Privacy Preserving Recommendations for Social Networks

Abstract

Social recommendation is an advanced service of social networking platforms that is provided to their users. Social recommendation uses profiles and connections to generate personalized suggestions of contents, advertisements, people, pages, or interest groups. Since individual sensitive information is possibly involved in elaborating a recommendation, it may be inferred by an adversary in some situations. In this work, we design a differentially private setting to prevent social recommendations from disclosing sensitive information. Our recommendation system targets users of online social networks by leveraging their attributes and relationships. Unlike other approaches, we rely on both profile similarity and homophily properties. Therefore, our system estimates the frequency of friends who share some attribute values and applies non-negative matrix factorization to derive recommendations such as hobbies, movies, etc. We demonstrate the effectiveness of the proposed approach through experiments on real-world datasets and evaluation according to utility measures.