Privacy-Preserving Federated Cross-Domain Social Recommendation

By combining user feedback on items with social networks, cross-domain social recommendations provide users with more accurate recommendation results. However, traditional cross-domain social recommendations require holding both data of ratings and social networks, which is not easy to achieve for both information-oriented and social-oriented websites. To promote cross-domain social network collaboration among the institutions holding different data, this chapter proposes a federated cross-domain social recommendation (FCSR) algorithm. The main innovation is applying Random Response mechanism to achieve sparsely maintained differential privacy for user connections and proposing Matrix Confusion Method to achieve efficient encrypted user feature vector updates. Our experiments on three datasets show the practicality of FCSR in social recommendation and significantly outperforms baselines.