Annotated Bibliography

Computing and Informatics, Vol. 33, 2014, 1095-1115

A M E T H O D O F E V A L U A T IN G T R U S T A N D R E P U T A T IO N F O R O N L IN E T R A N S A C T IO N

Sunhong K im, Wookju H a , Jiwan Se o , Sangyong H an*

School of Computer Science and Engineering Chung-Ang University Heuk seok-dong, Dongjak-gu Seoul 156-756, Korea e-mail: shkimOec. cse. cau. ac. kr, hansyOcau. ac. kr

Mucheol K im

.4 Group o f Industry-University Cooperation Surighyul University 53, Sungkyuldaehak-ro Manan-gu, Anyang, Korea e-mail: mucheol.kimQgmail. com

Abstract. The widespread use of the Internet and evaluater-based technologies has transformed the way business is conducted. Traditional offline businesses have increasingly become online, and there are new kinds of businesses th at solely exist online. Unlike offline business environments, interpersonal trust is generally lacking- in online business settings. Trading partners might feel insecure about the exchange of products and services over the net as they have limited information about each other’s reliability or about the product quality. Considering that enough trust needs to be created to get the online buyer and seller to take actions, trust is a precious asset in online transactions. In order to address the issue of evaluating trust and reputation in online transaction environments, this paper makes use of a social network that graphically represents interpersonal relationships. This paper proposes computational models that systematically evaluate the quantitative level of trust and reputation based on the social network. A method that combines the evaluated trust and reputation levels is also proposed to increase the reliability of online transactions.

corresponding author

1096 S. Kim. W. Ha. J. Seo. S. Han. M. Kim

Keywords: Social network, trust, reputation, online transaction, e-commerce, fuzzy logic

Mathematics Subject Classification 2010: 68M10

1 IN T R O D U C T IO N

W ith the widespread use of the Internet, online business has grown dramatically over the past few years. Traditional offline businesses have increasingly become available online, and numerous new business opportunities are introduced in online settings. For instance, online transactions th a t allow the exchange of products and services entirely electronically can offer the advantages of reduced costs and increased con­ venience [1]. While technical issues related to online transactions such as security and network availability have improved and reached a more or less “steady stage” , sociological aspects such as trust and reputation still require extensive research. In online transaction environments, trust relationships are difficult to establish be­ cause there is no physical contact or interaction between people who are involved in such transactions (sellers, buyers and administrators) [4]. Considering th at on­ line users often mention the lack of trust as one reason for not transacting online, trust is an important component of successful online transactions. In existing online transaction systems, information such as user’s personal information, transactions’ histories and previous customers’ comments or ratings are given to help online users make decisions related to trust. Most of these systems, however, provide only an intuitive approach to trust and reputation without much understanding of these concepts. The provided information is often incomplete, ambiguous and unreliable. Hence, more systematic approaches to formalize and transform the sociological con­ cept “trustworthiness" into quantitative, computational information are required for reliable online transactions [5].

A social network is a graphical representation of a social structure made of individuals or organizations th a t are tied (connected) by one or more specific types of interdependencies [10]. The social network views social relationships in terms of network theory via nodes and ties. Nodes are the individual actors within the networks, and ties are the relationships between the actors. Since ties (or patterns of ties) can be interpreted in many different ways, social networks are useful to map various, complex relationships between the members of social systems. Moreover, transforming qualitative, social concepts into quantitative information is relatively simple in social networks th a t are mathematical structures (i.e., network theory). Thus, social networks have become a popular topic of study, and several applications and technologies th at use social networks have emerged in the last few years.

This paper proposes the computational models that systematically evaluate quantitative trust and reputation levels based on the social network, and it pro­ poses a method th at combines the evaluated trust and reputation values. The

A Method of Evaluating Trust and Reputation for Online Transaction 1097

rest of this paper is organized as follows. Section 2 presents related works and background information. Section 3 describes computational methods for evaluating trust and reputation. Section 4 introduces the computational model that evaluates the trustor’s trust toward the trustee in the social network. Section 5 describes the computational model that evaluates the collective reputation (public opinion) about the trustee in the social network. Section 6 explains the method th at combines the evaluated trust and reputation levels to support reliable online transactions. In Sec­ tion 7, the experiments conducted to evaluate the accuracy of the proposed models and methods in evaluating trust and reputation are presented. Finally, conclusions and future research directions are given in Section 8.

2 RELATED WORK AND BACKGROUND

2.1 FOAF (Friend Of A Friend)

A social network represents interpersonal relationships in terms of the network the­ ory. Every member (node) has his/her own human connections in the network, either direct or indirect. In the social network depicted in Figure 1, the node titled “Person” has two directly connected nodes and four indirectly connected nodes. The nodes th at are not directly connected to the node “Person” can still be linked to it through intermediate nodes (i.e., FOAF applied) [26).

O Direct Connection o Indirect Connection Figure 1. Social network graph

In a social network, there can be more than one path between two nodes. Each of such paths has its own trust score. This section describes the computational methods for evaluating trust and reputation in a social network. This section is composed of two parts

1. the computational model th at evaluates the trust score of a single path and 2. the method th at combines the trust scores of multiple paths.

2.2 Trust-Based System

Trust and reputation th at are established in a natural manner through social con­ tacts and activities play a significant role in business. With the rise of online markets.

1098 S. Kim, W. Ha, J. Seo, S. Han, M. Kim

the roles of such social and psychological factors in online business have attracted considerable research interest. Trust is an essential component of building any rela­ tionship between individuals/organizations. Reputation is the opinion (or expecta­ tion) of the public toward a person based on his/her actions [6]. Trust and reputation exert their influence on every activity and technology involving interactions between people, and serve as a barometer to estimate the degree of trustworthiness of the potential counterparts [3]. Lately, there has been increasing research on formalizing trust and reputation via computational models.

Marsh proposed a computational model for trust that is applicable to the domain of Distributed Artificial Intelligence (DAI) [2], In this model, trust is represented as a subjective real number between —1 and +1. The model is simple but exhibits problems at the extreme values, and it has trouble dealing with negative trust values. A contribution of this work is its detailed exploration of the possibilities ot future work in the issue of formalizing tru st as a computational cdncept.

Zacharia et al. proposed reputation mechanisms th at rely on collaborative rating and personalized evaluation of the various ratings assigned to each user in the context of electronic commerce [7j. Their mathematical formulation dynamically evaluates the user's reputation with respect to a certain topic or criterion, instead of storing and using the net rating scores as they are.

Gao et al. proposed a comprehensive multidimensional model, which contains crucial factors having been researched and commonly accepted by precious schol­ ars [28]. They analyzed the intrinsic character and importance of each factor, includ­ ing the interaction between consumer trust and purchase intention. In addition they show the ranking results of their model on the five famous E-commerce websites.

Kim et al. proposed an identity management-based social trust model in order to mediating information sharing and information protection in online social net­ works [29]. This model solves the sparsity problem by using relationship model between users, quantified through the chronological records of users. Furthermore, the proposed social trust model has minimized unnecessary information leakages through active identity management.

Resnick et al. analyzed a way to increase the reliability of a system using a feed­ back rating mechanism in online transactions [1]. In their model, reputation is taken to be a function of the cumulative positive and non-positive ratings of a seller or buyer. In addition, trust by one agent of another is evaluated by an implicit mecha­ nism in which the ratings that an agent receives from others are taken into account. Their algorithm was designed to be applicable to eBay reputation systems.

Mui et al. distinguished the difference between trust and reputation, and pro­ posed a mathematical model to calculate agents’ trust and reputation on a prob­ abilistic basis [3]. They defined reputation as a quantity relative to the particu­ lar embedded social network of the evaluating agent and encounter history, and an agent’s reputation score is evaluated based on the accumulated positive feedback from previous transactions. In addition, they provide a mechanism th at evalu­ ates trust of the trustor toward the trustee from the reputation data about the trustee.

A Method of Evaluating Trust and Reputation for Online Transaction 1099

Golbeck et al. presented an algorithm for aggregating and evaluating reputation and trust ratings on a semantic web-based social network [8, 12]. In their work, trust between two individuals th a t are not directly connected in the network is evaluated based on locally-calculated trust ratings of intermediate nodes. In addition, they proposed a quantitative model th a t evaluates reputation by combining the trust scores of the searched paths. The proposed method was applied to the TrustMail system, an email client that looks up the mail sender in the reputation network and provides a trust rating for each email message.

3 COMPUTATIONAL METHODS FOR EVALUATION

3.1 Trust and Reputation

Definining and extracting appropriate trust factors in online environment are im­ portant issues. Figure 2 shows a schematic representation of trust and reputation viewed in this paper. As shown in a), trust is a particular degree of the subjective probability an individual has toward another one [30], which affects the decision of whether or not to transact. In Figure 2, b) represents the notion of reputation. Re­ putation is what is generally said or believed about a person’s or thing’s character or standing [31]. Like trust, reputation is built through interpersonal interactions, and it can also be evaluated by combining several individually occurred trust ratings.

o trustor trustee

a) b)

Figure 2. a) Trust and b) reputation

A social network is created to represent the interpersonal or inter-organizational relationships in the online transaction system of interest. Each node in the network represents a user of the online transaction system, and ties represent the contacts between the users. A quantitative score assigned to each tie indicates the level of trust between the users th at are connected by that tie. This trust score is di­ rected, i.e., it is the trustor’s trust toward the trustee, but not the other way round. An assumption made is th at every user in a social network gives a subjective trust score to other users who are directly connected in the network. The given trust score is a rational number between 0 and 10. The trust score 0 is considered as “no connection” . The FOAF (Friend of a Friend) theory is adopted to evaluate trust and reputation about strangers [26]. That is, an online transaction user eval­ uates the trust and reputation information about another user th at s/he does not

Social Network

1100 S. Kim, W. Ha, J. Seo, S. Han, M. Kim

know well by searching paths to th a t stranger in the network, starting from his/her acquaintances.

3.2 A Computational Model for Inferring the Path Trust Score

As described above, there can be multiple paths between two nodes in a social network. The evaluated trust score of a certain path searched in a social network is called the “path trust score” . Since a set of nodes comprising a path is different in each path, the evaluated path trust score varies according to the path.

Direct Connection

O—O —O —O —"O trustor trustee

Figure 3. A path between two nodes of interest

Figure 3 shows a path between two nodes of interest (from the trustor to the trustee) searched in the social network. The following symbols are defined to eval­ uate the trust score of the searched path:

• Td - trust score in a direct connection • Ti - trust score in an indirect connection • Tr - evaluated trust score (path trust score).

Direct Connection(Td) Indirect C onnection^)

Q — €>— Person

Calculated Trust(Tr)

Figure 4. Schematic diagram of single path trust computation

In Figure 4, the path has three nodes on it. Ui’s trust score toward U2 connected directly in the network is a trust score in a direct connection, whereas the trust score of U2 toward U3 is considered as a trust score in an indirect connection. The trust score evaluated based on Td and 7) is called a path trust score. Tr is evaluated using the Equation (1):

j, — ^ ~ - ^ i ) ) ^ V m a x

As trust is represented as a rational number between 0 and 10, the maximum possible trust score denoted as Vmax is 10.0. In the proposed trust evaluation model, the

A Method of Evaluating Trust and Reputation for Online Transaction 1101

significance of “direct” trust (Td) and that of “indirect” trust (Tf) are not taken equally. Intuitively, it is reasonable that Td rated directly by the trustor should play a more significant role in trust computation than T, derived indirectly. Thus, a weighting scheme th at adds Td’s complement (i.e., Vmax - Td) to 7) is used to adjust the significance level between Td and 7). For example, if the trust score of U\ toward U2 (Td) is 9, it is presumed th at the distrust level of U\ toward U2 is symmetrically 1 (i.e., Td s complement is 1). This distrust level toward U2 becomes a weight value assigned to 7) th at is a trust score given by U2. The path trust score Tr is evaluated by multiplying Td with 7) weighted by Td s complement, and then dividing the multiplication result by Vmax-

This trust computation process is applied recursively for the consecutive nodes on a single path in the network until all the intermediate nodes between the trustor and the trustee are covered. T hat is, an evaluated path trust score Tr evaluated using Equation (1) becomes a new Td for the adjacent node in the next cycle of computa­ tion. Table 1 shows the pseudo code of the proposed trust evaluation method.

func C a lc u la te P a th T r u s t (User , n e x tU ser) { if n e x tU s e r . next != T arg etU ser th en

d i r e c t T r u s t = U ser, t r u s t i n d i r e c t T r u s t = nex tU ser . t r u s t U ser, t r u s t = d i r e c t T r u s t *

( i n d i r e c t T r u s t + (MaximumV alue- d i r e c t T r u s t ) ) / MaximumV alue nextU ser = n extU ser . next C a lc u la te P a th T r u s t (User , nextU ser )

e lse o u tp u t U s e r . t r u s tJ______________________________

Table 1. Pseudo code of the evaluating the path trust score

As described in this section, the trust score of each path between a trustor and a trustee in a social network can be evaluated using the proposed trust evaluation model for a single path.

3 .3 A M e t h o d o f C o m b in in g M u ltip le P a t h T r u s t S c o re s

This section proposes a method that combines the trust scores of multiple paths between two nodes of interest (i.e., a trustor and a trustee). The trust scores to be combined can be either trust scoi*es given explicitly by the trustor (i.e., they are attached to the path in the direction of the connection between the trustor and the trustee) or evaluated trust scores evaluated using the trust evaluation method for a single path. As a set of nodes engaged in a certain path is different, the trust scores of multiple paths to be combined are different from one to the other. Instead of evenly aggregating all the trust scores, the combining method proposed in this section employs a weighting scheme th at puts more weight on a path having its trust

1102 S. Kim, W. Ha, J. Seo, S. Han, M. Kim

score close to the average trust score of the multiple paths concerned, The purpose of this weighting scheme is to prevent the resulting combined trust value from being directly influenced by a few extreme trust scores:

• Tr - combined trust score • Ti - trust score associated with ith path • Wi - weight assigned to ith path

trustee

Result Value(TR) * ‘ ' ‘ ' --- >

Figure 5. Schematic diagram of combining trust values of multiple paths

As shown in Figure 5, there is more than one path to be combined to evaluate the trustee’s tru st score T r . T represents the tru st score of 1th path, and W* is the weight given to «th path. Equations (2)-(4) are used to calculate the combined trust score T r .

W i

VFi

Tr

(Fmax - |aVg(TB) - T |)2 Wi

E (Wi) E cn *

( 2)

(3)

(4)

The trust score of an individual path th at is close to the average trust score of all the searched paths is given a greater weight. This weighting scheme reduces the influence of a few extreme trust scores, and thus, allows more reliable trust evaluation. The sum of all the weights assigned to the paths should .be 1.0. Table 2 shows the pseudo code of the proposed combining method.

4 COMPUTATIONAL MODEL FOR INFERRING TRUST

Based on FOAF implying the ability to access information through the “grape vine” of network members [26], the proposed trust evaluation model can evaluate the trust score of any member in the network by searching the paths to th at member.

As shown in Figure 6, there are several paths connecting the trustor to the trustee. The trust score of each single path can be evaluated using the method described in Section 3.1, and the evaluated path trust scores are combined using the

A Method of Evaluating Trust and Reputation for Online Transaction 1103

fu n c C o m b in eT ru st ( s e t of T r u s tT o S in k ) { fo r each Trust, Tj {

T j. wi = S q u are (MaximumValue — A b s o lu te V a lu e ( A verage ( s e t of Tj) — T j))

}

fo r each T r u s t Tj { Tj.W i = T j.w i / A verage ( Tj .Wi)

}

fo r each T r u s t Tj { C o m b in e d T ru st V alue += T j. T r u s t V alue * T j. IT;

}

o u tp u t C o m b in e d T ru st Value I_____________

Table 2. Pseudo code of the combining method

Figure 6. Trust evaluation model

combining m ethod presented in Section 3.2. In this way, a subjectively perceived level of tru s t toward th e tru stee is systematically transform ed into quantitative information (tru st scores). Table 3 shows the pseudo code of th e proposed tru st evaluation model.

fu n c I n f e r r i n g T r u s t ( S ource , T a r g e tU s e r ) { fo r each P a th S o u rce to T a r g e tU s e r P {

Pj. T r u s t = C a l c u l a t e P a t h T r u s t ( S ource , P j.N e x tU s e r) }

o u tp u t C o m b in eT ru st ( s e t o f P j. T r u s t ) }

Table 3. Pseudo code of the combining method

1104 S. Kim, W. Ha, J. Seo, S. Han, M. Kim

5 COMPUTATIONAL MODEL FOR INFERRING REPUTATION

For successful online transactions, one should be able to assess the trustworthiness of trading partners. Trust and reputation are typical factors related to trustworthiness. This section addresses the factor “reputation” and proposes a formal model that evaluates reputation in a social network.

Every member in a social network has its own reputation in th at particular domain. Reputation refers to a judgment of trustworthiness toward a certain net­ work member, made collaboratively by other members in the same network. Re­ putation is an objective and collective concept that gathers more than one mem­ ber’s personal trust. This section proposes a computational model th at evaluates a trustee’s reputation by combining the trust scores of other members toward the trustee.

As shown in Figure 7, the node of interest “Person” has a group of mem­ bers who are directly connected to (the primary group denoted as Gi). Next, the node has a group of members who are indirectly connected via one interme­ diate node. This group is called the secondary group and is denoted as G2. Simi­ larly, the node continues to have the next group of adjacent nodes th at are linked through an increased number of intermediate nodes each time. Those groups are collectively denoted as Gi. Other members in the same network (both directly and indirectly connected) have subjective trust scores toward the node Person, so the Person’s reputation can be evaluated by combining those individual trust scores.

In the proposed reputation evaluation model, the trust scores of other mem­ bers in the network toward the member of interest, Person, are evaluated first. For the nodes in Gi, the directly given trust scores are used as they are. To get the trust scores of the members belonging to the groups other than G\ (i.e., indirectly connected nodes), the trust evaluation model presented in Section 4 is used. The computed trust scores of multiple members are then combined to derive a collective

A Method of Evaluating Trust and Reputation for Online Transaction 1105

reputation score about the node Person. As described below, the proposed reputa­ tion evaluation model uses a variable that limits the range of the connection to be considered in the reputation evaluation:

• l - variable for the connection range limitation.

The value of the variable / is used to limit the range of the connection to be considered in reputation evaluation. With this value, the nodes that are too distant to be significant are excluded, and those closely linked to the node of interest are concerned. According to the given value of l, only the nodes belonging to the groups within the boundary of this value (i.e., from Gx to Gf) are taken into account in evaluating reputation, thereby reducing the computational load. Table 4 presents the pseudo code of the proposed reputation evaluation model.

func I n f e r r i n g R e p u ta t io n ( T a rg e tU s e r) { for each User Ui a d ja c e n t to source w ith in l { Ui. T ru stT o T arg e tU se r = I n f e r r i n g T r u s t (Ui , Ta r get Us e r ) } o u t p u t Com bineTrust ( s e t of [/*. T r u s t )J_________________

Table 4. Pseudo code of the reputation evaluation model

6 A METHOD OF COMBINING TRUST AND REPUTATION FOR RELIABLE ONLINE TRANSACTIONS

In conventional marketplaces, the trustworthiness of trading partners is estimated over trust evidences such as direct experiences from former encounters, witness infor­ mation and information about past transactions. However, such information sources are not available or are very limited in an online setting where there are no direct, physical contacts. The computational models based on social networks described in Sections 4 and 5 can be used to systematicallv evaluate trust and reputation in such online settings.

This section presents the method that combines the evaluated trust and rep­ utation for reliable online transactions. As described earlier, trust is a subjective judgment of trustworthiness between two trading partners, and reputation is a col­ lective assessment of someone’s trustworthiness made by multiple members in the network. According to the characteristics of the trades in online transactions, some­ times the subjective attributes associated with trust might be more important, and there might be other cases in which the objective attributes associated with reputa­ tion play a more significant role. There can also be some cases where both subjective and objective attributes should be evenly considered. To address this point, this pa­ per makes use of the fuzzy logic [24, 25] and constructs a computational model that

1106 S. Kim, W. Ha, J. Seo, S. Han, M. Kim

com bines tr u s t an d re p u ta tio n . In th is m odel, tw o tru stw o rth in e ss factors “t r u s t ” an d “r e p u ta tio n ” becom e fuzzy d escriptors.

Trust = {Low, Medium, High, Very High} Reputation = {Very Low, Low, Medium, High, Very High}

Figure 8. Fuzzy graph: a) tru st, b) reputation

F ig u re 8 shows th e fuzzy g rap h s reg ard in g tr u s t an d re p u ta tio n . Such fuzzy g rap h s can b e flexibly c o n stru c te d according to th e m em bers an d th e ty p e s of on­ line tra n sa c tio n s applied. In th e fuzzy set graphs, th e fuzzy d escrip to r T ru st has four m em bership values - low, m edium , high an d very high. The fuzzy d escrip to r R e p u ta tio n has five m em bership values - very low, low, m edium , high an d very high. T h e tr u s t an d re p u ta tio n scores p ro d u ced using th e proposed evalu atio n m odels are m a p p e d to th o se fuzzy m em bership values, an d th e m ost relevant one (i.e., having th e highest m ap p in g value) is selected.

T able 5 shows a fuzzy rule base w hich defines th e fuzzy sets an d m em bership values shown in F ig u re 9. T h e re are 20 cases w ith reg ard to two fuzzy d escriptors, 4-scale T ru st a n d 5-scale R e p u ta tio n , an d th e resu lt value of each case is listed. T h e com bined re su lt value ta k in g into acco u n t b o th tr u s t an d re p u ta tio n can be derived from th e tab le.

In ad d itio n , a fuzzy g ra p h as shown in F igure 9 is created to ev alu ate a q u a n tita ­ tiv e level of tru stw o rth in e ss by sim u ltaneously considering tr u s t a n d re p u ta tio n . In th e g rap h , th e re are seven m em bership values - very low, low, ra th e r low, m edium , ra th e r high, h ig h an d very high. T h e re su lt value is d e term in ed by com bining th e ra ­ tios of tw o fuzzy d escrip to rs T ru st a n d R e p u ta tio n . In th is way, th e tru stw o rth in e ss o f a n online u ser can b e q u a n tita tiv e ly ev aluated.

7 EXPERIMENTAL RESULT

T h e accuracy of th e prop o sed evalu atio n m odels for tr u s t and re p u ta tio n are eval­ u a te d th ro u g h e x p e rim e n ta l sim ulations. As m entio n ed earlier, every m em ber in

A Method of Evaluating Trust and Reputation for Online Transaction 1107

Trust Reputation Result 1 Low vLow vLow 2 Low Low Low 3 Low Med rLow 4 Low High Med 5 Low vHigh rHigh 6 Med vLow Low 7 Med Low rLow 8 Med Med Med 9 Med High rHigh

10 Med vlliph High 11 High vLow rLow 12 High Low Med 13 High Med rHigh 14 High High High 15 High vHigh vHigh 16 vHigh vLow rLow 17 vHigh Low Med 18 vHigh Med rHigh 19 vHigh High High 20 vHigh vHigh vHigh

Table 5. Fuzzy rule base (vLow = very low, rLow = rather low, rHigh = rather high, vHigh = very high)

a social network has a trust rating toward other directly connected members (a ra­ tional number between 0 and 10.0), and this trust score is directed (i.e., the trustor’s trust level toward the trustee, not vice versa).

7.1 E x p e r im e n t 1

Trust is the trustor’s subjective judgment of trustworthiness toward the trustee, so its accuracy is evaluated by solely considering the relationships between the trustor and the trustee. A social network is created to evaluate the proposed trust evaluation model, and several attributes are given to each member in the social network. In the created social network, each member has 20 directly connected nodes on average, and the trust scores given to those directly connected nodes are determined by assessing the similarity of the attributes attached to the two nodes. Based on the given trust scores of directly connected nodes, the proposed trust evaluation model evaluates the trust score of a trustor toward a trustee. The accuracy of the proposed model is then evaluated by comparing the evaluated trust score with the trust score directly given to the trustee earlier based on the similarity of the associated attributes. The experimental settings are as follows:

1108 S. Kim, W. Ha, J. Seo, S. Han, M. Kim.

Result = {Very Low, Low, Rather Low, Medium, Rather High, High, Very High}

Figure 9. Result graph

• th e n u m b er of m em bers in th e netw ork: 500 • th e average n u m b e r of d ire c tly co n n ected m em bers: 20 • th e n u m b e r of sim ulations- (experim ent rep etitio n s): 1000.

T h e p roposed tr u s t ev alu atio n m odel is com pared to T id a lT ru st, conventional tr u s t evalu atio n m odel [12]. T id a lT ru st, a tr u s t netw ork inference alg o rith m , is used as th e basis for g e n eratin g p red ictiv e ra tin g s personalized for each user. T h e accuracy of th e recom m ended ra tin g s is shown to o u tp erfo rm b o th a sim ple average ra tin g a n d th e ra tin g s p ro d u ced by a com m on recom m ender system algorithm .

Average accuracy Our model 97.2% TidalTrust 93.4 %

Table 6. Trust evaluation model experiment

As shown in T able 6, th e accuracy of th e proposed m odel increases by 3.8 % com ­ p ared to T id a lT ru st. In T id a lT ru st, evaluation accuracy decreases as th e len g th of th e connection p a th betw een tw o nodes increases. O n th e o th e r h an d , th e p roposed m odel m a in ta in s th e accuracy irresp ectiv e of th e p a th length. It is n o ta b le t h a t th e re p u ta tio n ev alu atio n m odel p roposed in th is p a p e r evaluates re p u ta tio n by com ­ b in in g th e tr u s t scores ev alu ated using th e p roposed tr u s t evalu atio n m odel, so th e accuracy of th e tr u s t evalu atio n m odel show n in th is ex p erim en t also d em o n stra te s th e accuracy of th e p ro p o sed re p u ta tio n evaluation m odel to som e degree.

7.2 E xperim ent 2

In th e second ex p erim en t, th e accuracies of th e proposed tr u s t evalu atio n m odel, re p u ta tio n ev alu atio n m odel an d th e m e th o d of com bining th e ev alu ated tr u s t scores

A Method of Evaluating Trust and Reputation for Online Transaction 1109

are examined. Each member in the social network receives a presumed exact trust score (called “standard value”). The standard value serves as a barometer against which the trust score evaluated using the proposed trust evaluation model is com­ pared, in order to assess the accuracy of the proposed model. As in Experiment 1, each node in the network gives a subjective trust score to the directly connected node. The difference here is that the given trust score is relative to the standard value Sy and the assigned rating accuracy R A- The trust scores given to the directly connected nodes are used in evaluating trust and reputation using the proposed eval­ uation models and combining methods. The evaluated trust and reputation scores are compared to the standard values, so as to evaluate the accuracy of the proposed evaluation models and combining method:

• Sy - standard value • Ra ~ rating accuracy.

500 members (nodes) are created in the network. The standard value for each member is given to form a normal distribution with the average 5.0 (i.e., the given standard values of 500 nodes cluster around trust score 5.0). Once the standard value Sy is assigned to every member in the network, a trust rating score (a rational number between 0 and 10.0) is given to the directly connected node based on RA and the node’s Sy. For example, if RA is 100%, then the given trust score toward a directly connected node is the same as Sy of that node. As RA decreases by 10%, the difference between the trust score given to the directly connected node and its Sy increases by 1.0 (±0.5). As shown below, the accuracy of the evaluated trust and reputation scores is evaluated by varying R x

• avg(Sy) - 5.0 • R a range - 0 ~ 100.0 (%) • the number of network members: 500 • the average number of directly connected members: 20 • the number of simulations (experiment repetitions): 1000.

Figure 10 shows the comparison of the evaluated trust and reputation scores to the standard values as the simulations are repeated by increasing R a. When R a is low, the accuracy of the evaluated trust and reputation is relatively low; but their accuracy improves as R A increases. This result indicates th at R a directly set by users considerably influences the accuracies of the proposed evaluation models. In evaluating the trust score of a single path, the evaluated trust score varies accord­ ing to the path chosen, and thus, the resulting accuracy graph does not increase consistently.

Table 7 shows the average difference (or average error) between Sy and the trust and reputation scores evaluated using the proposed evaluation models as the simulations continue by increasing Ra . A s R a increases, the difference be­ tween Sy and the evaluated trust and reputation decreases. In other words, as R a

1110

Result Accuracy (%)

i#o

Trust

S. Kim. W. Ha. J. Seo. S. Han, M. Kim

80

60

40

20

0 Rating Accuracy (%)

— Trust

Reputation

— Reputation

Figure 10. Accuracy of the evaluated tru st and reputation

Ra Average Error: Trust Average Erroi•: Reputation 80% 81.142% (0.9429) 79.746% (1.0127) 85% 85.694% (0.7153) 85.164% (0.7418) 90% 90.542% (0.4729) 90.918% (0.4541) 95% 92.832% (0.3584) 93.290% (0.3355) 100% 96.198% (0.1901) 96.410% (0.1795)

Table 7. Average accuracy by R a (average error)

increases, the evaluation accuracy of the proposed models increases and becomes nearly identical to R a • When R a is 100%, the accuracy of the proposed trust eval­ uation model is 96 %, and the accuracy of the proposed reputation evaluation model is 97%.

Trust, Reputation Our model 96.198% 96.410%

J. Golbeck’s model 91.286% 92.572%

Table 8. Comparison w ith other evaluation model (Ra - 100%)

A Method of Evaluating Trust and Reputation for Online Transaction 1111

Table 8 shows the comparison of the proposed models for evaluating tru st and reputation to conventional evaluation methods, trust and the reputation evalua­ tion model proposed by Golbeck, when RA is 100% [8]. That is, the trust and reputation scores evaluated using the evaluation model proposed in this paper and those evaluated using conventional methods are compared with SV - It is ex­ pected th at the evaluated scores should be equivalent to Sy because RA is 100%. The results in Table 4 show th at the error ratio of the proposed trust and rep­ utation evaluation models to conventional models is 0.4, so the proposed models can yield more accurate trust and reputation evaluation than the conventional me­ thods.

8 C O N C L U S IO N

This paper has presented the computational models th at evaluate trust and reputa­ tion from a social network representing human relationships. Trust and reputation are an antecedent to a successful online transaction, so the proposed evaluation models for trust and reputation can contribute to promoting online transactions th at offer many advantages in terms of cost and convenience. This paper has ana­ lyzed how trust and reputation are acquired and how they are used in traditional offline environments, and this paper proposed formal models to systematically eval­ uate trust and reputation in online transaction environments. In addition, this paper has proposed a method that flexibly combines the evaluated trust and re­ putation according to the characteristics of the transactions. The proposed eval­ uation models can serve as a framework th at transforms the sociological concept “trustworthiness” into quantitative information applicable in online systems. The proposed models contribute also to increasing the overall reliability of a social net­ work by offering an accurate way to gauge network members’ trust and reputation levels.

One of the practical limitations of this work is that it requires explicit trust ratings to evaluate tru st and reputation. The trust scores given by a user are based on the user’s subjective judgment, so the accuracy of the given trust, scores varies depending on which user is in charge of the ratings. The proposed evaluation models for trust and reputation are demonstrated in terms of the accuracy of the evaluation process, but they do not currently address the subjective probability regarding trust. To improve this problem, generalizing and refining the valid range of explicit trust ratings will be studied in our future works.

A ck n o w led g em en ts

This research was supported by the Chung-Ang University Research Scholarship Grants in 2012 and the Basic Science Research Program through the National Re­ search Foundation of Korea (NRF) funded by the Ministry of Education, Science and Technology (No. 2014018334).

1112 S. Kim, W. Ha, J. Seo, S. Han, M. Kim.

REFERENCES

[1] R esnick, P .— Zeckhauser, R.: Trust Among Strangers in Internet Transactions: Empirical Analysis of eBay’s Reputation System. Advances in Applied Microeco­ nomics, Vol. 11, 2002, pp. 127-157.

[2] Marsh, S.: Formalizing Trust as a Computational Concept. Ph. D. thesis. University of Stirling, Scotland, 1994.

[3] Mui, L.— Mohtashemi, M.—Halberstadt, A.: A Computational Model of Trust and Reputation for E-Businesses. Proceedings of the 35th Annual Hawaii International Conference on System Sciences (HICSS’02), 2002, Vol. 7, p. 188.

[4] Gefen, D.— Staub, W.: Consumer Trust in B2C E-Commerce and the Importance of Social Presence: Experiment in E-Products and E-Services. Omega, Vol. 32, 2004, No. 6, pp. 407-424.

[5] Steward, K. J.: Trust Transfer on the World Wide Web. Organization Science, Vol. 14, 1999, pp. 459-464.

[6] McKnight, D .H .—Chervany, N .L .: What Trust Means in E-Commerce Cus­ tomer Relationships: An Interdisciplinary Conceptual Typology. International Jour­ nal of Electronic Commerce, Vol. 6, 2002, No. 2, pp. 35-53.

[7] Zacharia, G .— Moukas, A.— Maes, P.: Collaborative Reputation Mechanisms for Electronic Marketplaces. Decision Support Systems, Vol. 29, 2000, pp. 371-388.

[8] G o l b ECK, J . —H e n d l e r , J.: Accuracy of Metrics for Inferring Trust and Reputa­ tion in Semantic Web-Based Social Networks. EKAW (Engineering Knowledge in the Age of the Semantic Web), 2004, pp. 116-131.

[9] G olbeck, J .— Hendler, J.: Reputation Network Analysis for Email Filtering. Proceedings of the First Conference on Email and Anti-Spam (CEAS), Vol. 44, 2004, pp. 54-58.

[10] Adamic, L .A .— Adar, E.: How to Search a Social Network. Social Networks, Vol. 27, 2005, No. 3, pp. 187-203.

[11] G olbeck, J . —Hendler, J.: Inferring Binary Trust Relationships in Web-Based Social Networks. ACM Transactions on Internet Technology, Vol. 6, 2006, No. 4, pp. 497-529.

[12] G o l b e c k , J.: Computing and Applying Trust in Web-Based Social Networks. Doc­ toral Dissertation, University of Maryland, College Park, 2005.

[13] D umais, S .T .— Cutrell, E .— Cadiz, J . J . —J ancke, G .— Sarin, R .—R ob­ bins, D. C.: Stuff I’ve Seen: A System for Personal Information Retrieval and Re- Use. SIGIR’03 (Special Interest Group on Information Retrieval), 2003, pp. 72-79.

[14] Borgatti, S. P .— C ross, R.: Relation View of Information Seeking and Learning in Social Networks. Management Science, Vol. 49, 2003, No. 4, pp. 432-445.

[15] K atz, Y .—G olbeck, J.: Social Network-Based Trust in Prioritized Default Logic. Proceedings of the 21st National Conference on Artificial Intelligence, Vol. 2, 2006, pp. 1345-1350.

[16] SZOMSZOR, M.: Semantic Modelling of User Interests Based on Cross-Folksonomy Analysis. ISWC 2008, LNCS, Vol. 5318, pp. 632-648.

A Method of Evaluating Trust and Reputation for Online Transaction 1113

[17] Ziegler, C.-N.— Golbeck, J.: Investigating Interactions of Trust and Interest Similarity. Decision Support Systems, Vol. 43, 2007, No. 2, pp. 460-475.

[18] Paolucci, M.— Soudry, J . —Srinivasan, N.—Sycara, K.: A Broker for OWL-S Web Services. In: Cavedon, L., Maamar, Z., Martin, D., Benatallah, B. (Eds.): Ex­ tending Web Services Technologies: The Use of Multi-Agent Approaches. Kluwer, 2005.

[19] Sycara, K .—W ido ff, S.— Klusch, M.—Lu, J.: LARKS: Dynamic Matchmak­ ing Among Heterogeneous Software Agents in Cyberspace. Autonormous Agents and Multi-Agent Systems, Vol. 5. 2002, pp. 173-203.

[20] Golbeck, J.: Generating Predictive Movie Recommendations from Trust in Social Networks. 4th International Conference on Trust Management (iTrust’06), LNCS, Vol. 3986, 2006, pp. 93-104.

[21] O ’Donovan, J .— Smyth, B.: Trust in Recommender Systems. Proceedings of the 10th International Conference on Intelligent User Interfaces (IUI ’05), ACM, New York, 2005, pp. 167-174.

[22] G olbeck, J .— Hendler, J.: Film Trust: Movie Recommendations using Trust in Web-Based Social Network. Consumer Communications and Networking Conference (CCNC 2006), IEEE 2006 3rd Publication, 2006.

[23] Hotta, H.—Hagiwara, M.: Profiling System Using Social Networks for Recom­ mendation. Proceedings of International Symposium on Advanced Intelligent Sys­ tems, 2007, pp. 207-274.

[24] Zadeh, L. A.: Fuzzy Logic = Computing with Words. IEEE Transactions on Fuzzy Systems, Vol. 4, 1996, No. 2, pp. 103-111.

[25] Zadeh, L. A.: Fuzzy Sets as a Basis for a Theory of Possibility. Fuzzy Sets and Systems, Vol. 1, 1978, pp. 3-28.

[26] G raves, M.— C onstabaris, A.— Brickley, D.: FOAF: Connecting People on the Semantic Web. Cataloging & Classification Quarterly, Vol. 43, 2007, No. 3-4, pp. 191-202.

[27] K uter, U.— G olbeck, J.: SUNNY: A New Algorithm for Trust Inference in Social Networks Using Probabilistic Confidence Models. Proceedings of the 22nd National Conference on Artificial Intelligence (AAAI ’07), Vol. 2, 2007, pp. 1377-1382.

[28] Gao, H.—Yao, J.: A Multi-dimensional Trust Model in E-Commerce. In 2012 IEEE International Conference on Computer Science and Automation Engineering (CSAE), 2012, Vol. 2, pp. 653-657.

[29] Kim, M.—Seo, J .—Noh, S.— Han, S.: Identity Management-Based Social Trust Model for Mediating Information Sharing and Privacy Enhancement. Security and Communication Networks, Vol. 5, 2012, No. 8, pp. 887-897.

[30] G randison, T .—Sloman, M.: A Survey of Trust in Internet Applications. Com­ munications Surveys & Tutorials, IEEE, Vol. 3, 2000, No. 4. pp. 2-16.

[31] J osang, A.—Ismail, R .— Boyd, C.: A Survey of Trust and Reputation Sys­ tems for Online Service Provision. Decision Support Systems, Vol. 43, 2007, No. 2, pp. 618-644.

1114 S. Kim, W. Ha, J. Seo, S. Han, M. Kim

S u n g h o n g K im received his B. Sc. degree in computer science and engineering from Chung-Ang University, South Korea, in 2008. Having got a B. Sc. degree, he took a master course in E-Commerce Application Laboratory, School of Computer Science and Engineering, Chung-Ang University. His research interests include trust and reputation model, decision making, and fuzzy theory.

M u c h e o l K im is an Assistant Professor in a group of Industry University Cooperation at Sungkyul University in Korea. He was a Senior Researcher in Korea Institute of Science and Technology Information (KISTI), Daejeon, Korea. He received the B.Sc., M. Sc. and Ph. D. degrees from the School of Computer Science and Engineering at Chung-Ang University, Seoul, Korea in 2005, 2007 and 2012, respectively.

W o o k ju H a received his Bachelor of Science in computer scien­ ce and engineering from Chung-Ang University, South Korea, in 2012. Having got a Bachelor of Science degree, he took a master course in E-Commerce Application Laboratory, School of Com­ puter Science and Engineering, Chung-Ang University. His re­ search interests include trust and reputation model, and know­ ledge ecosystem.

J iw a n Seo received his Bachelor of Science in 2010 and Master degrees in computer science and engineering, Chung-Ang Univer­ sity in 2012; he is Ph. D, candidate in E-Commerce and Internet Application Laboratory, School of Computer Science and Engi­ neering, Chung-Ang University. His research interests include online social networks, trust and reputation model, and big data analysis.

A Method of Evaluating Trust and Reputation for Online Transaction 1115

S a n g y o u n g H an is a Professor of the School of Com puter Scien­ ce and Engineering, ChungAng University, Korea. He received Bachelor of Engineering from College of Engineering, Seoul Na­ tional University in 1975, and the Ph. D. degree in 1984. Between 1984 and 1995, he worked at Poughkeepsie Lab. and Watson Re­ search Center in IBM, USA. His research interests include web technologies, web services, semantic web, information retrieval and multimedia.

Copyright of Computing & Informatics is the property of Institute of Informatics and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use.