A Social Reputation Model for Electronic Marketplaces Sensitive to Subjectivity, Deception and Change

This thesis examines the topic of designing electronic marketplaces populated by intelligent software agents acting on behalf of buyers and sellers. In particular, we explore the challenge of having buying agents learn rich representations of other agents providing information about sellers (known as advisors) in order to make effective decisions about which selling agents are best for purchasing goods. This constitutes a social model of reputation in electronic marketplaces. The interpretation of the information provided by advisors is complicated by the inherent subjectivity of each advisor, the possibility that advisors may deliberately provide misleading evaluations to deceive competitors and the dynamic nature of seller and advisor behaviours that may naturally change seller evaluations over time. We use the work of Tran and Cohen to develop the Regan-Tran-Cohen (RTC) social model of reputation in order to gain insight into how to best model sellers and advisors while addressing subjectivity, deception and change. This includes an approach for identifying reputable and disreputable advisors, as well as adjusting for straightforward subjective differences. The lessons learned from the RTC model are used to design a probabilistic model called BLADE (Bayesian Learning to Adapt to Deception in E-marketplaces). The BLADE model incorporates the concept of epistemic uncertainty to capture the buying agent’s confidence in its current model of the seller (and its model of advisors). Each seller is modeled in terms of a set of properties and each advisor is represented in terms of its evaluation function, mapping observed seller properties to a single rating. The beauty of BLADE is its ability to make use of information provided by advisors in order to learn about sellers, regardless of whether the advisor is subjectively different or deceptive. We compare the BLADE model to TRAVOS and BRS, two related probabilistic approaches for constructing a social model of reputation in multi-agent systems. We find that when seller ratings provided by advisors are inconsistent and there is little information contained in these ratings, this is reflected in uniformly poor performance by all three models. However, when a majority of advisors provide deceptive, yet consistent ratings, BLADE is able to outperform the other models, which tend to discount or disregard the information provided by advisors who are considered to be deceptive. We also demonstrate that the BLADE model is able to quickly adjust its model in order to cope with changing seller and advisor behaviour. We argue as well that BLADE provides a solid foundation for future research in two main directions: to incorporate mechanisms in electronic marketplaces in order to discourage deception and as part of a decision-theoretic framework for calculating policies dictating the optimal action for the buying agent.