Meta-Path and Matrix Factorization Based Shilling Detection for Collaborate Filtering

Nowadays, collaborative filtering methods have been widely applied to E-commerce platforms. However, due to its openness, a large number of spammers attack those systems to manipulate the recommendation results to earn huge profits. The shilling attack has become a major threat to collaborative filtering systems. Therefore, effectively detecting shilling attacks is a crucial task. Most existing detection methods based on statistical-based features or unsupervised methods rely on a priori knowledge about attack size. Besides, the majority of work focuses on rating attack and ignore the relation attack. In this paper, motivated by the success of heterogeneous information network and oriented towards the hybrid attack, we propose an approach DMD to detect shilling attack based on meta-path and matrix factorization. At first, we concatenate the user-item bipartite network and user-user relation network as a whole. Next, we design several meta-paths to guide the random walk to product node sequences and utilize the skip-gram model to generate user embeddings. Meanwhile, users’ latent factors are decomposed by matrix factorization. Finally, we incorporate these embeddings and factors to joint train the detector. Extensive experimental analysis on two public datasets demonstrate the superiority of the proposed method and show the effectiveness of different attack strategies and various attack sizes.