Recommender Systems: A Study of Cold-Start and Attack Resilience

Abstract

My thesis focuses on two key challenges in Recommender System: the Cold Start Problem and Data Poisoning attacks within the user-clustering framework. We explored utilizing user clustering to address the Cold Start Problem, analyzed its impact on data poisoning attacks, and devised a detection method for robust recommendation systems.

By addressing the Cold Start issue in Chapter 3, we revisited the challenge of making recommendations for new users due to the absence of any historical preference data. We proposed a novel Cluster-based Bandit (CB) algorithm that achieves fast learning in cold-start users. CB suggests that for fast learning, we want to initially ask the user to rate those items for which the information to distinguish between a group pair is the largest. Identifying such items, called distinguisher items, can quickly identify the correct cluster a user belongs to. Once the correct cluster is identified, smart recommendations for new users can be utilized by the collective wisdom of comparable users within the cluster. We demonstrate that even without extensive user data, our method improves the first user experience and offers good recommendations.

In Chapter 4, we examined the security ramifications of user clustering with regard to data poisoning threats. In order to modify recommendations and affect user behaviour, these malicious attacks entail the injection of skewed or false preference data. Our analysis showed that Matrix Factorisation (MF) could be vulnerable to targeted data poisoning attacks where attackers concentrate on particular user clusters. We specifically examined the process underlying how the user and item feature matrices U, V resulting from MF, alter following the injection of fake ratings. We demonstrated how these modifications aid in expanding the reach of targeted attacks by assessing the changes in these user and item feature matrices following the attack. Our findings lead us to the conclusion that it is the target item feature vector V that significantly contributes to altering the rating of the targeted item.

Finally, in Chapter 5, based on the conclusions from the last chapter, we presented an effective detection method for safeguarding user clusters in recommenders against data poisoning attacks. We presented a novel Item Vector Deviation (IVD)-based detection method that is based on the deviation in the item feature vector following injection of set of new ratings.