A genre trust model for defending shilling attacks in recommender systems

Abstract

Shilling attacks have been a significant vulnerability of collaborative filtering (CF) recommender systems, and trust in CF recommender algorithms has been proven to be helpful for improving the accuracy of system recommendations. As a few studies have been devoted to trust in this area, we explore the benefits of using trust to resist shilling attacks. Rather than simply using user-generated trust values, we propose the genre trust degree, which differ in terms of the genres of items and take both trust value and user credibility into consideration. This paper introduces different types of shilling attack methods in an attempt to study the impact of users’ trust values and behavior features on defending against shilling attacks. Meanwhile, it improves the approach used to calculate user similarities to form a recommendation model based on genre trust degrees. The performance of the genre trust-based recommender system is evaluated on the Ciao dataset. Experimental results demonstrated the superior and comparable genre trust degrees recommended for defending against different types of shilling attacks.