QoS-driven selection of web service considering group preference

Computer Networks 93 (2015) 111–124

Contents lists available at ScienceDirect

Computer Networks journal homepage: www.elsevier.com/locate/comnet

QoS-driven selection of web service considering group preference Hei-Chia Wang∗, Wei-Pin Chiu, Suei-Chih Wu The Institute of Information Management, National Cheng Kung University, Tainan, Taiwan

a r t i c l e

i n f o

Article history: Received 12 January 2015 Revised 22 August 2015 Accepted 19 October 2015 Available online 26 October 2015 Keywords: Web service Quality of service Group decision TOPSIS

a b s t r a c t The aim of this paper is to present an approach for selecting Web services during the process of building an orchestration/composition, on the basis of some Quality of service (QoS) features. The specific proposal is to consider decision maker’s preferences, and aims at guaranteeing consistency of the selection considering such preferences. The composite service has multiple tasks and each task can be fulfilled by a group of services. Then the optimization is to select a service for each task so that the overall evaluation function can be optimized. With the traditional group decision method, the Technique for Order Preference by Similarity to Ideal Solution (TOPSIS) process, the decision makers must construct the decision matrix. Once web service candidates become unavailable or new web services emerge, the alternatives change, which means that the decision makers must reset the large number of evaluated values. In this paper, a group selection process named Group Decision Web Service Selection, GDWSS, was proposed. First, weights for each task are collected from the different decision makers and global optimization parameters are generated. Next, an easy preference setting named WSS-TOPSIS, adapted from the TOPSIS method, is proposed to reduce the amount of preference data input. Furthermore, GDWSS allows group decision makers to set preferences with a lighter workload and to consider group preferences for the system to reduce the exception difference existing among decision makers. Finally, the scenarios are replicated to illustrate the effectiveness of this approach. © 2015 Elsevier B.V. All rights reserved.

1. Introduction Web services have shown increasingly more demands [1]. One of the reasons for this increase is that web services can be combined dynamically to fulfill complex requirements [2,3]. With so many web services available, the system integration from different companies or departments across the supply chain, under effective and careful consideration of participants’ needs, has become an important issue in system development [4–6]. However, many of their functions are similar. Besides, a single web service can provide only ∗

Corresponding author. Tel.: +88662757575#53724. E-mail addresses: [email protected] (H.-C. Wang), [email protected] (W.-P. Chiu), [email protected] (S.-C. Wu). http://dx.doi.org/10.1016/j.comnet.2015.10.014 1389-1286/© 2015 Elsevier B.V. All rights reserved.

limited functions, so it appears difficult to fully meet users’ complex needs [7]. Thus, it is preferable to combine web services that complement one another to satisfy complex needs [8]. In addition to selecting a web service oriented toward a single task process, global optimization should be considered to make web service composition more effective [9]. To solve this problem, Quality of service (QoS), in terms of such factors as execution time, price and reliability, is used as a selection criterion in web service selection [10]. But there are countless web service composition options and the QoS of composite web service is difficult to control [11]. Zeng et al. [12] contemplated the dynamic transformation of web services and selected web services using integer programming. Danilo and Pernici [13] further improved Zeng [12] to overcome the problem of confliction to the global limit occurring during the process of sub path merging. Huang, Lan and

112

H.-C. Wang et al. / Computer Networks 93 (2015) 111–124

Yang [14] computed the QoS value of web service compositions using a different computing method based on composition process type and selected a web service composition that was globally optimized. Chang Guofeng [15] applied Bayesian Networks to obtain each web service QoS value to satisfy users’ requirements. Lin, Lo, et al. [16] utilized the features of QoS, combining fuzzy methods to achieve group consensus. However, most web service selection research has assumed that the decision maker does not understand the process of web service composite structure development. Therefore, there is no difference in the importance of each task [17–20]. Furthermore, the abstract web service composition can be regarded as an enterprise operating process during which users should have differing degrees of importance to each operating process. Consequently, the composite web service should be selected by considering the differing importance of the process in each combination. There are many studies [11–14,21–27] presenting different frameworks and methods to assist users in selecting the desired web service by considering the opinions of IT staffs but not the preferences of different stakeholders. Web service composition should consider these preferences guardedly because each decision maker may have different preferences when selecting web services. To determine individual preferences during the web service selection process, Lo [28] proposed a method of group decision making with regard to the selection of single web services. Zhang et al. [29] proposed a method of group decision making for web service execution paths. Considering the dynamic changes that web service has experienced [12] and the various QoS domains, it will be more difficult to determine group web service preferences. To consider the needs of the majority during the process of web service composition, the Technique for Order Preference by Similarity to Ideal Solution (TOPSIS) is employed to select the final web service composition because in most cases, multiple users will participate in decision making [28,30,31]. The TOPSIS can aid decision makers in not only ranking alternatives but also generating feasible alternatives to address multiple attributes or multiple criteria [18,32]. However, two challenges need to be solved when employing TOPSIS: • This method is easily affected by outliers and imprecise or uncertain data, possibly leading disorder among alternatives [33,34]. The final solution is selected from a set of existing alternatives instead of after consideration of the decision makers’ opinions. Furthermore, decision maker must set the evaluation for each criterion for every alternative during the group decision process. The evaluation may be imprecise because there are too many alternatives and criteria [35]. • The decision makers must construct the decision matrix. When web service candidates become unavailable or new web services emerge, the alternatives will change, which means that the decision makers must reset the large number of evaluated values. To solve the abovementioned problems. A selection process named Group Decision Web Service Selection, GDWSS, was proposed, allowing different tasks to have individual domain-related attributes. These attributes can still be aggregated to obtain the domain-related attribute value of the

web service composition even if each task has different types of domain-related attributes. Alternative sorting allows for a significant degree of consensus, meaning that the group selection result can match the preferences of the majority of decision makers. With the overall GDWSS process, the final solution has not been selected from the default alternatives, but alternatives are generated by decision makers’ preferences and the final solution is chosen from the alternatives. This method has the following features. • Considers the global optimization of web service composition and the decision maker’s preferences simultaneously during the web service composition process. • Generates alternatives using the decision makers’ preferences and choosing the alternative closest to the group’s preferences. • Proposes the WSS-TOPSIS method, adapted from the TOPSIS, which replaces the evaluation of the decision maker with QoS to reduce the workload. The remainder of this paper is organized as follows. Section 2 includes a brief introduction to decision making methods and a review of the QoS. The method developed is detailed in Section 3. In Section 4, the method used to examine the proposed model is introduced. Finally, the research results and implications are presented and discussed on Section 5. 2. Literature review The aim of this paper is to present an approach for selecting Web services during the process of building an orchestration/composition, on the basis of some QoS features. The specific proposal is to consider decision maker’s preferences, and aims at guaranteeing consistency of the selection considering such preferences. 2.1. The methods of decision making Group decision-making is more complex than individual decision making due to different opinions. When making multiple-criteria decisions, many factors need to be considered, such as the selection of the lowest-cost, fastest, most reliable web service. But considerations will differ over time. Some multi-criteria decision making methods for web service selection were applied, including Integer Programming (IP), Simple Additive Weighting (SAW) and the TOPSIS. IP is a branch of linear programming. The most important feature of IP is that the variables are all integers because many of decisions and variables in real cases are inherently integers. Huang et al. [14], Meng et al. [36] and Zeng et al. [12] applied IP to compute the assessed value of web service compositions and select web services by optimizing the assessed value because this method is good for global allocation of tasks to services. This programing has three main elements, namely, decision variables, the objective function and constraints. The primary goal is to transfer the variables of the objective function to find the maximum or minimum values. However, IP problem may takes very long processing time for computers to find the optimal solution with the increase of problem size. For this reason, IP problem is limited

H.-C. Wang et al. / Computer Networks 93 (2015) 111–124

to use in real scenarios, especially in time-critical scenario [37]. SAW is known as the weighted sum method and is often used multi-attribute decision technique for classical multiple attribute decision making (MADM) [38]. It is simple and benefits from additive property for calculating final score of alternatives. In other words, each alternative is assigned a value which is the weighted sum of these attributes after alternatives are calculated by some attributes. SAW is utilized to select the optimal web service [12,14,36]. This method is divided into two steps. First, the attributes must be normalized to integrate attribute values with different units for computing or comparing. Second, attribute values are multiplied by the value of the attribute weights. The sum of attribute values is the assessment score of the object. But when information is not crisp anymore, SAW maybe hard to work [39]. TOPSIS method is proposed by Hwang [40]. This method is a multiple criteria method to deal with multi-attribute or multi-objective decision making problems effectively. This multiple criteria method can identify solutions from a finite set of alternatives [41]. The basic concept of the TOPSIS is that the distance of the chosen alternative between the positive ideal solution (PIS) and negative ideal solution (NIS). The stronger the similarity is, the more strongly the solution will meet the user’s preferences and a descending order from superior to inferior will be achieved. The PIS is the set of attributes with the best performance. In other words, the NIS is the set of attributes with the worst performance. Usually, no solution can be found that is the shortest possible distance to the PIS and the farthest possible distance from the NIS. The selection processes for the following steps are presented in [42]. 1. Construct the decision matrix: there are m alternatives and n attributes, and xij is the assessed value of i-th alternative under the j-th attribute. The decision matrix is X=(xij )mxn . 2. Normalization: get rij by vector normalization is given x

as ri j = mi j

2 i=1 xi j

, i = 1, . . . , m; j = 1, . . . , n.

3. Weight: multiplied r by the corresponding attribute weights is given as vi j = w j × ri j , i = 1, . . . , m; j = 1, . . . , n. 4. Define the positive ideal solution, PIS, A∗ , negative ideal solution, NIS, A− . Prior to doing this, we must define the benefit attribute first (J1 is the set of the benefit attributes) and the cost attribute (J2 is the set of the cost attributes) are given as

(v∗1 , . . . , v∗j , . . . , v∗n ) = {(max vi j | j ∈ J1 ), (min vi j | j ∈ J2 )|i = 1, . . . , m}

A∗ =

i

A− = =

i

(v−1 , . . . , v−j , . . . , v−n ) {(min vi j | j ∈ J1 ), (max vi j | j ∈ J2 )|i = 1, . . . , m}, i

i

5. Using Euclidean distance to computing the distance from the alternative to the positive ideal solution(Si∗ ) and the distance from the alternative to the negative ideal solution(Si− ) are

given as Si∗ =



n j=1

113



n j=1

(vi j − v∗j )2 ,i = 1, . . . , m., Si− =

(vi j − v−j )2 ,i = 1, . . . , m.

6. Computing the similarity between the alternative and the positive ideal solution is given as Ci∗ = Si− /(Si∗ + Si− ), i = 1, . . . , m. 7. Sorting alternatives by C∗i in descending order, the solution with the highest C∗i will be selected for the optimal solution. The TOPSIS has been widely used in recent years in various fields to allow groups to choose better solutions with multiple attributes [43]. It was established as one of the most appropriate methods of determining traversal rank and selecting web services using the fuzzy TOPSIS when a group of users have different opinions on evaluation. [28]. The fuzzy TOPSIS methods involve considerable number of arithmetical operations on fuzzy numbers to deal with imprecise information. Therefore, this method can reduce the computational complexity during the decision making process in an efficient way. 2.2. Web services with a QoS model The web service provider describes the functionality requirement, input and output and provides QoS-related information on the web Service using Web Services Description Language (WSDL) and publicizes the web service with the Universal Description Discovery Integration (UDDI) Registry [44], which is an intermediary party. Although there are many non-functional QoS attributes, there is currently no standardized description framework containing all aspects of a service’s non-functional characteristics [14]. In view of this issue, most relevant studies involve the selection of four specific QoS criteria for selecting a web service: response time, availability, execution price and reputation to be considered [45]. Commonly used QoS criteria are listed in [11,12,14,21,25,36,46] including response time, availability, execution price, reliability, reputation, throughput and data quality. For the sake of illustration, the adopted model has a limited number of criteria. On the other hand, these criteria can be replaced or added easily to extend the quality model without altering fundamentally the selection and composition methods [47]. Moreover, some specific QoS attributes maybe adopted in some specific domain [48]. The four QoS criteria of this paper are described briefly as follows. • Response time is to evaluate a service’s response time to a request typically comprises the measurement of the execution time and waiting time. • Availability is the degree to which a service is operational and accessible when it is required for use. • Execution price is the cost of service for a request. • Reputation is a measure of web service trustworthiness. It mainly depends on end user’s experiences of using the service; different end users may have different opinions on the same service. 2.3. Global optimization and web service selection The simplest method of selecting a web service composition is to choose the best web service for each task process.

114

H.-C. Wang et al. / Computer Networks 93 (2015) 111–124

However, this approach is only used for local optimization and the resulting global quality of the execution is not ideal [10,12]. Sometimes the methodology is used rather loosely because the selection method must consider global optimization [49–53]. Zeng et al. [12] presented an abstract web service composition using execution paths by which a web service was employed during an abstract task to formulate a service execution plan. IP was used to find the best composition configuration and the set of constraints was harnessed to determine the web service composition fitting the global restrictions. A critical path was used to compute the execution time for web service composition. The critical path task cannot be used to select a web service when considering other factors but not the time factor to choose the best service. Zeng [12] showed that although global optimization requires higher computational costs and more information to set user preferences, weights and restrictions, his study of global optimization allows for the selection of the web service composition with the best QoS. Ardagna and Pernici [13] improved Zeng [12] to resolve the problem of confliction to the global limit during the sub path merging process and exploited IP to choose the best web service composition in a dynamic environment. Huang et al. [14] and Sun, and Zhao [10] computed the QoS value of web service compositions using different computing methods (add, multiply, maximum, minimum) depending on the composition process type (sequence, parallel, synchronization, arbitrary cycles and so on.). Criteria and their importance were established using multi-objective decision-making and a weighted sum model (WSM). The QoS of the web service composition was calculated using IP and summed to obtain an aggregated value for non-split processes. Branch-and-bound was used to calculate the overall web service composition result for split processes. Canfora et al. [21] applied genetic algorithms to select web services and compared this method with the IP method to summarize the following conclusions: the genetic algorithm was superior to IP in that there were many candidates for each abstract service; however, this usually occurred when the function of the abstract services was simple and easy to develop, as with weather forecasts or currency conversions. Thus, the IP can be designed to quickly select for a small number of professional-oriented web services. Compared with other methods, although top-k ranking strategy is used to reflect a user’s personalized requirement, this method is not flexible than TOPSIS [24]. Zhang et al. [54] focus on the topic that one user in some specific composition scenarios instead of a group-oriented algorithm for Web service composition. Khanfir [55] integrate ranks classification without considering user’s preferences to classify selected services. After reviewing relevant literatures, some methods could be used to choose the compromise solution with multi-attributes, understanding their advantages and disadvantages. However, the studies above only considered system designers’ opinions and did not acknowledge the importance of different tasks. TOPSIS is adopted here for choosing solution because it’s theoretical rigorousness and is a sound logic that represents the human rationale in selection. From the QoS-related research works, some criterions are listed systematically usually used to select web services. These criterions can be added for choosing web services to satisfy decision-makers requirement. Therefore, a method was

Fig. 1. The concept of alternatives generation and selection.

Fig. 2. Group decision web service selection.

proposed to eliminate the problem of prior studies’ lack of consideration of differing tasks to improve the problem that the previous research did not consider. 3. Group decision web service selection The goal of the research is to propose a method of web service selection that considers group preference, as shown in Fig. 1. The generation of web service composition depends on decision makers’ preferences through consideration of the QoS of the alternatives, attribute weights and task weights to recommend the most acceptable solution. The ultimate product is a system that is composed of web services. As shown in Fig. 2, within the dotted rectangle is the Group Decision Web Service Selection (GDWSS) architecture. All notations are described in detail in Appendix A. • Decision Maker, DM: DMk is the kth decision makers responsible for providing preferences for web service composition, which is the main input source in the system. • Service provider: provided web services, described QoS attribute value and functions, and registered web services in UDDI.

H.-C. Wang et al. / Computer Networks 93 (2015) 111–124

t2

Table 1 The attribute weight for decision maker DMk .

t5

t1 t3

t4

t6

Fig. 3. Statechart service composition.

• Composition of abstract services: it is used to describe the functional requirement of users. • Services matchmaking: find the web services that meet the requirements of decision makers, which means there will be at least one web service for each abstract service node. • preprocessing: find the common attributes of the same domain and analysis candidates of Web services to get the QoS attribute values. • Preference setting: decision-makers set the importance of each abstract service process and each QoS attributes. • Computing alternatives: The WSS-TOPSIS method was proposed based on the method of TOPSIS. The assessed values of alternatives are generated from decisionmakers’ preference, the QoS attribute values, the weights of QoS attributes and the weight of tasks. The alternatives with the highest assessed value would be returned to the decision makers. The overall procedures in Fig. 2 is described in detail in Sections 3.1–3.5. The scenario was adapted from Zeng [12] to illustrate our selection process. Six tasks were included in the scenario: (t1 ) attraction searching, (t2 ) flight booking, (t3 ) hotel booking, (t4 ) geographic distance computing, (t5 ) bicycle rental and (t6 ) car rental, as shown in Fig. 3. S is the abstract services set; S is composited by N tasks, S:{t1 ,…,ti, …,tN } and wsi, j is the j-th physical web service for task ti , which means that at most, mi web services can be selected for task ti to finish the function; this set of web services is WSi :{wsi,1 , wsi,2 , wsi,3 ,…wsi,m }. 3.1. Preprocessing The purpose of preprocessing is to find the common attributes of the same domain and analyze Web service candidates to derive the QoS attribute values. The QoS is determined by the definitions of the service providers, which describe policies, including privacy and features of WSDL [19]. Part of the QoS can be monitored and collected by a third-party organization; for example, reputation and availability can be determined using UDDI. The QoS of time (CT ), price (CP ), reputation (CR ) and availability (CA ) are used as the evaluated criterions for web service selection. These criterions are called common attribute which can be replaced depending on the decision makers’ requirements. 3.2. Preference setting The difference between WSS-TOPSIS and the traditional TOPSIS is that the latter requires decision makers to set an evaluated value for each attribute for all alternatives. To lessen the burden on the decision makers, the former method was proposed to collect the weight of QoS attributes and the weight of tasks.

115

Criteria

CT

CP

CR

CA

CD

DMk

Wk,T

Wk,P

Wk,R

Wk,A

Wk,D

Table 2 The domain-related attribute weight for decision maker DMk . Task

t1

…

ti

…

tN

DMk (Decision maker)

Wk,D1,1 : Wk,D1,d

… … …

Wk,Di,1 : Wk,Di,d

… … …

Wk,DN,1 : Wk,DN,d

SUM

Wk,D1

…

Wk,Di

…

Wk,DN

1

i

N

Table 3 The task weight for decision maker DMk . Task

t1

…

ti

…

tN

DMk

TWk,1

…

TWk,i

…

TWk,N

There are two types of QoS attribute weight: common attribute weight, which is introduced in advance and domainrelated weight. The QoS of domain-related (CD ) is also used as another evaluated criterions for web service selection. Table 1 provides the attribute weight for decision maker k, 0 < = Wk, T , Wk,P , Wk,R , Wk,A , Wk,D ≤ 1 and Wk,T + Wk,P + Wk,R + Wk,A + Wk,D = 1, k = 1,…, K. • The domain-related weights are that there are 0 to di domain-related attribute weights for each domain task and the sum of the domain-related attribute weights is equal to Wk,D and 0 ≤ TW kD1, . . . , TW kDd, . . . , TW kDdi ≤  W kD, di d=1 TW kDd = W kD, as shown in Table 2. The different of common attribute and domain-related weight is that the former can be collected by a thirdparty organization and the later can be collected by searching the same or similar web service of domain. • QoS constraint: decision makers set the constraint against the QoS attributeThe weight of tasks: the web service composition includes many task processes and each decision makers will assign different degrees of importance to each tasks; hence, the decision makers, DMk , must set the task weight, TWk,I , for task ti and 0 ≤ TW k, 1, . . . , TW k, i, . . . , TW k, N ≤  1, N i=1 TW k, i = 1, as shown in Table 3. 3.3. Alternatives computing Alternatives with the best overall performance are generated for each decision maker based on individual preference. Next, IP and reference to prior research [12,13] are used to design an algorithm – Attribute and Task weight Integer Programming (ATIP) – to compute the web service composition. To strengthen our accurate results, service composition processes, which can be regarded as business processes, are suggested here. Decision makers should assign different degrees of importance to different tasks. ATIP considers the importance of different tasks to select the web service that best meets the decision maker’s preferences and allows tasks to

116

H.-C. Wang et al. / Computer Networks 93 (2015) 111–124 Table 4 QoS aggregating function.

CT

Sequence

Switch

s

h

f =1

s

CP

ws f, j,T

CR CA

f =1

p f ws f, j,T

f =1

p f ws f, j,P

h

f =1 ws f, j,P 1 s f =1 ws f, j,R s s f =1 ln ws f, j,A

1 h

Flow

h

f =1

max{ws f, j,T f ∈{1,...,pl } }  pl ws f, j,P f =1 1  pl ws f, j,R pl f =1  pl ln ws f, j,A f =1

p f ws f, j,R

min{ws f, j,A f ∈{1,...,h} }

Symbol s h pi pl wsf,j,T wsf,j,P wsf,j,R wsf,j,A

The number of processes with a sequence relationship. The number of processes with a selected relationship. The probability that ti will be selected using process h. The number of processes with a flow relationship. The time attribute value of wsf,j in the sequence/ switch/ flow process. The price attribute value of wsf,j in the sequence/ switch/ flow process. The reputation attribute value of wsf,j in the sequence/ switch/ flow process. The availability attribute value of wsf,j in the sequence/ switch/ flow process and the application of a logarithm to transform the value into linear form.

have different domain-related attributes. There are three elements of using IP: defined variables, the objective function and constraint setting. 1. Defined variables: The number of variables is the  number of candidate web services; there are N i=1 mi variables, mi are the number of candidate web services, yi,j are the jth web service variables for task ti (yi,j are the values of wsi,j ). The variables are constrained to 0 and 1. When yi,j is equal to 1, it means that wsi,j has been selected. 2. Objective function: The goal is to recommend a web service composition, also called the service execution plan, for the decision maker. The objective function is composed by the Common Attribute Weight (CAW), the Domain Attribute Weight (DAW) and the Task weight Value (TV), as illustrated in (1). The CAW computing method is taken from [23].

Objplan = max(CAW + DAW + TV)

(1)

I. CAW: The common attributes are aggregated to obtain the service composition attribute values and the process normalization and weight are used to obtain the CAW. During the aggregation process, the formula is referenced and modified [21] to calculate the aggregated QoS values, as shown in Table 4. For different attribute units, the attributes must be normalized to form the same unit and direction. The attributes can be classed into benefit attributes and cost attributes. Benefit attributes are the larger, superior attributes, as illustrated in (2); cost attributes are the smaller, inferior attributes, as illustrated in (3). Original_qv is the value before normalization and max_qv is the value that indicates the best value under the same attribute; min_qv is the value that exhibits the worst value under the same attribute. After normalization and weighting, the CAW is obtained, as illustrated in (4).

QV =

QV =

 original_qv−min_qv max_qv−min_qv

,

1,

 max_qv−original_qv max_qv−min_qv

1,

,

if max_qv = min_qv if max_qv = min_qv

(2)

if max_qv = min_qv if max_qv = min_qv

(3)

CAW =

CAWCT MAX − CAWCT CAWCT MAX − CAWCT min × Wk,T +

CAWCPMAX − CAWCP CAWCPMAX − CAWCP min

× Wk,P +

CAWCR − CAWCR min × Wk,R CAWCRMAX − CAWCR min

+

CAWCA − CAWCA min × Wk,A CAWCAMAX − CAWCA min

(4)

QV: the value of attribute after normalization, Original_qv: the value of attribute before normalization max_qv: the maximum value of attribute under the same QoS min_qv: the minimum value of attribute under the same QoS II. DAW: To consider the different domain-related attributes of different tasks, the DAW is computed as in (5).

 DAW =

 i∈S

di Nor(wsi, j,Cd )  d=1 × Wk,CD i∈S di

j∈W Si

(5)

III. TV: The QoS value is normalized first. Then, the normalized values are multiplied by the corresponding attribute weights and added and finally, multiplied by the weight of the task to obtain the TV, as shown in (6). TV =

⎧ ⎨

  i∈S j∈W Si

⎡

⎤

⎫

Nor(wsi, j,CT ) × Wk,T + Nor(wsi, j,CP ) × Wk,P ⎬ × yi, j ×⎣+Nor(wsi, j,CR ) × Wk,R + Nor(wsi, j,CA ) × Wk,A⎦ × TWk,i i ⎩ ⎭ + dd=1 Nor(wsi, j,Cd ) × Wk,Cd

(6) 3. Constraint setting: The sum of the variables within the same task is equal to 1, which means that only one web service will be selected within the task, as illustrated in Eq. (7).

 j∈W Si

yi j = 1,

∀i ∈ S

(7)

H.-C. Wang et al. / Computer Networks 93 (2015) 111–124

117

Table 5 The characteristics comparison of WSS-TOPSIS and TOPSIS. Item

WSS-TOPSIS

TOPSIS

Generation of alternatives Weight setting

In accordance with decision makers’ preferences

The distance of the chosen alternative between the positive ideal solution (PIS) and negative ideal solution (NIS). Only collecting common attribute of decision makers

Collecting the task weight to consider the importance the decision makers attribute to different tasks. Combination of QoS, attribute weights and task weight.

Evaluated value

Table 6 The evaluated value of Aa according to decision maker k.

QoS

Table 7 The decision matrix of decision maker k.

Task

t1

…

ti

…

tN

Aggregation

CT CP CR CA

xk,a,1,T xk,a,1,P xk,a,1,R xk,a,1,A

…

xk,a,i,T xk,a,i,P xk,a,i,R xk,a,i,A

…

xk,a,N,T xk,a,N,P xk,a,N,R xk,a,N,A

zk,a,T zk,a,P zk,a,R zk,a,A

…

…

The evaluation of each attribute is given depends on performance of the web service by decision makers

The detail calculation of CAW, DAW, and TV are presented in appendix B. After finding the composition using the max Objplan , the alternative is also found against the decision maker.

Alternative

Alternative

1. Construct the decision matrix Alternative: {A1 ,…, Aa ,…, Ak } alternatives are generated from K decision makers. During this process, common attributes are only used for selection. The normalized attribute weights, task weights and QoS values are combined to obtain the evaluated value, xk,a,i,c , as shown in Table 6. c can be {T, P, R, A}, as illustrated in (8). However, the common attribute weights are only considered here, so the domain-related weight must be reallocated to common attributes. Then, the QoS values of each task are aggregated to produce zk,a,c using (9) to formulate the decision matrix of decision

CT

CP

CR

CA

A1 : Aa : AK

zk,1,T : zk,a,T : zk,a,T

zk,1,P : zk,a,P : zk,a,P

zk,1,R : zk,a,R : zk,a,R

zk,1,A : zk,a,A zk,a,A

Table 8 The aggregated decision matrix.

3.4. Alternatives ranking This step involves the alternatives ranking process, used to determine the alternative that most strongly matches the group’s preferences. During the traditional TOPSIS process, the decision makers must construct the decision matrix. For example, in [28], when there are five alternatives and four criterions, the decision makers must set twenty evaluated values. When the web service candidates become unavailable or new web services emerge, the alternatives will change, which means that the decision makers must reset the large number of evaluated values. The selection of web services differs from other decision problems regarding the QoS, which can be regarded as an evaluated reference. Consequently, the WSS-TOPSIS method was proposed to combine the QoS, attribute weights and task weight to produce the evaluated values and eliminate the workload required for significant preference setting. Once, when the web service candidates change, new alternatives are regenerated and a new final solution is determined using prior preference settings. The characteristics comparison of WSS-TOPSIS and TOPSIS are shown in Table 5 and the ranking process proceeds as follows:

Criteria

Criteria

CT

CP

CR

CA

A1 : Aa : AK

g1,T : ga,T : gK,T

g1,P : ga,P : gK,P

g1,R : ga,R : gK,R

g1,A : ga,A gK,A

maker k, as shown in Table 7. Finally, the aggregated decision matrix is obtained, as shown in Table 8, using the average corresponding evaluated values, as illustrated in (10). DMk  sQoSCc weight

  xk,a,i,c =



Wk,D Wk,c + 4

 ×

TWk,i



×QV

(8)

DMk  sti weigh

zk,a,c =

N 

xk,a,c,i

(9)

i=1

ga,c =

K 1 zk, a,c ; a = 1, . . . , K; c = T, P, R, A. K

(10)

k=1

2. Define the positive ideal solution, as illustrated in (11) and negative ideal solution, as illustrated in (12).

A∗ = (max (g1i , . . . , gai , . . . , gKi ), i = 1, 2, 3, 4) a

= (g∗1 , g∗2 , g∗3 , g∗4 )

(11)

A− = (min (g1i , . . . , gai , . . . , gKi ), i = 1, 2, 3, 4) a

− − − = (g− 1 , g2 , g3 , g4 )

(12)

3. The advantage of Euclidean distance is the computation of the differences between two elements, two sets and so on in the two-dimensional or three-dimensional space. Thus, this method is used to calculate the distance from the alternative to the positive ideal solution, as illustrated

118

H.-C. Wang et al. / Computer Networks 93 (2015) 111–124 Table 11 The web service compositions of each decision maker.

Table 9 The attribute weights. Criteria

CT

CP

CR

CA

CD

Task

t1

t2

t3

t4

t5

t6

DM1 DM2 DM3

0.05 0.6 0.3

0.6 0.1 0.1

0.2 0.2 0.3

0.05 0.05 0.2

0.1 0.05 0.1

A1 A2 A3

ws1,2 ws1,1 ws1,3

ws2,3 ws2,3 ws2,1

ws3,3 ws3,3 ws3,1

ws4,1 ws4,1 ws4,1

ws5,2 ws5,2 ws5,3

ws6,2 ws6,2 ws6,3

Table 12 The normalized QoS values of each alternative.

Table 10 The task weights. Task

t1

t2

t3

t4

t5

t6

Alternative

Task

t1

t2

t3

t4

t5

t6

DM1 DM2 DM3

0.5 0.3 0.2

0.1 0.2 0.2

0.2 0.2 0.2

0.06 0.1 0.1

0.06 0.1 0.1

0.08 0.1 0.2

A1

CT CP CR CA CT CP CR CA CT CP CR CA

0.4 0.25 1 0.625 1 0 0.71 1 0 1 0 0

0 0 1 1 0 0 1 1 1 1 0 0.56

0 0 1 1 0 0 1 1 1 1 0 0

1 1 1 1 1 1 1 1 1 1 1 1

1 0 1 0 1 0 1 0 0 1 0 1

1 0 1 0 1 0 1 0 0 1 0 1

A2 ws1,1 ws1,2 ws1,3

ws5,1 ws5,2 ws5,3

ws4,1 ws4,2

t5

t1 t2 ws2,1 ws2,2 ws2,3 ws2,4

t4

t3 ws3,1 ws3,2 ws3,3

A3

t6 ws6,1 ws6,2 ws6,3

Fig. 4. The web service scenario composition.

in (13) and the distance from the alternative to the negative ideal solution, as illustrated in (14).



Sa∗

=

4 

(gai − g∗i )2 ,a = 1, . . . , K.

(13)

i=1

 Sa− =

4 

(gai − g−i )2 ,a = 1, . . . , K.

(14)

i=1

3. Compute the similarity between the alternative and the positive ideal solution, as illustrated in (15).

Ca∗ = Sa− /(Sa∗ + Sa− ), 4.

a = 1, . . . , K.

(15)

Sort the alternatives using Ci∗ in descending order; the solution with the highest Ci∗ will be selected as the optimal solution.

3.5. Scenarios The scenarios are replicated from [12] as an example to introduce our proposed method. Assume that there are six task processes and three decision makers; DM1 has a greater emphasis on cost savings, DM2 has a greater emphasis on operational efficiency and with DM3 , the decision makers are more comprehensive. Every decision maker understands the task processes and has indicated his or her QoS attribute weights and task weights, as shown in Tables 9 and 10. There are at least 2 web service candidates that can process the task functional requirement, as shown in Fig. 4. The QoS values are provided by the web service providers or captured from the UDDI. During the generation of alternatives, the web service composition, which renders the Objplan a maximum, must be determined. The service composition is applied, for which the 1st candidate is selected and the weights of DM1 are

regarded as an example to compute the CAW. The values of CAW, DAW, and TV are 0.33, 0.02 and 0.495 respectively. Table 11 provides the web service compositions generated by the three decision makers. Table 12 provides the normalized QoS values of the service compositions. The evaluated values for which every decision maker has different alternatives are obtained by combining Tables 9, 10, and 12 and Eq. (8). Table 13 is the evaluated result of DM1 using (9). Table 14, a simplified result obtained using Table 13, is a decision matrix of DM1 . An aggregated decision matrix is obtained using (10) and the PIS/NIS is defined using (11, 12), as shown in Table 15. By computing C∗ and sorting the alternatives in descending order using C∗ , the C∗ with the highest value will be chosen as the final solution, as shown in Table 16. 4. The experiment In our experiment environment, the PC has the configuration: Intel (R) Core (TM) 2 Quad CPU [email protected] with DDR2RAM memory 2GB, Windows XP and java7 is used to build the system to obtain the final selection results. The optimization problem is solved by running CPLEX12.1, which was developed by IBM. The architecture is shown in Fig. 5. 4.1. Data simulation Despite the availability of many ready-made web services, the collection of a large number of complete QoS values is still quite time consuming and labor intensive. Many studies [13,14,23,25] on QoS-based web service selection have made good use of simulated data for experiment and verification. The simulation of our experiment is similarly processed and referred to in [13,25,56]. Comparing to [9] on multi-user condition, the QoS values in history records are used. But the difference is that our experiment consider decision maker’s preferences, and aims at guaranteeing consistency of the selection such preferences. The composite service has multiple

H.-C. Wang et al. / Computer Networks 93 (2015) 111–124

119

Table 13 The evaluation results for DM1 . Alt.

Task

t1

t2

t3

t4

t5

t6

Sum

A1

CT CP CR CA CT CP CR CA CT CP CR CA

0.015 0.07813 0.1125 0.02344 0.0375 0 0.08 0.0375 0 0.3125 0 0

0 0 0.0225 0.0075 0 0 0.0225 0.0075 0.0075 0.0625 0 0.00417

0 0 0.045 0.015 0 0 0.045 0.015 0.015 0.125 0 0

0.0045 0.0375 0.0135 0.0045 0.0045 0.0375 0.0135 0.0045 0.0045 0.0375 0.0135 0.0045

0.0045 0 0.0135 0 0.0045 0 0.0135 0 0 0.0375 0 0.0045

0.006 0 0.018 0 0.006 0 0.018 0 0 0.05 0 0.006

0.03 0.11563 0.225 0.05044 0.0525 0.0375 0.19286 0.0645 0.027 0.625 0.0135 0.01917

A2

A3

Data simulation Data Generator

Table 14 The decision matrix of DM1 .

GDWSS Normalization

WS QoS

A1 A2 A3

CT

CP

CR

CA

0.03 0.0525 0.0825

0.11563 0.0375 0.15313

0.225 0.19286 0.41786

0.05044 0.0645 0.11494

ATIP DM preference Task information

WSS-TOPSIS

Cplex Fig. 5. The architecture of experiment.

Table 15 The decision matrix and PIS/NIS.

A1 A2 A3 PIS NIS

CP

CR

CA

0.14775 0.205 0.18375 0.205 0.14775

0.05135 0.02042 0.13021 0.130208 0.020417

0.25417 0.23119 0.1572 0.254167 0.157202

0.07801 0.09067 0.08521 0.090667 0.07801

Table 16 The alternatives ranking result.

A1 A2 A3

tasks and each task can be fulfilled by a group of services. Then the optimization is to select a service for each task so that the overall evaluation function can be optimized. In the common attribute value simulation process, the complexity of web services affects execution time and price [25], so execution time and price are multiplied by a complex coefficient, γ . Time is composed of execution time and waiting time. Execution time is randomly generated in a uniform distribution in the range of [0.5, 8] and multiplied by γ . Waiting time refers to the time users spend waiting for access to services. High-performance web service necessitates long wait times because there are many users. Wait time is randomly generated in a uniform distribution with a range of [1.5, 10] and then multiplied by the normalized value, which is multiplied by the reputation and availability values. Price is affected by reputation, availability and time, which will be higher with better quality [13,25]. Likewise, price is also assumed to be higher with more complex functions. Thus, the above factors would be multiplied to obtain the price. Reputation is randomly generated in uniform distribution in the range of [0.8, 0.99]. Availability is randomly generated in a uniform distribution with a range of [0.95, 0.999].

CT

S∗

S−

C∗

Rank

0.098264 0.11217 0.099415

0.10178 0.094403 0.115767

0.50879 0.456997 0.537996

2 3 1

4.2. Experiment results and comparison To verify that ATIP still has the characteristics of global optimization, ATIP is compared with the SAW method, which considers only local optimization. The 100 web services are generated for each task and the improvement in time and price occurring under different weights is compared, as shown in Table 17. The web services selected using ATIP are faster and cheaper than those chosen using SAW and the improvement rates are all above 7.41%. Some values of the improvement are negative under different composite services because the performance of SAW is superior to ATIP according to the relations of the criterions in time or price. After considering domain-related weights and task weights, it is obvious that ATIP still has the characteristics of global optimization. The web service with the higher domain-related attribute value will be selected after considering the domain-related attribute of each task and the web services that more strongly correspond to decision makers’ preferences will be selected

120

H.-C. Wang et al. / Computer Networks 93 (2015) 111–124 Table 17 Differences in web service composition. Time

Price

Composite service

SAW

ATIP

Improvement (%)

SAW

ATIP

Improvement (%)

1 2 3 4 5 6 7 8 9 10 Average

10.28 17.03 15.95 13.54 16.92 11.33 13.64 17.08 16.55 16.07 14.84

10.28 16.48 15.26 9.34 17.12 7.81 14.67 15.67 15.07 15.69 13.74

0.00 3.23 4.32 31.01 −1.21 31.04 −7.53 8.26 8.95 2.36 7.41

8.71 21.54 18.69 16.03 20.02 14.05 14.37 20.11 16.25 20.10 16.99

10.55 18.54 17.93 10.41 18.54 8.76 15.03 18.54 16.23 18.45 15.30

−21.18 13.93 4.05 35.06 7.38 37.65 −4.54 7.82 0.12 8.21 9.94

Table 18 Selected results of PIP and ATIP. Method

Alternative

t1

t2

t3

t4

t5

t6

PIP

A1 A2 A3 A4 A5 A6 A7 A8 A9 A10 A1 A2 A3 A4 A5 A6 A7 A8 A9 A10

40 80 80 80 80 40 40 80 40 80 40 80 80 40 80 40 40 80 40 80

29 2 2 2 2 2 2 2 2 2 29 2 11 29 2 29 11 2 2 11

84 62 68 84 62 27 62 62 62 62 84 62 68 84 62 27 62 62 62 62

67 67 67 67 67 67 67 67 67 67 38 67 67 41 67 41 67 67 67 67

12 54 54 54 54 54 54 54 54 54 12 54 54 12 54 12 54 54 54 54

82 86 17 82 86 2 86 86 86 86 82 86 17 82 86 33 17 86 86 86

ATIP

Table 19 The domain-related attribute values of PIP and ATIP. Alternative

A1 A3 A4 A6 A7 A10 average

Domain-related att. PIP

ATIP

Improvement

6.57 4.70 5.32 4.02 3.37 3.72 4.62

6.74 5.15 6.77 5.84 4.79 4.17 5.58

2.58% 9.52% 27.36% 45.50% 42.33% 12.02% 20.84%

in terms of the tasks to which the decision makers pay attention. The 100 web services selected using previous experiments and the decision makers’ weights are applied for the comparison of the two methods, PIP and ATIP. PIP is the method that selects web services without considering domain-related attributes and task weights. Selected results are shown in Table 18. The gray-colored blocks are the tasks that produced different selected results. The domain-related attribute values of PIP and ATIP are shown in Table 19. The domain-related attribute values of

ATIP exhibit better performance than PIP for the different attribute domains. According to Table 20, approximately 62% of the tasks for which the PIP and ATIP selections differ are more important than the other tasks. ATIP selects the web services that can satisfy the decision makers’ preferences to a greater extent than PIP because the normalized value, which has been weighted by ATIP, is higher. During the group decision-making process, consensus is an important indication of group agreement or reliability. To fully reflect the actual behavior of the group, a final decision should be made at a significant level of consensus [30]. Therefore, the consensus of our ranking results, which is produced by the WSS-TOPSIS, is calculated to verify that the group results are consistent with the views of the majority of the decision makers. By using 100 web services used in previous experiments and decision makers’ weights, which were collected from five decision makers, the alternatives ranking is obtained, as shown in Table 21. The Group column contains the overall group ranking. That is, the DM1 column contains the DM1 ranking and so on. In addition, Spearman’s rank correlation coefficient is applied to obtain the consensus, ρ , as illustrated in (16) and the T-value is then obtained using (17) and the p-value. Finally, the hypothesis testing is processed: H0, the decision maker’s ranking, is irrelevant to the ranking of group, as shown in Table 22. According to Table 21, the rankings of DM1, DM2 and the Group are the same, ρ is equal to 1 and consensus is apparent. The ρ between Group and DM3 and DM4 is 0.9 and the p-value is 0.037387, so the H0 is rejected with 95% confidence, which means that the ranking between the Group and DM3 and DM4 has significant consensus. The ρ between the Group and DM5 is −0.6, which means that the Group and DM5 rankings are partially conflicting; the p-value is 0.284768, with 95% confidence, so the H0 cannot be rejected. In summary, the ranking result can match the preferences of the majority of decision makers.

ρGroup,a = 1 −

6



t−valueGroup,a =

2

(rGroup − raj ) j , a = 1, . . . , K. n(n2 − 1)

j

(16)

ρGroup,a  ×

K−2 , 2 1 − ρGroup,a

a = 1, . . . , K.

(17)

H.-C. Wang et al. / Computer Networks 93 (2015) 111–124

121

Table 20 The PIP and ATIP comparison. Decision Maker

Task

Weight

PIP ws no.

PIP QoS

ATIP ws no.

ATIP QoS

DM1(A1) DM3(A3) DM4(A4)

t4 t2 t1 t2 t4 t5 t2 t4 t5 t6 t2 t6 t2 8/13=0.62 62% over 0.16

0.16 0.2 0.17 0.1 0.23 0.21 0.13 0.1 0.09 0.05 0.2 0.2 0.19

67 2 80 2 67 54 8 67 54 2 2 86 2 Avg.= 0.62

0.78 0.67 0.57 0.38 0.72 0.63 0.44 0.74 0.66 0.65 0.62 0.61 0.64

38 11 40 29 41 12 29 41 12 33 11 17 11 Avg.= 0.78

0.81 0.77 0.62 0.86 0.82 0.87 0.83 0.80 0.83 0.69 0.71 0.75 0.73

DM6(A6)

DM7(A7) DM10(A10) Overall result

Table 21 The alternatives ranking result. Alt.

Group

DM1

DM2

DM3

DM4

DM5

A1 A2 A3 A4 A5

4 1 2 3 5

4 1 2 3 5

4 1 2 3 5

4 2 1 3 5

3 1 2 4 5

3 4 5 1 2

5. Conclusion and future work In this paper, a method of web service selection considering group preference – GDWSS – is proposed to present an approach for selecting Web services during the process of building an orchestration/composition, on the basis of some Quality of service (QoS) features. The specific proposal is to consider decision maker’s preferences, and aims at guaranteeing consistency of the selection such preferences. The composite service has multiple tasks and each task can be fulfilled by a group of services. Then the optimization is to select a service for each task so that the overall evaluation function can be optimized. GDWSS allows different tasks to have individual domain-related attributes. These attributes can still be aggregated to obtain the domain-related attribute value of the web service composition even if each task has different types of domain-related attributes. Alternative sorting allows for a significant degree of consensus, meaning that the group selection result can match the preferences of the majority of decision makers. With the overall GDWSS process, the final solution has not been selected from the

default alternatives, but alternatives are generated by decision makers’ preferences and the final solution is chosen from the alternatives. Furthermore, due to each decision maker must understand the process of web service composition and determine the importance of each task. This method significantly reduces the burden of decision makers with regard to preference setting. Once web-service candidates become unavailable or new web services emerge, a new solution can be regenerated without resetting using existing preference weights. According to the experiment, the web services selected using ATIP are faster and cheaper than those selected using SAW and the improvement rates are all above 7.41%. ATIP allows for the selection of web services that can satisfy decision makers’ preferences to a greater extent than PIP because the normalized value, which has been weighted using ATIP, is higher. Through a series of tests, the p-value is 0.284768, with 95% confidence, so the H0 cannot be rejected. Therefore, the ranking result will be more convincing to the decision makers. Currently, our proposed method involves processing that decision makers can consider to select local or global optimization services in accordance with individual preferences and thus meet their requirements and strengthen their method of automatically obtaining the normalized values of QoS at runtime. However, the users have to give too many weights and reduce complexity remain topic for further research. In addition, another goal of the on-going researches will toward paradigm of service choreographies in selecting web services to seek collaboration.

Table 22 Spearman correlation coefficients. Decision maker

DM1 DM2 DM3 DM4 DM5

Group P correlation coefficient

p-value Significance (2-tailed)

1 1 0.9 0.9 −0.6

0.037387∗ 0.037387∗ 0.284768

122

H.-C. Wang et al. / Computer Networks 93 (2015) 111–124

Appendix A Notation summary. Symbol

Description

Symbol

Description

K

Total number of decision makers

h

The number of processes with a selected relationship.

DMk

The k-th decision maker, k=1 to K

Pi

The probability that task ti will be selected using process h.

S

The abstract services set

pl

The number of processes with a flow relationship.

t

The process of task

wsf,j,T

The time attribute value of wsf,j in the sequence/ switch/ flow process.

N

S is composited by N tasks

wsf,j,p

The price attribute value of wsf,j in the sequence/ switch/ flow process.

mi

Web services can be selected for task ti , i = 1 to N

wsf,j,R

The reputation attribute value of wsf,j in the sequence/ switch/ flow process.

wsi,j

The jth physical web service for task ti

wsf,j,A

The availability attribute value of wsf,j in the sequence/ switch/ flow process and the application of a logarithm to transform the value into linear form.

CT

The QoS of time

QV

The value of attribute after normalization

CP

The QoS of price

original_qv

The value of attribute before normalization

CR

The QoS of reputation

max_qv

The maximum value of attribute under the same QoS

CA

The QoS of availability

min_qv

The minimum value of attribute under the same QoS

CD

The QoS of Domain-related

Nor

Normalization

di

The number of ith task of domain-related attribute for each domain task, I = 1 to N

xk,a,i,c

The normalized attribute weights, task weights and QoS values are combined to obtain the evaluated value for decision maker k, a = 1 to K, I = 1 to N, c=T or P or R or A

Di,j

The jth domain attribute of the ith task, i = 1 to N, j = 1 to di

zk,a,c

The QoS values of each task are aggregated to produce for decision maker k, a=1 to K, c=T or P or R or A

Wk,T

The weight of CT for decision maker k

ga,c

The average corresponding evaluated values, a = 1 to K, c = T or P or R or A

Wk,P

The weight of CP for decision maker k

A∗

The set of positive ideal solution

Wk,R

The weight of CR for decision maker k

A−

The set of negative ideal solution

Wk,A

The weight of CA for decision maker k

S∗ a

The distance of solution a to positive ideal solution

Symbol

Description

Symbol

Description

Wk,D

The weight of CD for decision maker k

S− a

The distance of solution a to negative ideal solution

WkDi,di

The weight of Di,di for decision maker k, I = 1 to N

C∗ a

The degree of similarity between solution a and positive ideal solution

TWk,i

The task weight of decision maker k, for task ti , I = 1 to N

γ

Task complexity

CAW

The Common Attribute Weight

τ

The simulated value of time attribute

DAW

The Domain Attribute Weight

ϕ

The simulated value of price attribute

TV s

Task weight Value The number of processes with a sequence relationship

ε α

The simulated value of reputation attribute The simulated value of availability attribute

H.-C. Wang et al. / Computer Networks 93 (2015) 111–124

123

Appendix B The process of computing CAW. Criteria

Normalization

CT

CAWCT MAX −CAWCT CAWCT MAX −CAWCT min

=

40−28 40−26

CP

CAWCPMAX −CAWCP CAWCPMAX −CAWCP min

=

1000−910 1000−690

CR

CAWCR −CAWCR min CAWCRMAX −CAWCR min

=

0.955−0.93 0.985−0.93

CA

CAWCA −CAWCA min CAWCAMAX −CAWCA min

=

−0.183+0.32336 −0.0604+0.32336

= 0.857 = 0.29

DM1 weight

CAW

W1,T = 0.05

0.857 × 0.05 + 0.29×0.6 + 0.455×0.2 + 0.5324×0.05 = 0.33

W1,P = 0.6

= 0.455

W1,R = 0.2

= 0.5324

W1,A = 0.05

The process of computing DAW.

CD1,1 CD2,1 CD3,1 CD5,1 CD6,1

15 0.2 0.65 0.2 0.2

Normalization

DM1 weight

DAW

(30-15)/(30-15) = 1 (0.2-0.2)/(1-0.2) = 0 (0.65-0.65)/(0.83-0.65) = 0 (0.2-0.2)/(1-0.2) = 0 (0.2-0.2)/(1-0.2) = 0

W1,D = 0.1

(1 + 0 + 0 + 0 + 0)/5×0.1 = 0.02

The process of computing TV.

WS

CT

CP

CR

CA

CD

ws1,1 ws2,1 ws3,1 ws4,1 ws5,1 ws6,1 TV

W1,T 0.05 1 1 1 1 0.33 0.33 0.495

W1,P 0.6 0 1 1 1 0.6 0.6

W1,R 0.2 0.71 0 0 1 0.43 0.43

W1,A 0.05 1 0.56 0 1 0 0

W1,D 0.1 1 0 0

References [1] G. Delac, et al., A methodology for SIP and SOAP integration using application-specific protocol conversion, Acm Trans. Web 6 (4) (2012) 28. [2] S.-H. Li, et al., Semantic-based transaction model for web service, Inf. Syst. Front. 15 (2) (2013) 249–268. [3] P. Wang, K.-M. Chao, C.-C. Lo, On optimal decision for QoS-aware composite service selection, Exp. Syst. Appl. 37 (1) (2010) 440–449. [4] J.C.P. Cheng, et al., A service oriented framework for construction supply chain integration, Autom. Constr. 19 (2) (2010) 245–260. [5] R. Gronmo, M.C. Jaeger, Model-driven methodology for building QoSoptimised web service compositions, in: L. Kutvonen, N. Alonistioti (Eds.), in Distributed Applications and Interoperable Systems, 2005, pp. 68–82. Editors. [6] W. Zhang, Y. Xu, Implementation of Agile supply chain information integration system in manufacturing industry based on service-oriented architecture and web service, in: H. Zhang, G. Shen, D. Jin (Eds.), in Advanced Research on Information Science, Automation and Material System, Pts 1-6, 2011, pp. 1145–1148. [7] S.-C. Liu, S.-S. Weng, Applying genetic algorithm to select web services based on workflow quality of service, J. Electron. Commer. Res. 13 (2) (2012) 157–172. [8] T. Yu, Y. Zhang, K.-J. Lin, Efficient Algorithms for Web Services Selection with End-to-End QoS Constraints, Acm Transactions on the Web 1 (1) (2007). [9] S. Wang, et al., Multi-user web service selection based on multi-QoS prediction, Inf. Syst. Front. 16 (1) (2014) 143–152. [10] S.X. Sun, J. Zhao, A decomposition-based approach for service composition with global QoS guarantees, Inf. Sci. 199 (2012) 138–153. [11] T. Neubauer, C. Stummer, Interactive selection of Web services under multiple objectives, Inf. Technol. Manag. 11 (1) (2010) 25–41. [12] L.Z. Zeng, et al., QoS-aware middleware for Web Services Composition, IEEE Trans. Softw. Eng. 30 (5) (2004) 311–327. [13] D. Ardagna, B. Pernici, I.C. Soc, Global and local QoS constraints guarantee in web service selection, in: IEEE International Conference on Web Services, vols 1 and 2, Proceedings. 2005, 2005, pp. 805–806.

0 0.2

Summing up the weighted normalized attribute values

DM1 task weight

After multiplied by the task weight

0.342 0.678 0.65 1 0.462 0.482

0.5 0.1 0.2 0.06 0.06 0.08

0.17 0.068 0.13 0.06 0.028 0.039

[14] A.F.M. Huang, C.-W. Lan, S.J.H. Yang, An optimal QoS-based Web service selection scheme, Inf. Sci. 179 (19) (2009) 3309–3322. [15] G. Chang, Using Bayesian Networks to Measure Web Service QoS, in: Proceedings of the 2012 International Conference on Communication, Electronics and Automation Engineering, G. Yang, 2013, pp. 1233–1238. [16] W.-L. Lin, et al., Multi-group QoS consensus for web services, J. Comput. Syst. Sci. 77 (2) (2011) 223–243. [17] H.-S. Shih, H.-J. Shyur, E.S. Lee, An extension of TOPSIS for group decision making, Math. Comput. Model. 45 (7-8) (2007) 801–813. [18] Z. Yue, A method for group decision-making based on determining weights of decision makers using TOPSIS, Appl. Math. Model. 35 (4) (2011) 1926–1936. [19] Z. Yue, Developing a straightforward approach for group decision making based on determining weights of decision makers, Appl. Math. Model. 36 (9) (2012) 4106–4117. [20] K. Jabeur, J.M. Martel, S. Ben Khelifa, A distance-based collective preorder integrating the relative importance of the group’s members, Group Decis. Negot. 13 (4) (2004) 327–349. [21] G. Canfora, et al., An approach for QoS-aware service composition based on genetic algorithms, in: GECCO 2005: Genetic and Evolutionary Computation Conference, Vols 1 and 2, 2005, pp. 1069–1075. [22] E. Karakoc, P. Senkul, Composing semantic Web services under constraints, Exp. Syst. Appl. 36 (8) (2009) 11021–11029. [23] P. Xiong, Y. Fan, M. Zhou, QoS-aware Web service configuration, IEEE Trans. Syst. Man Cybern. Part a-Syst. Hum. 38 (4) (2008) 888–895. [24] Y. Hao, Y. Zhang, J. Cao, A novel QoS model and computation framework in web service selection, World Wide Web-Internet Web Inf. Syst. 15 (5-6) (2012) 663–684. [25] D. Ardagna, B. Pernici, Adaptive service composition in flexible processes, IEEE Trans. Softw. Eng. 33 (6) (2007) 369–384. [26] X.-Q. Fan, X.-W. Fang, C.-J. Jiang, Research on Web service selection based on cooperative evolution, Exp. Syst. Appl. 38 (8) (2011) 9736– 9743. [27] X. Zhao, et al., A hybrid clonal selection algorithm for quality of serviceaware web service selection problem, Int. J. Innov. Comput. Inf. Control 8 (12) (2012) 8527–8544.

124

H.-C. Wang et al. / Computer Networks 93 (2015) 111–124

[28] C.-C. Lo, et al., Service selection based on fuzzy TOPSIS method, in: 2010 IEEE 24th International Conference on Advanced Information Networking and Applications Workshops, 2010, pp. 367–372. [29] L. Zhang, H. Zou, F. Yang, Multi-attribute group decision making-based decentralized web service selection, in: 2010 International Conference on Service Sciences, 2010, pp. 82–87. [30] C. Kahraman, et al., Information systems outsourcing decisions using a group decision-making approach, Eng. Appl. Artif. Intell. 22 (6) (2009) 832–841. [31] M. Behzadian, et al., A state-of the-art survey of TOPSIS applications, Exp. Syst. Appl. 39 (17) (2012) 13051–13069. [32] M. Anisseh, et al., Fuzzy extension of TOPSIS model for group decision making under multiple criteria, Artif. Intell. Rev. 38 (4) (2012) 325–338. [33] Z. Yue, Extension of TOPSIS to determine weight of decision maker for group decision making problems with uncertain information, Exp. Syst. Appl. 39 (7) (2012) 6343–6350. [34] R. Khorshidi, et al., Selection of an optimal refinement condition to achieve maximum tensile properties of Al-15%Mg2 Si composite based on TOPSIS method, Mater. Des. 46 (2013) 442–450. [35] Y.-L. Chen, L.-C. Cheng, An approach to group ranking decisions in a dynamic environment, Decis. Support Syst. 48 (4) (2010) 622–634. [36] S. Meng, F. Arbab, QoS-driven service selection and composition using quantitative constraint automata, Fundam. Inf. 95 (1) (2009) 103–128. [37] X. Zhao, Z. Wen, X. Li, QoS-aware web service selection with negative selection algorithm, Knowl. Inf. Syst. 40 (2) (2014) 349–373. [38] E.K. Zavadskas, et al., Sensitivity analysis of a simple additive weight method, Int. J. Manag. Decis. Mak. 8 (5-6) (2007) 555–574. [39] M. Modarres, S. Sadi-Nezhad, Fuzzy simple additive weighting method by preference ratio, Intell. Autom. Soft Comput. 11 (4) (2005) 235–244. [40] S.J. Chen, C.L. Hwang, F.P. Hwang, Fuzzy multiple attribute decisionmaking - methods and applications, Lect Notes Econ. Math. Syst. 375 (1992) 1–531. [41] Z. Longchang, Z. Hua, Y. Fangchun, A dynamic web service composition algorithm based on TOPSIS, J. Netw. 6 (9) (2011) 1296–1304. [42] K.P. Yoon, D.C.-L. Hwang, Multiple Attribute Decision Making: An Introduction (Sage University Paper series on Quantitative Applications in the Social Sciences, 07-104), CA: Sage, Thousand Oaks, 1995. [43] P. Li, et al., Sensitivity analysis of TOPSIS method in water quality assessment: I. Sensitivity to the parameter weights., Environ. Monit. Assess. 185 (3) (2013) 2453–2461. [44] J. Yang, et al., A web services-based approach to develop a networked information integration service platform for gear enterprise, J. Intell. Manuf. 23 (5) (2012) 1721–1732. [45] O. Hilari, Marc, Quality of service (QoS) in SOA systems. A systematic review., Univ. Polit. Catalunya (2009) 1–102. [46] S. Ran, A model for web services discovery with QoS, SIGecom Exch 4 (1) (2003) 1–10. [47] J. O’Sullivan, D. Edmond, A. ter Hofstede, What’s in a service? Towards accurate description of non-functional service properties, Distrib. Parallel Databases 12 (2-3) (2002) 117–133. [48] I. Sora, D. Todinca, C. Avram, Translating user preferences into fuzzy rules for the automatic selection of services, in: Proceedings 2009 5th International Symposium on Applied Computational Intelligence and Informatics, 2009, pp. 497–502. [49] S.L. Liu, et al., A dynamic web service selection strategy with QoS global optimization based on multi-objective genetic algorithm, in: Grid and Cooperative Computing - Gcc 2005, Proceedings, H. Zhuge and G.C. Fox, Editors, 2005, pp. 84–89.

[50] Y. Ma, C. Zhang, Quick convergence of genetic algorithm for QoS-driven web service selection, Comput. Netw. 52 (5) (2008) 1093–1104. [51] D. Ardagna, B. Pernici, Global and local QoS guarantee in Web service selection, in: C. Bussler, A. Haller (Eds.), in Business Process Management Workshops, 2006, pp. 32–46. [52] M. Alrifai, T. Risse, W. Nejdl, A hybrid approach for efficient web service composition with end-to-end QoS constraints, Acm Trans. Web 6 (2) (2012) 31. [53] S.-Y. Hwang, C.-H. Lee, Reliable web service selection in choreographed environments, Decis. Support Syst. 54 (3) (2013) 1463–1476. [54] L.-c. Zhang, C.-j. Li, Z.-l. Yu, Dynamic web service selection group decision-making based on heterogeneous QoS models, J. China Univ. Posts Telecommun. 19 (3) (2012) 80–90. [55] E. Khanfir, et al., A web service selection framework based on user’s context and QoS, in: 2014 IEEE International Conference on Web Services, 2014, pp. 708–711. [56] N.H. Priya, A.M. Shobana Priya, S. Chandramathi, QoS based selection and composition of web services-a fuzzy approach, J. Comput. Sci. 10 (5) (2014) 861–868. Hei-Chia Wang was born in Tainan, Taiwan. He graduated from Chung-Yuan Christian University, Taiwan. He received the M.E. and Ph.D. degrees from University of Manchester, G.B. Now, he is a professor of Department of Industrial and Information Management, National Cheng Kung University, Taiwan. His research interests are knowledge discovery, elearning, bioinformatics, computer network, and distributed software engineering.

Wei-Pin Chiu was born in Kaohsiung, Taiwan. He graduated from Central Police University, Taiwan. He received the M.E. degrees from National Chung-Ceng University, Taiwan. Now, he is a Ph.D. candidate of Department of Industrial and Information Management, National Cheng Kung University, Taiwan. His research interests are time series forecasting, data-mining, and semantic web.

Suei-Chih Wu was born in Taichung, Taiwan. She graduated from National University of Kaohsiung, Taiwan. She received the M.E. degrees from National Chung-Ceng University, Taiwan. Her research interests are web services, and semantic web.