KC-GCN: A Semi-Supervised Detection Model against Various Group Shilling Attacks in Recommender Systems

Abstract

Various detection methods have been proposed for defense against group shilling attacks in recommender systems; however, these methods cannot effectively detect attack groups generated based on adversarial attacks (e.g., GOAT) or mixed attack groups. In this study, we propose a two-stage method, called KC-GCN, which is based on $k$ -cliques and graph convolutional networks. First, we construct a user relationship graph, generate suspicious candidate groups, and extract influential users by calculating the user nearest-neighbor similarity. We construct the user relationship graph by calculating the edge weight between any two users through analyzing their similarity over suspicious time intervals on each item. Second, we combine the extracted user initial embeddings and the structural features hidden in the user relationship graph to detect attackers. On the Netflix and sampled Amazon datasets, the detection results of KC-GCN surpass those of the state-of-the-art methods under different types of group shilling attacks. The F1-measure of KC-GCN can reach above 93% and 87% on these two datasets, respectively.