以期望值最大化於仲介式口碑系統之信任學習

Abstract

A reputation system predicts a user’s reputation in a way similar to the word-ofmouth in the real world. Each user sends feedbacks to the system, and the system learns a trust model predicting each user's reputation. The prediction builds up trust relationship between each pair of users and it can reduce a user's losses in a transaction. Our system learns user trust by using Expectation-Maximization algorithm (EM algorithm). EM algorithm can learn the unobservable trust of a user from observable feedbacks sent by users, with the probabilistic model describing the relationship between the known and unknown. The model assumes the existence of a buyer's rating bias which is reflected in a buyer's feedbacks in order to better predict a user's reputation, especially when there are few feedbacks available. Our reputation system predicts both user's reputation and rating bias in a broker-based architecture. EM learning is done inside each broker who only receives feedbacks from its own group of users. Inter-broker communication can reduce the errors brought by the seperation of user feedbacks, while the broker-based architecture keeps the system scalable and avoids drawbacks of a centralized system. EigenTrust is resilience to various attacks in a P2P environment, and we use it to manage our inter-broker communication where the inter-broker relation is in a P2P fashion. We implement a simulator to verify our model, and the experiment result shows that our system can predict better than the simple averaging method. Our system is also less sensitive to the change of feedback types and the increase of users. Therefore, our model can accurately learn a user’s trust in a broker-based system.