Robustness analysis of multi-criteria recommender systems against power user attacks

Recommendation systems offer new product predictions to users using their past liking feedback. The accuracy of predictions is tied to the system's understanding of user preferences. Multi-criteria recommendation systems further aim to provide more personalized recommendations by collecting user evaluations in more than one category. With the increasing commercial adoption of recommendation systems, there is a growing need to safeguard these systems. One of the common methods used in attacking recommendation systems is the shilling attack, which involves adding fake user profiles to the system to manipulate suggestions. In power user attack, where some users are considered to be able to influence large user groups, selected power users are copied, thus changing predictions. Effectiveness of those attack on traditional recommender systems have shown in studies. But robustness of multi-criteria recommender systems against power user attack has not yet been studied. Within scope of this thesis, it is discussed how to apply power user attack to multicriteria recommender systems and robustness of common recommender system algorithms against power user attack analyzed. In the experimental studies on real multicriteria data sets, difference between predictions before and after power user attack have been presented by common method in literature called prediction shift. Also, the robustness of common multi-criteria recommender system algorithms against those attacks has been analyzed with two new metrics. The success of attacks according to recommender system algorithms, power user detection methods, and sub-criteria has been analyzed using ANOVA and Kruskal-Wallis statistical methods.