RMPD: Method for Enhancing the Robustness of Recommendations with Attack Environments

Abstract

Personalized item recommendation has become a hot topic research among academic and industry community. But lots of purposeful fraudsters maybe perform different attacks on the recommender system to insert fake ratings, which could reduce the authenticity and reliability of recommendations. For a recommender system with fraudsters, it is crucial to detect malicious ratings and reduce the proportion of fraudster's ratings. This paper presents a method Prediction and Detection of Rating Matrix(RMPD) combining rating prediction and attack detection. The detection results of the attackers are applied to the rating prediction, thereby controlling the contribution and proportion of attackers to the rating prediction component both in training and learning, and then implementing more accurate item rating projections. The method will also solve the problem of data sparsity in the recommender system to some extent. The superiority of the proposed method in predicting recommendation performance compared with other baseline methods is demonstrated on real-world datasets. The ablation experiment proves the necessity of the components.