(A) robust recommendation system against review quality manipulation

Abstract

Recommendation systems influence decision making and have become attractive targets of manipulation. Collaborative filtering, widely adopted in recommendation systems, exploits the observed reviews given by users to provide personalized recommendations under the assumption that all users honestly rate items. Unfortunately, shilling attacks which inject fake reviews can easily manipulate the systems with the naive assumption. Several approaches have been proposed to mitigate the effect of shilling attack. Recently, some researchers have been interested in the fact that most recommendation systems encourage users to write reviews, as well as rate the helpfulness of reviews written by other users. With the assumption that users will evaluate the helpfulness of fake reviews as low, they proposed recommendation systems considering the helpfulness as the quality of reviews. However, those systems are vulnerable to attacks that inject fake helpfulness ratings to improve the quality of fake reviews. We propose a robust recommendation system to overcome such review quality manipulation attacks. The proposed approach estimates the true quality of reviews even in the presence of both injected fake reviews and helpfulness ratings. Experimental results on a real-world dataset demonstrate the robustness of our method. The proposed approach yields up to 20 times more robust recommendation results than the approaches that do not consider review quality manipulation.