Trust management through reputation mechanisms Giorgos Zacharia MIT Media Laboratory 20 Ames Street, Room 305 Cambridge MA 02139 USA 1-617-253-9833
[email protected] 1. Introduction This paper proposes reputation mechanisms that rely o n collaborative ratings and personalized evaluation of the various ratings assigned to each user. Reputation is usually defined as the amount of trust inspired by a particular person in a specific setting or domain of interest [15]. In "Trust in a Cryptographic Economy" [17] reputation is regarded as asset creation and it is evaluated according to its expected economic returns. Reputation is a multidimensional value. Therefore, individuals may have different reputations in different domains. For example, a Unix guru will naturally have a high rank regarding Linux questions, while he/she may not enjoy that high a reputation for questions regarding Microsoft’s operating systems. These individual reputation standings develop through interactions among a loosely connected group of people that share the same interests. Reputation building is a social process dependent on the past interactions between the users of a system. In agent mediated marketplaces, where agents negotiate on behalf of their users, the trustworthiness of the agents is a reflection of the reputation of their owners. In other words, the pairwise trust among the agents is derived from the respective reputations of the owners of the agents. Our research focuses on developing methods through which we can automate the social mechanisms of reputation building for the purposes of an agent-mediated marketplace for services. Our previous work on reputation mechanisms [20] was necessitated by the problems we faced in our Agent mediated marketplaces, Kasbah[1] and MarketMaker [14]. Kasbah and MarketMaker are ongoing research projects to help realize a fundamental transformation in the way people transact goods—from requiring constant monitoring and effort, to a system where software agents do much of the bidding and negotiating on their user's behalf. In these systems, a user wanting to buy or sell a specific good creates an agent, gives i t some strategic direction, and sends it off into the agent marketplace. Kasbah agents pro-actively seek out potential buyers or sellers and negotiate with them on their creator's behalf, trying to make the "best deal" possible [1]. Other examples of online market-places include ONSALE Exchange [16] and eBay [4]. OnSale Exchange and eBay are online auction sites where sellers can list their items for sale and buyers compete in an auction-like bidding system to buy the posted items. Consumer to consumer electronic trading systems like Kasbah [1], MarketMaker [14], eBay [4] and OnSale Exchange [16] create online market places that bring together users unknown to each other. These kinds of online marketplaces introduce two major issues of trust among their users:
1.
A potential buyer has no physical access to the product of interest while he/she bids or negotiates. Therefore, the seller could misrepresent the condition or the quality of his/her product or service in order to ask for higher price.
2.
The seller or buyer may not abide by the agreement reached at the electronic marketplace asking later t o renegotiate the price, or even refuse to commit the transaction.
An approach, which could solve the above-mentioned problems, would incorporate in the system a reputation mechanism. This would allow each user to customize his/her pricing strategies according to the risk implied by the reputation of the potential counterparts. For the purposes of Kasbah and MarketMaker, we implemented two complementary reputation mechanisms [20]: 1.
Sporas is a simple reputation mechanism which can be implemented irrespective of the number of rated interactions, and
2.
Histos is a more complex reputation mechanism that assumes that the system has been bootstrapped (probably by using Sporas) so that there is an abundance of rated interactions to create a dense web of pairwise ratings.
In this paper, we discuss the requirements of pairwise reputation mechanisms for trust management in multi-agent systems. We evaluate Sporas and Histos in the context of an agent mediated marketplace for services.
2. Related Work One approach to building a reputation mechanism is to have a central agency that keeps records of the recent activity of the users on the system, very much like the scoring systems of credit history agencies [7]. This central agency also keeps records of complaints by users in textual formats and even publishes warnings against possibly malicious users, much like the local Better Business Bureaus in the US [2]. However, adopting such a solution requires a lot of overhead on behalf of the providers of the online communities. Other proposed approaches are more distributed. For example, approaches such as Yenta [8] and Weaving a Web of Trust [12] require that users rate themselves and either have a central agency or other trusted users verify their trustworthiness. Both systems require the a priori existence of social relationships among the users of their online community. GroupLens [18] is a collaborative filtering solution for rating the content of Usenet articles and presenting them to the user in a personalized manner. Collaborative filtering is a technique used to detect patterns among the opinions of different users and to make recommendations to people, based on others who have shown similar taste. Collaborative filtering mechanisms such as HOMR, Firefly [19] and
GroupLens essentially automate the process of "word of mouth" to produce an advanced, personalized marketing scheme. In GroupLens, users are clustered together according to the ratings they give to several articles. Each user is then presented with the articles he/she has not seen, along with their ratings which are equal to the average of the ratings given to the articles by users in the same cluster. The most relevant computational methods to our system, t o our knowledge, are the reputation mechanisms of OnSale Exchange [16] and eBay [4]. OnSale allows its users to rate and submit textual comments about traders and overall reputation value of a trader is the average of his/her ratings through his/her usage of the OnSale system. In eBay, traders receive +1, 0 or –1 as feedback for their reliability in each auction and their reputation value is calculated as the sum of those ratings over the last six months. In OnSale, the newcomers have no reputation until someone eventually rates them, while on eBay they start with zero feedback points. In both sites the reputation value of a trader is available, with any textual comments that may exist, to the potential bidders. Due to the inadequacies of such simple mechanisms, most online auction sites have sought complimentary solutions like escrow and insurance services. However these solutions are expensive, so the pairwise rating schemes remain the primary trust management tool for users of online auctions.
3. Desiderata for online reputation systems While the above discussed reputation mechanisms have some interesting qualities, we believe they are not perfect for maintaining reputations in online communities and especially in online marketplaces. This section describes some of the problems of online communities and their implications for reputation mechanisms. Reputation mechanisms try to solve the problem of trust management in multi-agent systems where agents represent human individuals with financial goals. Hence, we assume that the inter-agent communication links are secure and that no user can falsify his/her identity to mimic an established user of the system. However, we honor users' privacy b y allowing pseudonymous identities. Therefore users may switch identities and join the system as beginners if they believe that they will be better off by doing so. In online communities, it is relatively easy to change one's identity [6][13]. Thus, if a user ends up having a reputation value lower than the reputation of a beginner, he/she would have an incentive to discard his/her initial identity and start from the beginning. Hence, it is desirable that while a user's reputation value may decrease after a transaction, it will never fall below a beginner's value. Friedman and Resnick [9] suggest that it is economically efficient to give the beginners the option to purchase reputation points for a monetary value. However, in their model they need to either charge for names i n the first place or enforce persistent pseudonymous identities. We therefore decided for the reputation mechanisms described in the following section that a beginner cannot start with an average reputation. We also want to make sure that even if a user starts receiving very low reputation ratings, he/she can improve his/her status later at almost the same rate as a beginner. For example, if the reputation value is evaluated as the arithmetic average of the ratings received since the user joined the system, the users who perform relatively poorly in the beginning will tend t o
get rid of their bad reputation history by adopting a new identity. Another problem is that the overhead of performing fake transactions in both Kasbah and OnSale Exchange is relatively low (OnSale does not charge any commission on its Exchange service). Therefore two friends might decide to many fake transactions, rating each other with perfect scores so as t o increase their reputation values. Even if we allow each user to rate another only once, another way to falsely increase one's reputation would be to create fake identities and have each one of those rate the user's real identity with perfect scores. A good reputation system would avoid both of these problems. A possible solution to these problems would be to make sure that the ratings given by users with an established high reputation in the system are weighted more than the ratings given by beginners or users with low reputations. In addition, the reputation values of the users should not be allowed to increase at infinitum as in eBay. In this system, a seller may cheat 20% of the time but he/she can still maintain a monotonically increasing reputation value with a rate of 60%. In multiagent systems, where the agents' goal is the utility maximization of their human owners, the reputation values have to represent the expected probability for the completion of the delegated tasks. Reputation mechanisms have to be able to quantify the subjective [3] expectations of the owners of the agents, based on their past activity on the system. Finally, we have to consider the effect of the memory of our system [15]. The larger the number of ratings used in the evaluation of reputation values the higher the long-term prediction accuracy of the mechanism. However, since the reputation values are associated with humans and humans change their behavior over time it is desirable to disregard very old ratings. Thus, we ensure that the predicted reputation values are closer to the current behavior of the individuals rather than their overall performance.
4. Sporas: A reputation mechanism for loosely connected online communities Sporas provides a reputation service based on the following principles: 1.
New users start with a minimum reputation value, and they build up reputation during their activity on the system.
2.
The reputation value of a user does not fall below the reputation of a new user no matter how unreliable the user is.
3.
After each rating, the reputation value of the user i s updated based on the feedback provided by the other party to reflect his/her trustworthiness in the latest transaction.
4.
Two users may rate each other only once. If two users happen to interact more than once, the system keeps the most recently submitted rating. That way we avoid artificially inflated reputations through a two-party collusion. It is still possible to achieve a multi-party collusion which is faced by Histos, described in the next section.
5.
Users with very high reputation values experience smaller rating changes after each update. This approach is similar to the method used in the Elo [5] and the Glicko [11] system for pairwise ratings.
Each user has one reputation value, which is updated as follows:
1 t  F(Ri )• Riother +1 • (Wi +1 - E (Wi +1 )) q 1 1 R F (R ) = 1 and E (Wt +1 ) = t -(R - D ) Rt +1 =
1+ e
D
s
Equation 1 Sporas formulae Where, t is the number of ratings the user has received so far, q is a constant integer greater than 1, Wi represents the rating given by the user i, other
R
is the reputation value of the user giving the rating
D is the range of the reputation values, s is the acceleration factor of the damping function F. The smaller the value of s, the steeper the damping factor F(R).
As we can see from Equation 1, the change in the reputation value of the user receiving a rating of Wi from user Ri other, i s proportional to the reputation value Ri other of the rater him/herself. Thus if a user creates multiple new identities and has them rate him/herself the effect is not as significant as when he/she is rated by an established trustworthy user. Suppose for example that the minimum reputation allowed i n the system is 1/20th of the maximum reputation. Then if a user tries to fake his/her reputation through perfect ratings from newly created fake identities, he/she would need around 20 such identities to balance each rating submitted by a perfectly trustworthy person. The damping function F(R), ensures that the reputations of trustworthy persons are more robust against temporary malicious attacks. The temporal effect of the damping function should not be too long. Otherwise it would give an incentive to users to be trustworthy at the beginning of their activity on the system so that they can abuse the protection of the damping function later on. The expected rating of a user is his/her current reputation value over the maximum reputation value allowed in the system. Thus if the submitted rating is less than the expected one the rated user loses some of his/her reputation value. The value of q determines how fast the reputation value of the user changes after each rating. The larger the value of q, the longer the memory of the system. Thus, just like credit card history schemes [7], even if a user enters the system being unreliable in the beginning, if he/she improves later, his/her reputation value will not suffer forever from the early poor behavior. Similarly if an established trustworthy person starts behaving unreliably he/she should not be able to abuse his/her past high ratings for too long.
5. Histos: A reputation mechanism for highly connected online communities Sporas provides a global reputation value for each member of the online community, which is associated with them as part of their identity. Besides the online agent mediated interaction, our users will eventually have to meet each other physically i n order to commit the agreed transaction, or they may even know each other through other social relationships. The existing social relationships as well as the actual physical transaction process create personalized biases on the trust relationships between those users. The PGP Web of Trust [10] uses the idea that as social beings we tend to trust someone trusted b y someone we trust more than we trust a total stranger. Figure 1 Change of reputation for 10 different users over 100 ratings with q=10 New users start with reputation equal to 0 and can advance u p to the maximum of 3000. The reputation ratings vary from 0.1 for terrible to 1 for perfect. Since the reputation of a user in the community is the weighted average of non-negative values, i t is guaranteed that no user can ever have reputation value lower than that of a beginner. Also the weighed average schema guarantees that no user exceeds the maximum reputation value of 3000. If a user has a persistent real reputation value, the iteration of Equation 1 over a large number of ratings will give an estimate very close to that value [Figure 1].
Following this approach, we built a more personalized system. In Weaving a Web of Trust [12], what matters is that there is a connected path of PGP signed webpages between two users. In our case, we consider the different reputation ratings connecting the users of our system. We can represent the pairwise ratings in the system as a directed graph, where nodes represent users and weighted edges represent the most recent reputation rating given by one user to another, with direction pointing towards the rated user. If there exists a connected path between two users, say from A to AL, we can compute the more personalized reputation value of AL for A. When the user A submits a query for the Histos reputation value of a user AL we perform the following computation:
a)
b)
The system uses a Breadth-First-Search-like algorithm to find all directed paths connecting A to AL that are of length less than or equal to N. As described above we only care about the chronologically q most recent ratings given to each user. Therefore, if we find more than q connected paths taking us to user AL , we are interested only in the most recent q paths with respect to the chronological order of the rating events represented b y the last edge of the path. We can evaluate the personalized reputation value of AL if we know all the personalized reputation ratings of the users at the last node of the path before AL . Thus, we create a recursive step with at most q paths with length at most N-1.
If the length of the path is only 1, it means that the particular user, say C, was rated by A directly. The direct rating given t o user C is used as the personalized reputation value of C for A. Thus, the recursion terminates at the base case of length 1. To calculate the personalized reputation values, we use a slightly modified version of the reputation function described above. For each user AL , with m connected paths coming towards AL from A, we calculate the reputation of AL as follows: t 1 t Rt +1 = Â F (Ri +1 )• Riother • Wi +1 Â Ri +1 +1 q ' t -q ' t -q ' q ' = min (q , m ) and m = deg( AL )
(
)
agent mediated marketplaces where user anonymity is allowed. We believe incorporating reputation mechanisms in online communities may induce social changes in the way users participate in the community. Further evaluation is required to measure the effects of the proposed mechanisms.
8. References [1] Chavez, A, and Maes, P. Kasbah: An Agent Marketplace for Buying and Selling Goods. Proceedings of the First International Conference on the Practical Application of Intelligent Agents and Multi-Agent Technology (PAAM'96). London, UK, April 1996 [2]
Better Business Bureau. http://www.bbb.org
[ 3 ] Castelfranchi, C., Falcone, R., Principles of Trust for MAS: Cognitive Anatomy, Social Importance, and Quantification, Workshop in Deception, Fraud and Trust in Agent Societies, Second International Conference on Autonomous Agents (Agents '98), Minneapolis, St. Paul, May 9-13, 1998 [4]
eBay http://www.ebay.com
[5] Elo, A. E., The Rating of Chessplayers, Past and Present, Arco Publishing, Inc. (New York) 1978. [ 6 ] Donath, J., Identity and Deception in the Virtual Community, Communities in Cyberspace, Kollock, P. and Smith M. (eds). London: Routledge, 1998. [7]
Fair Isaak Co. http://www.fairisaac.com
[8] Foner, L., Yenta: A Multi-Agent, Referral Based Matchmaking System, First International Conference on Autonomous Agents (Agents '97), Marina del Rey, California, February 1997. [ 9 ] Friedman, E., and Resnick, P., The Social Cost of Cheap Pseudonyms: Fostering Cooperation on the Internet, Proceedings of the 1998 Telecommunications Policy Research Conference.
Equation 2 Histos formulae
[10] Garfinkel, S. PGP: Pretty Good Privacy, O'Reilly and Associates, 1994.
Where deg (AL ) is the number of connected paths from A to AL with length less than or equal to the current value of L. In the base case where L=1, since we have a connected path it means that A has rated AL him/herself, the personalized value for AL is naturally the rating given by A.
[11] Glickman, M. E., Paired Comparison Models with Time-Varying Parameters, PhD Thesis, Harvard University, May 1993.
In order to be able to apply the Histos mechanism we need a highly connected graph. If there does not exist a path from A to AL with length less than or equal to N, we fall back to the simplified Sporas reputation mechanism.
[13] Kollock, P., The Production of Trust in Online Markets, Advances in Group Processes (Vol. 16), edited by E. J. Lawler, M. Macy, S. Thyne, and H. A. Walker. Greenwich, CT: JAI Press. 1999.
6. Marketplace for Services
[15] Marsh, S. P., Formalising Trust as a Computational Concept, PhD Thesis, University of Stirling, April 1994.
We are applying the mechanisms described above in the context of a service marketplace. The experiment conducted i n this domain assumes that users with different areas of expertise enter the marketplace at the same time and start off with minimum reputation values. Since we allow user anonymity, the marketplace incurs the inefficiency that all users underprice themselves until their reputation values reflect their real competency and reliability. We are running simulations in order to investigate the microeconomic efficiency of these reputation mechanisms in conjunction with pricing strategies that adapt to variations in users' reputations.
7. Conclusion Collaborative filtering methods have been around for some years now, but they have focused on content rating and selection. We have developed two collaborative reputation mechanisms that establish reputation ratings for the users themselves, and presented two systems that can be used i n
[12] Khare, R., and Rifkin A., Weaving a Web of Trust, summer 1997 issue of the World Wide Web Journal (Volume 2, Number 3, Pages 77-112).
[14] MarketMaker. http://maker.media.mit.edu
[16] OnSale Exchange. http://www.onsale.com/exchange.htm [17] Reagle, J. M. Jr., Trust in a Cryptographic Economy and Digital Security Deposits: Protocols and Policies, Master Thesis, Massachusetts Institute of Technology, May 1996. [18] Resnick, P., Iacovou, N., Suchak, M., Bergstrom, P., Riedl J., GroupLens: An Open Architecture for Collaborative Filtering of Netnews, from Proceedings of ACM 1994 Conference on Computer Supported Cooperative Work, Chapel Hill, NC: Pages 175-186 [ 1 9 ] Shardanand, U., and Maes, P., Social Information Filtering: Algorithms for Automating 'Word of Mouth’, Proceedings of the CHI-95 Conference, Denver, CO, ACM Press, May 1995. [20] Zacharia, G., Moukas, A., and Maes, P., Collaborative Reputation Mechanisms in Electronic Marketplaces, Proceedings of the 33r d Hawaii International Conference of System Sciences, IEEE Press, January 1999.
