Robust Reputation Management Using Probabilistic Message Passing

In a typical reputation management system, after each transaction, the buyer (who receives a service or purchases a product) provides its report/rating about the quality of the seller for that transaction. In such a system, the problem of reputation management is to compute two sets of variables: 1. the (global) reputation parameters of entities who act as sellers, and 2. the trustworthiness parameters of the entities who act as the raters (i.e., buyers). In this paper, for the first time, we introduce an iterative probabilistic method for reputation management. The proposed scheme, referred to as RPM, relies on a probabilistic message passing algorithm in the graph-based representation of the reputation management problem on an appropriately chosen factor graph. In the graph representation of the problem, the sellers and buyers are arranged as two sets of variable and factor nodes, respectively, that are connected via some edges. Then, the reputation and trustworthiness parameters are computed by a fully iterative and probabilistic message passing algorithm between these nodes in the graph. We provide a detailed evaluation of RPM via computer simulations. We observe that RPM iteratively reduces the error in the reputation estimates of the sellers due to the malicious raters. Finally, comparison of RPM with some well- known and commonly used reputation management techniques (e.g., Averaging Scheme, Bayesian Approach and Cluster Filtering) indicates the superiority of the proposed scheme both in terms of robustness against attacks (e.g., ballot-stuffing, bad-mouthing) and computational efficiency.