Development of robust prediction models for recommender systems

Recommender systems play an important role in addressing the problem of information overload. A growing body of literature recognizes the importance of recommender systems. However, these systems are subject to shilling attacks that adversely affect the performance of recommender systems. Many works have been presented to detect these attacks and have achieved successful results. However, most of the works have presented offline methods. Therefore, there is a need for online and instantaneous attack detection methods to improve the robustness of these systems. In this thesis, we propose a novel online shilling intrusion detection to enable instantaneous intervention in recommender systems. Quality control charts are used to detect suspicious users. The detected suspicious users are checked with a density-based clustering algorithm. Correlation analysis is used to find user similarities in the checking process. Recall, precision and F1-scores are given to show the performance of the proposed method. The main contribution of this work is that shilling attacks are detected among suspicious users instead of the whole dataset. Furthermore, the attack time is used as an additional attribute in the detection process. To ensure the robustness of the recommender systems, a classification algorithm is trained with the detected attacks to facilitate the detection of similar attacks in the future. Here, oversampling is used to compensate for the number disparity between attack and real profiles. Again, in the absence of attack profiles, single class support vector machines were used for attack detection.