Unsupervised and supervised methods for the detection of hurriedly created profiles in recommender systems

摘要

Recommender systems try to provide users with accurate personalized suggestions for items based on an analysis of previous user decisions and the decisions made by other users. These systems suffer from profile injection attacks, where malicious profiles are generated in order to promote or demote a particular item introducing abnormal ratings. The problem of automatic detection of such malicious profiles has been recently addressed by a great number of authors in the literature using supervised and unsupervised approaches. In this paper, we propose a framework to identify anomalous rating profiles, where each attacker (outlier) hurriedly creates profiles that inject into the system an unspecified combination of random ratings and specific ratings, without any prior knowledge of the existing ratings. This attack is a superset of the two different attacks (Uniform and Delta) proposed in Harper et al. (ACM Trans Interact Intell Syst 5(4):19, 2016) making the attack model more realistic and its detection more challenging. The proposed detection method is based on several attributes related to the unpredictable behavior of the outliers in a validation set, on the user-item rating matrix, on the similarity between users and on the filler items. In this work, we propose a new attribute (RIS) to capture the randomness in item selection of the abnormal profiles. In this work, three different systems are proposed: (1) a probabilistic framework that estimates the probability of a user to be an outlier by combining several features in a completely unsupervised way. (2) An unsupervised clustering system based on the k-means algorithm that automatically spots the spurious profiles. (3) A supervised framework that uses a random forest classifier for cases where labeling sample data is available. Experimental results on the MovieLens and the Small Netflix datasets demonstrate the high performance of the proposed methods as well as the discrimination accuracy of the proposed features.