Improving computational trust representation based on Internet auction traces

摘要

Computational trust representations are used by Trust Management (TM) systems to elicit information from users about the behavior of others. In most practically used TM systems, simple computational trust representations dominate, such as the three-valued discrete scale of “negative”, “neutral” and “positive” used in reputation systems of Internet auctions. This paper asks the question: what is the appropriate system for computational representation of human trust? In order to find an answer, we study a large trace of feedbacks and textual comments from a reputation system of an Internet auction. We discover that users systematically try to add information in the textual comments. Text-mining and NLP approaches reveal a taxonomy of non-positive feedbacks and an importance order on the categories of non-positive behavior. This importance order is further supported by survey data. Based on these observations, we propose and evaluate a complete, new computational trust representation system inspired by the work of Yager. This system is complemented by operators that can be used to produce rankings of most trusted agents. The operator used to create rankings selects Pareto-optimal agents with respect to the multiple criteria revealed by our trace analysis. The proposed system takes into account all criteria utilized by auction users to evaluate behavior, and the relative importance of these criteria. The proposed system is compared to the Detailed Seller Rating system introduced by eBay.