Shilling Attack Prevention for Recommender Systems Using Social-based Clustering

Abstract

A Recommender System (RS) is a system that utilizes user and item information to predict the feeling of users towards unfamiliar items. Recommender Systems have become popular tools for online stores due to their usefulness in confidently recommending items to users. A popular algorithm for recommender system is Collaborative Filtering (CF). CF uses other users' profiles to predict whether a user is interested in a particular object. This system, however, is vulnerable to malicious users seeking to promote items by manipulating rating predictions with fake user profiles. Profiles with behaviors similar to "victim" users alter the prediction of a Recommender System. Manipulating rating predictions through injected profiles is referred to as a shilling attack. It is important to develop shilling attack prevention frameworks for to protect the trustworthiness of Recommender Systems. In this thesis, we will demonstrate a new methodology that utilizes social information to prevent malicious users from manipulating the prediction system. The key element in our new methodology rests upon the concept of trust among real users, an element we claim absent among malicious profiles. In order to use trust information for shilling attack prevention, we first develop a weighting system which makes the system rely more on trustworthy users when making predictions. We then use this trust information to cluster out untrustworthy users to improve rating robustness. The robustness of the new and classic systems is then evaluated with data from a public commercial consumer RS, Epinions.com. Several complexity reduction procedures are also introduced to make implementing the algorithms mentioned possible for a huge commercial database.