Strategic Attacks on Recommender Systems: An Obfuscation Scenario

International audienceUnderstanding user behavior in the context of recommender systems remains challenging for researchers and practitioners. Inconsistent and misleading user information, which is often concealed in datasets, can inevitably shape the recommendation results in certain distorted ways despite utilizing recommender models with enhanced personalizing capabilities. Naturally, the quality of data that fuels those recommenders should be extremely reliable and free of any biases that might be invisible to a model, irrespective of its type. In this article, we introduce two modern forms of noise that are intrinsically hard to detect and eliminate; one is malicious in nature and will be termed Burst while the other is unique in that it forms its own category and will be referred to as Opt-out. Additionally, with the aim of segregating the nature of noise behind such threats, we present a distinct case study on Burst and Opt-out to illustrate how the detection of those threats can be challenging compared to that of traditional noise and with the current detection methods. Finally, we expound on the ability of such threats to bias the output of recommenders in their own unique way while primarily retaining data that is not fundamentally erroneous