Robust Reputations For Peer-To-Peer Marketplaces

We have developed a suite of algorithms to address two problems confronting reputation systems for large peer-to-peer markets: data sparseness and inaccurate feedback. To mitigate the effect of inaccurate feedback - particularly retaliatory negative feedback - we propose EM-trust, which uses a latent variable statistical model of the feedback process. To handle sparse data, we propose Bayesian versions of both EM-trust and the well-known Percent Positive Feedback system. Using a marketplace simulator, we demonstrate that these algorithms provide more accurate reputations than standard Percent Positive Feedback.