Research On Robust Collaborative Recommendation Based On Attack Identification

Due to some malicious users artificially provide large scale fake ratings to the system, trying to make the system product advantage recommendation results for themselves, leading to the quality of the system recommended are seriously affected. And attack users have been looking for new attack methods to damage the system. So, solving the security problems of recommended system facing to different attacks is urgently needed. This paper conducts research on the problem of the recommender accuracy and the ability rejecting to attack for recommender system when facing attack by the way of analysing the features between existed and new attack, proposes a robust collaborative recommendation algorithm based on identification.Firstly, due to the existing robust algorithms are poor to face shilling attacks, and some robust recommendation algorithms in order to improve the robustness by sacrificing the accuracy. To solve this problem, we propose robust recommendation algorithm based on shilling attack identification. We first calculate the feature of user rating behavior to get the set of suspicious users. Then we can find the attacked item by statistical method and calculate the suspicious degree of users’ behavior by the attacked item to detect attack users. Finally, we combine results with probabilistic matrix factorization to improve the accuracy and robustness of recommendation.Secondly, to solve the problem of the existed robust recommendation algorithms are general for the existed shilling attack which lead to they are poor to face the power item attack, we devise robust recommendation algorithm based on power item attack identification. we first analysis the feature of the power item attack users to detect the attack users using popular item density of users and makes a judgment of the users whether or not have ever gave the attacked item ratings, if it is true, this item is not participate in iteration update of matrix factorization which can eliminate the influences on the attack users to the attacked item, and improve the recommender accuracy to a certain extent.Finally, we give the experimental evaluations based on Movielens and Netflix database and analysis of the robust algorithms which based on attack identification proposed in this paper, compare the performance between the proposed algorithms and other existied algorithms.