Research On Group Attack Detection Method Based On Dense Subgraph Mining In Recommender Systems

Recommender systems can provide precise and personalized services for people,which are vulnerable to attacks from malicious users.Some organizations usually inject shilling profiles into systems to boost or suppress target items driven by the interests.Group shilling attacks are more harmful to the recommender systems than individual shilling attacks.How to efficiently detect group shilling attacks that cause serious damage in the field of recommender system security has become a problem that needs to be solved urgently.Domestic and foreign researchers have proposed detection approaches based on supervised and unsupervised learning to reduce the impact of the group attacks on the recommendation system.However,the supervised detection methods need to train the model through labeled training samples and can only detect specific attack models.Moreover,the result of the detection is strongly dependent on the choice of the labeled samples and the model.Although the unsupervised detection method does not require the labeled training samples,it requires sufficient prior knowledge.For the limitations of these detection methods,this paper studies group attack detection algorithm based on dense subgraph mining.Firstly,the existing detection methods based on graph have a problem,which is the structure of divided subgraph is sparse.This paper proposes a group attack detection method based on triangle dense subgraph mining.In this method,the user relationship graph by analyzing the users' rating of items in the database and extracting the rating times and rating information to establish the relationship between users.The improved triangle dense subgraph mining method is used to mine dense subgraphs in the user relationship graph,and the PPR seed expansion algorithm is used to find candidate shilling groups.Based on which,the attack groups are obtained by calculating the suspicious degrees of the candidate shilling groups.Secondly,another work related to group shilling attack is based on community discovery.The existing algorithms usually neglect the topological structure of community.In this paper proposes a group attack detection method based on clique percolation method.In this method,the user relationship graph is constructed by introducing suspicion of product to analyze the possibility of an item's popularity being attacked,then find the similar relationship between users.The clique percolation method finds dense structure of the 3-clique community.The detection indicators are extracted from the two perspectives of group behavior characteristics and group network structure characteristics,which are measure the group suspicious degree of each candidate group,then generate a list of attack groups.Finally,the proposed method in this paper is experimented on Amazon dataset and Yelp dataset,and the experimental results are compared with the existing group attack detection methods to verify the proposed method effectiveness.