Multi-Agent Modeling of Risk-Aware and Privacy-Preserving Recommender Systems

Abstract

Recent progress in the field of recommender systems has led to increases in the accuracy and significant improvements in the personalization of recommendations. These results are being achieved in general by gathering more user data and generating relevant insights from it. However, user privacy concerns are often underestimated and recommendation risks are not usually addressed. In fact, many users are not sufficiently aware of what data is collected about them and how the data is collected (e.g., whether third parties are collecting and selling their personal information). Research in the area of recommender systems should strive towards not only achieving high accuracy of the generated recommendations but also protecting the user’s privacy and making recommender systems aware of the user’s context, which involves the user’s intentions and the user’s current situation. Through research it has been established that a tradeoff is required between the accuracy, the privacy and the risks in a recommender system and that it is highly unlikely to have recommender systems completely satisfying all the context-aware and privacy-preserving requirements. Nonetheless, a significant attempt can be made to describe a novel modeling approach that supports designing a recommender system encompassing some of these previously mentioned requirements. This thesis focuses on a multi-agent based system model of recommender systems by introducing both privacy and risk-related abstractions into traditional recommender systems and breaking down the system into three different subsystems. Such a description of the system will be able to represent a subset of recommender systems which can be classified as both risk-aware and privacy-preserving. The applicability of the approach is illustrated by a case study involving a job recommender system in which the general design model is instantiated to represent the required domain-specific abstractions.