Robustness of Recommendation Algorithms against Co-Rating Attacks

Online stores like Amazon feature a high amount of different products, which can often be intimidating for new customers to find the right items. Modern information systems need to able to deal with these amounts of data, which is often done by using the interaction of users with products to calculate what other users might like or want to buy. One family of methods to calculate these suggestions uses the similarity of users, items, and their interactions. The calculation of this similarity is also based on the interactions, therefore, recommender systems can be manipulated by intentionally creating such user interactions, e.g., viewing items together, rating items together, or purchasing products at the same time, to purposefully raise or lower the relevance of items for a user. The efficiency reached and the effort required to perform such attacks is dependent on the recommendation algorithm. In this thesis, we aim to measure this efficiency and effort for different attack types and recommendation algorithms to investigate the robustness against such attacks.