Dual Clustering Based Collaborative Filtering for Robust Recommendation

Collaborative filtering has emerged as an effective solution for handling the issue of excessive information in the past several years. However, shilling attacks, where attackers introduce fake ratings to manipulate recommender systems, have proven to be effective against collaborative filtering algorithms. This paper proposes a collaborative filtering method that miti-gates the impact of attacks while maintaining prediction accuracy through dual clustering on the dataset. Dual clustering divides all data into some clusters by the first clustering, and integrates the clusters to form less clusters by the second clustering. The intention of dual clustering aims to separate actual users and attack users into different clusters by the first clustering and increase the data volume of actual users within one cluster by the second clustering. The effectiveness of dual clustering is experimentally explored using a large dataset of movie ratings. The prediction accuracy is assessed through an evaluation approach that measures the error value between actual ratings and predicted ratings. Moreover, the robustness against attacks is evaluated by comparing the shift of prediction value and the shift of TopN value before and after attacks. The experimental results show that dual clustering can reduce the impact of the drawbacks in collaborative filtering.