversal problem of MCDM methods.