Group Shilling Attack Detection Method Based On Multi-view Learning In Recommender Systems

In the recommendation system,user rating data can not only help users themselves to obtain high-quality services,but also provide valuable reference for user groups with similar interests in the system.An item with a higher rating in the system has a higher probability of being recommended to similar user groups.Based on their own interests,unscrupulous merchants employ gray industries to inject a large number of fake use profiles into the system in an organized and premeditated manner,in order to improve the recommendation probability of their own items or suppress competitors’ items.This behavior seriously affects the service quality of the system.In order to detect these fake profiles,domestic and foreign researchers have carried out a lot of research and achieved good detection results.However,with the continuous improvement of attack strategies,these detection methods have the problems of a single data source and insufficient feature information,which makes the detection performance vulnerable to the evolution of the attack model,and cannot effectively distinguish all the differences between real users and attack users.Aiming at these problems,this paper proposes two multi-view learning-based methods for detecting group shilling attacks.Firstly,in view of the problems of single data source and insufficient feature information in existing detection methods,this paper proposes a group shilling attack detection method based on extreme value theory and triple relationship.The detection method calculates the user-item relevance degree according to the item’s rating time series,uses the adjusted cosine similarity to construct a user relationship graph for each data source.The unified user relationship matrix and triple user relationship matrix are obtained by using the node adaptive multi-graph fusion method based on extreme value theory.The greedy algorithm is used to iteratively select the group with the largest group density to generate the candidate group.The suspicious degree of each group is calculated by the group density and the multi-source group item suspicious degree,the attack group is obtained by using the hierarchical clustering algorithm.Secondly,aiming at the defect that existing detection methods require manual feature engineering to identify attack groups,this paper proposes a group shilling attack detection method based on multi-view clustering.The detection method calculates the user-item correlation degree according to the item’s rating time series and the item’s rating level time series,uses the adjusted cosine similarity to construct the user correlation graph matrix of each data source.The multi-view fusion model designed to automatically obtain the number of abnormal groups is used to fuse the user correlation graph matrix extracted from different data sources to generate a unified user correlation graph matrix include the information of candidate groups and attack groups.The attack group is then obtained according to the group size.Finally,experiments on Epinions synthetic dataset and Amazon real dataset show that the detection method proposed in this paper outperforms the four existing detection methods both in single-source and multi-source detection.