Detecting Anomalous Ratings Using Matrix Factorization For Recommender Systems

Personalization recommendation techniques play a key role in the popular E-commerce services such as Amazon, TripAdvisor and etc. In practice, collaborative filtering recommender systems are highly vulnerable to "shilling" attacks due to its openness. Although attack detection based on such attacks has been extensively researched during the past decade, the studies on these issues have not reached an end. They either extract extra features from user profiles or directly calculate similarity between users to capture concerned attackers. In this paper, we propose a novel detection technique to bypass these hard problems, which combines max-margin matrix factorization with Bayesian non-parametrics and outlier detection. Firstly, mean prediction errors for users and items are calculated by utilizing trained prediction model on test sets. And then we continue to comprehensively analyze the distribution of mean prediction errors of items in order to reduce the scope of concerned items. Based on the suspected items, all anomalous users can be finally determined by analyzing the distribution of mean prediction error on each user. Extensive experiments on the MovieLens-100K dataset demonstrate the effectiveness of the proposed method.