Improve Predictive Accuracy by Identifying Collusions in P2P Recommender Systems

According to the malicious collusion behavior of nodes in P2P networks, a method of collusion identification is proposed based on the idea of clustering. This method considers the behavior characteristics of collusion, screens the rating nodes of the target node by three attributes of the rating extremes, rating time and historical similarity. According to the size of the suspected degree of collusion, some sets of malicious collusive nodes are selected. Experiments on a real data set show that the accuracy of recommendation has been significantly improved after excluding the identified collusive nodes, which proves the effectiveness of the method proposed in this paper.