Robust Trust Management

Our research is within the area of artificial intelligence and multiagent systems. More specifically, we focus on the robustness issues [1] in reputation systems for electronic marketplaces and aim to address the following problems, • how to cope with dishonest advisors (buyers who provide misleading opinions), improving the robustness of the trust model. • how to choose an appropriate seller to perform transaction by querying advisors in an optimal way. To explain, in multi-agent based e-marketplaces, self-interested selling agents may act maliciously by not delivering products with the same quality as promised. It is thus important for buying agents to analyze their quality and determine which sellers to do business with, based on their previous experience with the sellers. However, realistically, in most e-marketplaces, buyers often encounter sellers with which they have no previous experience. In such cases, they can query other buyers (called advisors) about the sellers. But, advisors may act dishonestly by providing misleading opinions (unfair ratings) to promote low quality sellers or demote sellers with high quality. Hence, it is necessary to evaluate the quality of advisors' opinions to determine their reliability. While it is prima facie necessary to gather opinions about a seller, a buyer may not need to query all the advisors about the seller, since the cost of querying all the advisors may be greater than the value derived from a successful transaction with the seller. Thereby, it is necessary to design an optimal scheme to selectively query advisors and choose a quality seller to perform transaction, in order to maximize the utility of the buyer in the long run. 2. PROGRESS TO DATE Up to date we have proposed: 1) a biclustering based approach to identify dishonest advisors in a multi-criteria e-marketplace; 2) a POMDP based approach (called the SALE POMDP) to optimally select sellers in an e-marketplace. 2.1 The Biclustering Based Approach Existing trust models [2, 3] such as BRS, iCLUB, etc., which deal with the unfair rating problem are only designed to operate