Research On Identification Method Of The Fake Reviewers In Recommender Systems

Recommender system is also called personalized recommender system.The main recommendation algorithms include collaborative filtering recommendation,content-based recommendation and combined recommendation.The openness and interaction characteristics of collaborative filtering recommender system not only ensure the accuracy of recommendation results,but also make it vulnerable to external attacks.For the purpose of making illegal profits,the attackers try to change the recommendation result of the recommender system by injecting fake user profile information.Such attacks are called shilling attacks or profile injection attacks.The research shows that collaborative filtering recommendation algorithm is easily manipulated by bogus user profile,which leads to the decrease of users' satisfaction with the accuracy of recommendation results.The detection of shilling attacks has become a hot topic in the research field of recommender systems.This paper mainly verifies the influence of bogus user profiles on recommendation results,and studies the identification of fake users,that is,attack users in collaborative filtering recommender systems.First of all,this paper experimentally verified the influence of bogus attackers on user-based and item-based collaborative filtering recommendation results.Secondly,aiming at the problem that the existing classification attributes cannot differentiated between the shilling attackers and the normal users,this paper proposes Mean of User Popularity Degree(MUPD)classification feature to distinguish the real users from the false users effectively.Rating Deviation of Item(RDI)feature is proposed to effectively distinguish the common items and the attack items in the systems.Aiming at the low detection rate of existing supervised attack detection algorithms,this paper proposes discrimination index,and designs a shilling attack detection framework based on dynamic feature selection to reduce misjudgment.Through dynamic selection of classification attributes and further filtering and examination of the detection results,the framework can effectively detect the fake attackers in the recommender system.Finally,considering the natural noise,that is vandalism attack,and more complicated attack types in the recommender system,by introducing proximity operator and matrix completion,we design the L2,1-Norm regularized based matrix completion incorporating prior information model,and employ the bregman iteration based optimization algorithm to solve the above model,and thus propose the L2,1-Norm regularized matrix completion based shilling attack detection algorithms.In order to verify the effectiveness of the proposed two attack detection algorithms,this paper selectes data sets with different sparsity to test the algorithms and compared them with other attack detection algorithms.Experimental results show that dynamic feature selection based shilling attack detection algorithm proposed can effectively detect the common attack models,and the detection effect is obviously better than the contrastive algorithm.The L2,1-Norm regularized matrix completion based shilling attack detection algorithms proposed can not only detect all kinds of shilling attacks,but also detect the more complicated shilling attacks and the natural noise in the recommender systems.