The 28th International Conference on Distributed Computing Systems Workshops

Computing Reputation Metric in Multi-agent E-commerce Reputation System Anna Gutowska School of Computing and IT University of Wolverhampton, UK [email protected]

Dr Kevan Buckley School of Computing and IT University of Wolverhampton, UK [email protected]

Abstract

trusted in an online transaction. One of the important aspects in such decisions is a third party’s reputation, which is usually based on information about its past behaviour. There are many aspects however, which can influence reliability of the reputation rating calculated for a provider. They are often considered drawbacks of existing systems. Such aspects include time factors, transaction value, source of feedback, malicious incidents, unfair ratings, reluctance to submit feedback, and changing identities [7]. Current researchers discuss these issues in a reasonably detailed way and propose some solutions to these problems which are summarized in [7]. They tend to be still “one issue-centric”. The approach adopted here is more comprehensive as it addresses all of the above aspects. Past behaviour however, is not the only information source affecting trust/reputation rating of an online vendor. According to previous research, there are many issues influencing trust-based decisions. They have been presented in a form of Trust Taxonomy [1]. The taxonomy synthesises and brings together the viewpoints of trust from across different disciplines. Using this taxonomy, the system presented here does not rely only on information based on past behaviour but also incorporates other aspects affecting online trust. This yields an improved distributed agent-based reputation mechanism that considers additional trust-related factors such as existence of control mechanisms and supporting organisations, security and privacy strategies, technology and information-based factors as well as interaction. As a part of the proposed distributed agent-based reputation system, this paper presents an improved reputation metric model that works on the basis of the parameters identified above.

Trust and reputation systems have recently come into the focus for the virtual environment research to offer online user a decision support mechanisms. Many of them are agent-based as they represent a promising paradigm for open, distributed marketplaces. This paper presents an improved distributed agent-based reputation mechanism which contributes in measuring reputation of online providers. The system offers a comprehensive approach as it considers a number of the parameters that have a bearing on trust and reputation (unlike current approaches which tend to be “one issue-centric”). Moreover, it also extends the existing frameworks based on information about past behaviour, with other aspects affecting online trust, taken from the Trust Taxonomy [1], such as existence of control mechanisms and supporting organisations, security and privacy strategies, technology and information based factors as well as interaction.

Keywords

Reputation systems, Rating aggregation algorithm, Trust, E-commerce, Agents

1. Introduction Trust and reputation systems have come into the focus for the virtual environment research to offer an online user a decision support mechanism. The basic idea is to let parties rate each other and use the aggregated ratings about a given party to derive a trust or reputation rating. This can assist in deciding whether or not to engage in a transaction with this party. Many proposed reputation systems are agentbased [2-6] as software agents are a promising paradigm for open, distributed marketplaces. This paper focuses on the problem of trust, i.e. deciding whether another agent or third party can be

1545-0678/08 $25.00 © 2008 IEEE DOI 10.1109/ICDCS.Workshops.2008.38

255

the reputation rating about a particular provider the only source of information will be their feedback from the past and feedback received from other users. Also, when entering the system as a new member they are duty-bound to both. They rate transactions they were involved in and share this information with others when requested. The cooperation parameter mentioned above acts as an incentive mechanism to motivate users to share the reputation ratings. While computing the reputation rating for a provider the time frame and transaction values are taken into account (see Section 3.1.2).

2. Proposed Reputation System The purpose of the proposed agent-based distributed reputation system is to collect and provide reputation ratings about web services for its users.

2.1. The Framework There are two main roles considered in the framework: user agent i.e. agent representing a user (ai ) and provider (web service) (pk). A user agent collects for its user the distributed reputation ratings about a web service (provider). In return, a user provides its agent with ratings about a transaction in order to build the reputation database of the services. Agents create a network where they exchange transaction ratings about web services their users have dealt with, this is called a buyers’ coalition. In this way they are involved in a joint recommendation process. User agents and providers are engaged in a transaction process e.g. buying-selling, where money and products/services are involved. To assess the reputation of a provider, first, a user agent will use the information from the direct interactions it has had with that party and second, the ratings provided by other agents (indirect ratings) from the buyers’ coalition which have dealt with the provider. The proposed reputation system is distributed where each user agent will store their opinion about transactions with other parties as well as reputation ratings received from other users as a result of previous requests. This is in case, when there are few or no participants available at the time of rating request from others. Agents representing users would not obtain any benefits from lying when sharing the ratings about transactions, as they do not compete with each other but instead work together as the buyers’ coalition. Moreover, the users are not involved in any transactions with each other, but only in recommendation processes. Hence, there is no reputation computed for the users which could be used to choose agents to gather the feedback from. Therefore, the problem of lying witnesses (credibility of referees) or changing identity is not crucial in this case. However, the parameter that the user agents can be judged against is cooperation, i.e. if the agent shares its ratings of transactions with a provider when requested. The assumption is that it is in users’ interest to leave feedback after each transaction as that is the only way the reputation system will work. The participants are aware that if they want to calculate

2.2. Model Parameters There are two sets of parameters to be incorporated in the proposed model which will be used to calculate the reputation ratings. They are compulsory parameters and optional parameters. The set of optional parameters includes factors that may be of different importance to the users. Therefore, the decision to include none/some/all of them is left to them. 2.2.1. Compulsory Parameters. The choice of compulsory parameters results from an earlier literature review in the field of reputation systems [1, 7]. They are always included in calculation of the reputation ratings in the proposed model. The compulsory parameters are as follows: transaction ratings, source of feedback (direct interaction, another user or another reputation system), reputation lifetime (how old the ratings are), transaction value, number of transactions/ratings, number of malicious incidents. Further description can be found in [7]. 2.2.2. Optional Parameters. In addition to the parameters presented in the preceding section, a user may choose to include some or all of the optional parameters into calculations, which will influence the rating value of a provider. These parameters are taken form the Trust Taxonomy based on the result from the conducted survey. The optional parameters are: existence of trustmark seals, existence of payment intermediaries, existence of first party information, existence of privacy statements, existence of security/privacy strategies, existence of purchase protection/insurance, and existence of alternative dispute resolution. They are further described in [7].

256

Thus, the rating values received by agent ai already include the time and the transaction value component. Applying the weight of the feedback source, ai calculates the aggregated indirect rating value of provider pk. Combining the above with the aggregated direct rating calculated before and applying the weight of the number of malicious incidents agent ai comes up with compulsory rating of a provider. Information about optional parameters is not stored by agents but provided by a specialized agent, which is responsible for gathering this data from the web service of provider pk. In the proposed framework all the data is stored by individual user agents. There is no central party which holds a freely available database of transaction history. Thereby, the privacy issue is not a concern as the agents may choose or not to share information with other members of the buyer’s coalition.

2.3. Rating Model The rating model of the proposed framework (Figure 1) presents and explains how the compulsory and optional parameters form the reputation/trust ratings. The assumption is, there are no parties external to the buyer’s coalition included in this work.

3. Establishing the Reputation Metric Based on the parameters identified above a general reputation metric formula based on the weighted average we has been established. Let pk be a provider whose reputation value is calculated in any instance of time t. Similarly, let ai be an agent representing a user/buyer that belongs to the buyers’ coalition. represent the collaborated Let R(pk) reputation/trust rating of provider pk, based on the feedback provided by m agents which have dealt with provider pk. The reputation metric formula is defined as below:

Figure 1. Rating Model

R( pk ) =

Each agent collects the aggregated ratings for each provider it has had a transaction with (direct ratings). The rating value is based on the three components of the transaction ratings, i.e. transaction outcome, fulfilling signals and customers service, as well as weights for time the transactions took place and transaction values. Based on this information each agent calculates updated rating of a particular transaction with a particular provider. After each transaction the database is updated. When agent ai considers to engage in a transaction with provider pk, it uses direct ratings as well as requests the reputation ratings about this provider from the buyers’ coalition (indirect ratings). Other agents, who have dealt with provider pk, pass to agent ai the rating values of provider pk. Before the rating value is provided to agent ai, it is updated by including the time of the last transaction, which the agent sharing the information had with provider pk.

CR( pk ) + OR( pk ) 2

(1)

Where, CR(pk) is compulsory reputation metric and OR(pk) is optional reputation metric for provider pk. Further, the full rating scale of trust is 0 ≤R(pk) ≤1. If optional reputation metric is not calculated then the reputation metric takes the value of compulsory reputation metric i.e. R(pk) = CR(pk).

3.1. Compulsory Reputation Metric Compulsory reputation metric is based on the set of the compulsory parameters. Let AGRDi (pk) represent the aggregated direct rating calculated for all the transactions which the agent ai have had with the provider pk. Also, let AGRI(pk) represent the aggregated indirect rating calculated for the provider

257

pk based on AGRD(pk) provided by other agents who were engaged in transaction with pk before. Further, n(R(pk)) gives the total number of transactions that the provider pk has had with all agents by which it was reputed. Similarly, n(AGRDi (pk)) provides the total number of transactions that the provider pk has had with the agent ai, and n(AGRI(pk)) gives the total number of transactions that the provider pk has had with all other agents who shared the ratings with the agent ai. Respectively, m(R(pk)) gives the total number of agents that have reputed the provider pk, and m(AGRIi (pk)) constitutes the number of agents that provided their AGRD(pk) to agent ai. With these, we define the compulsory reputation CR(pk) of provider pk, as:

CR( p k ) =

m ( AGRI ( p k ))

j

AGRI ( p k ) = ws ⋅

(3)

ij

AGRDi ( p k ) =

( pk )

j =1

n( AGRDi ( p k ))

(5)

Where, ∆t(x) is the time difference between the current time (i.e. time of request) and the time when the transaction x took place. β is used to scale ∆t(x) and β > 1. The other weight wv is associated with the transaction value and is calculated using the formula below: (8) wv x = 1 − γ − v ( x )

3.1.1. Computing Aggregated Ratings. The aggregated direct rating value is calculated based on the data stored in the requesting agent ai database i.e. regarding its direct interactions. The formula looks as below:

∑UR

)

Where gx represents the rating for the transaction x between agent ai and provider pk. The transaction rating is the average of two components: fulfilling provider’s signals (y) and customer service (z), where both can take values [0; 1]. The weight associated with the reputation life time is defined as: (7) wt x = β − ∆t ( x )

Where, m is a number of malicious incidents of provider pk that occurred within the transactions taken into calculation. M is the set threshold of the number of malicious incidents above which the reputation value is reduced to minimum. In the equation above α is used to scale wm(pk) and α > 1. The rating scale for compulsory reputation metric is 0 ≤ CR(pk) ≤ 1.

n ( AGRDi ( pk ))

m( AGRI ( pk ))

k

3.1.2. Computing Updated Ratings. In general, each provider is reputed by an agent after each transaction by providing a transaction rating g. In addition, appropriate weights for reputation lifetime wt and transaction value wv are applied where 0 < wt, wv ≤ 1. Based on the above updated reputation rating URix(pk) is defined which is calculated by agent ai for transaction x in which ai was involved with provider pk : (6) URix ( p k ) = g x ⋅ wt x ⋅ wv x

AGRD( p k ) + AGRI ( pk ) ∗ wm( pk ) 2

0 ≤ m < M then wm( pk ) = α − m   m≥M then wm( p k ) = 0 

j =1

Where ws is the weight factor associated with the source of the feedback. Further, 0 < ws < 1. These two elements (i.e. (4) and (5)) constitute the main parts of the equation (2) which is used to calculate the compulsory reputation value of a provider.

(2) Where wm(pk) is the weight factor associated with the malicious incident component and is calculated as follows:

if if

∑ AGRD ( p

(4)

Where, v(x) is the value of transaction x, γ is used to scale v(x) and γ > 1.

Where URij(pk) stands for updated rating value of transaction j with the provider pk calculated by the agent ai (see equation (6)). The aggregated indirect rating values are calculated based aggregated direct ratings of other agents and are provided to a requesting agent along with information about the number of malicious incidents that have occurred. The aggregated indirect rating metric is defined as below:

3.2. Optional Reputation Metric The optional reputation metric is based on the set of optional parameters O = {o1, o2, o3, o4, o5, o6, o7}. Further, when chosen to be applied in calculation they take values [0; 1]. Let n(o(pk)) will be the number of optional parameters chosen to be included

258

in calculation. Based on the above, we define optional reputation metric for provider pk as follows: n ( o ( p k ))

OR( p k ) =

∑o ( p i

k

i =1

n(o( p k ))

To present the impact of weights applied, results for different cases are provided where α, β, and γ take different values, ceteris paribus (i.e. while other parameters such as transaction ratings g, time t and transaction value v stay unchanged). The evaluation parameters listed in Tables 1 and 2 are calculated when α=1.1, β=1.001, γ=1.115 and ws=0.9. The number of malicious incidents for all the transactions taken into account is 2. The parameters values have been arrived at experimentally. The values of updated ratings UR in Table 1 are calculated based on randomly chosen parameters of transaction ratings g, transaction time t and value v. In order to present how UR is obtained, the numbers for the first URi1 are shown in Table 2.

) (9)

The rating scale for optional reputation metric is 0 ≤ OR(pk) ≤ 1.

4. Evaluation and Results

In the following section experimentation with the established metrics system is presented. It is demonstrated with examples how the reputation is computed and what results can be inferred from the values. Table 1. Reputation evaluation parameters

collaborated rating, pure transaction ratings average with some additional information, as well as compulsory and optional reputation rating. For the data set presented in Table 1 and 2 the results available to the user would look like in Table 3.

Table 2. Updated rating evaluation parameters

Table 3. Results available to the user The value for the final collaborated reputation/trust rating of provider pk is obtained after applying the given values to the equations presented in Section 3. Thus, the value for R(pk) is 0.76 for the given values in the tables above. In a real world situation a score similar to that obtained above for provider pk would be utilized to decided on whether to engage in transaction with this provider or not. In order to show a wider picture of the reputation outcomes to the user placing a reputation request, the presented results will include, apart from the final

The final collaborated reputation rating is influenced, among other parameters, by time, value, and malicious incidents weights. Parameters α, β, and γ are applied to scale the weights so that they have controlled affect on the reputation value. To show that,

259

Fig. 2 presents the results from different scenarios where one of the parameters is taking values 1.001, 1.01, 1.1, and 1.115, ceteris paribus. The results are based on the same transaction data set used for Table 1 and 2.

based modelling toolkit written in Java. Based on that, further simulations will be carried out. One of the assumptions of the proposed system is that there are no external parties included in the framework. There is a possibility however, to extend this work in the future by information coming from other reputation systems or reputation authorities.

6. References [1] A. Gutowska and K. Bechkoum, "The Issue of Online Trust and Its Impact on International Curriculum Design", The Third China-Europe International Symposium on Software Industry-Oriented Education, Dublin, 6-7 February, 2007, pp. 134-140. [2] O. Bamasak and A. N. Zhang, "A distributed reputation management scheme for mobile agent-based E-Commerce applications", 2005 IEEE International Conference on eTechnology, e-Commerce and e-Service, Hong Kong, 29 March -12 April,, 2005, pp. 270-275.

Figure 2. Comparative view of reputation ratings for different values of α, β, and γ

[3] T. D. Huynh, N. R. Jennings, and N. R. Shadbolt, "Certified reputation: how an agent can trust a stranger", 5th International Conference on Autonomous Agents and MultiAgent Systems, Hakodate, Japan, 8-12 May, 2006, pp. 12171224.

In the simulations above, the values for transaction ratings g, transaction time t, and value v were randomly chosen. The aim was to present what results could be obtained from given values. Also, as the modeled system is modular in nature, it gives multiple reflections when the applied weights are changed. As a good trust model should introduce weight adaptiveness [8], this issue will be focused on in further simulation work.

[4] T. D. Huynh, N. R. Jennings, and N. R. Shadbolt, "An integrated trust and reputation model for open multi-agent systems", Journal of Autonomous Agents and Multi-Agent Systems, vol. 13, 2006, pp. 119-154. [5] J. Sabater and C. Sierra, "Reputation and social network analysis in multi-agent systems", The First International Joint Conference on Autonomous Agents and Multiagent Systems, Bologna, Italy, 15-19 July, 2002, pp. 475-482.

5. Conclusions Trust and reputation are important sources of information that we gather about each other in our daily lives. They seem even more important in the online world, due to the nature of E-commerce. This paper presents a framework for an improved agentbased reputation system. It establishes the comprehensive reputation metric based on the chosen set of trust parameters which extends the existing reputation frameworks based only on information on past behaviour with other aspects affecting online trust. A new system is to offer an online user a decision support mechanism, based on many aspects which are crucial in E-commerce activities, i.e. online shopping and information seeking. Implementation of the framework will be done using Repast which is a free and open source agent-

[6] J. Carbo, J. Molina, and J. Davila, "Comparing Predictions of SPORAS vs. a Fuzzy Reputation Agent System", 3rd WSES International Conference on Fuzzy Sets and Fuzzy Systems, Interlaken, Switzerland, 11-15 February, 2002, pp. 147-153. [7] A. Gutowska and K. Bechkoum, "A Distributed Agentbased Reputation Framework Enhancing Trust in eCommerce", The 11th IASTED International Conference on Artificial Intelligence and Soft Computing, Palma de Mallorca, Spain, 29-31 August, 2007, pp. 92-101. [8] Z. Liang and W. Shi, "Analysis of ratings on trust inference in open environments", Performance Evaluation, vol. 65, 2008, pp. 99-128.

260