Research On Trustworthy Recommendation Algorithm Based On Shilling Attack Detection And Matrix Factorization Model

Collaborative filtering recommendation algorithms are widely used in personalized recommendation systems. Because the collaborative recommendation systems rely on users’ ratings information, malicious users inject fake rating information into the system,making recommender systems generate recommendation results to their own benefit. The existence of shilling attack undermines the credibility of recommender systems. Therefore,how to ensure the robustness of collaborative recommender systems has become an important issue in the research field of recommender systems. This paper proposes a trustworthy collaborative recommendation algorithm based on shilling attack detection and matrix factorization model.Firstly, we propose corresponding unsupervised attack detection algorithms for standard attacks(Random Attack, Average Attack and Bandwagon Attack), PIA(Power Items Attack) and Ao P(Average over Popular Items) attack. We devise standard attack detection algorithm through improving the detection algorithm based on Principal Component Analysis. We propose improved weighted deviation degree of mean rating,deviation of item’s average rating, percentage of item’s maximum rating, and devise Ao P attack detection algorithm based on these statistical characteristics. We propose improved degree of average similarity with nearest neighbors, and devise PIA attack detection algorithm.Secondly, we propose trustworthy recommendation algorithm based on shilling attack detection and matrix factorization model. First of all, the type of shilling attack is identified based on statistical characteristics of attack profiles. Then the suspicious users and items can be flagged by calling corresponding detection algorithm. At last, we combine the proposed attack detection algorithm with improved matrix factorization model, the model parameters can be computed using stochastic gradient descent algorithm,and then predicted ratings can be calculated. The matrix factorization model combines with users’ implicit trust information, these implicit trust relationships data can be generated using trust metrics model.Finally, we devise corresponding experimental program, and conduct contrast experiments on comparing the proposed algorithm with the existing algorithms.Experimental results show that the proposed algorithm in this dissertation exhibits good recommendation precision and excellent robustness.