Robust Collaborative Filtering Algorithm Based On Attack Users Identifing And Bayesian Probabilistic Matrix Factorization

With the development of big data and e-commerce, collaborative filtering recommendation system is gradually permeating people’s daily life. However, due to the openness of recommender systems, malicious users can manipulate the recommendation results by injecting a number of fake profiles. This behavior has seriously affected the safety of recommendation system. In this paper, we propose a robust collaborative filtering method based on attack users identifying and Bayesian probabilistic matrix factorization. The specific content is shown as below.Firstly, for the problem of exiting attack users in data set, because malicious users can manipulate the recommendation results by injecting a number of fake profiles, we propose a clustering approach of suspicion users to reduce the influence of shilling attacks. For this purpose, we define the distance between users based on average rating popularity degree of users, and be used to cluster the attack users.Secondly, for the problem of false alarm ratio in the process of suspicion users clustering, we propose a attack users identifying approach based on suspicion users clustering and attacked item identifying to obtain the recommendation accuracy. We obtained the set of suspicion users based on clustering algorithm, however, the set of suspicion users contains some genuine users, so wo need to accurately identify and mark attack users in order to obtain the recommendation accuracy.Thirdly, for the preblem of low robustness of recommender systems, we devise a robust collaborative filtering recommendation algorithm by incorporating the user rating matrix and the indicator matrix of attack users into Bayesian probabilistic matrix factorization model to improve the robustness of recommender systems. In this algorithm, we excluded the attack users rated attacked item in the process of recommendation.Finally, to valisate the proposed algorithm, we simulate the experiment on Movie Lens 100 K dataset, and select three algorithm to compare with the proposed algorithm. The result show that the proposed algorithm outperforms the existing methods in terms of both recommendation accuracy and robustness.