Trust-aware Privacy-Preserving Recommender System

Recommender systems have achieved great success in providing product recommendations for online shopping. With recommender systems, customers can find their interested merchandise in a timely manner. It not only facilitates customers' purchases, but also promotes sales. While recommender systems can predict customers' preferences accurately, they suffer from privacy leakage in many aspects.

In this paper, we study an attack model in centralized recommender systems and present a privacy-preserving recommendation framework to neutralize such attack. In this model, an attacker holds some of the real customers' ratings and attempts to obtain their private preferences. The proposed framework is based on the weighted nonnegative matrix tri-factorization technique. It utilizes customers' trustworthiness to filter out unrelated ratings so that their privacy can be preserved. We performed experiments on two different datasets with respect to recommendation precision and privacy preservation level. The results demonstrate that our recommender system can distinguish between real customers and attackers to a great extent without compromising the prediction accuracy.