Mobasher Bamshad, Robin Burke, JJ Sandvig
The open nature of collaborative recommender systems allows attackers who inject biased profile data to have a significant impact on the recommendations produced. Standard memory-based collaborative filtering algorithms, such as k-nearest neighbor, have been shown to be quite vulnerable to such attacks. In this paper, we examine the robustness of model-based recommendation algorithms in the face of profile injection attacks. In particular, we consider two recommendation algorithms, one based on k-means clustering and the other based on Probabilistic Latent Semantic Analysis (PLSA). These algorithms aggregate similar users into user segments that are compared to the profile of an active user to generate recommendations. Traditionally, model-based algorithms have been used to alleviate the scalability problems associated with memory-based recommender systems. We show, empirically, that these algorithms also offer significant improvements in stability and robustness over the standard k-nearest neighbor approach when attacked. Furthermore, our results show that, particularly, the PLSA-based approach can achieve comparable recommendation accuracy.
Subjects: 12. Machine Learning and Discovery