Privacy-enhanced matrix factorization with shuffle differential privacy

Recommender systems collect and analyze user data to enhance the user experience. Nevertheless, when a recommender system has access to users' item sets or their ratings, several privacy concerns are raised. A variety of solutions have been proposed to improve the privacy of traditional recommender systems. However, the existing solutions in the literature, which are based on local differential privacy (LDP), suffer from inherently low performance. Recently, shuffle differential privacy (SDP), as an emerging extension of LDP, has been proven to have a privacy amplification effect due to its data anonymity. This shows potential for improving the trade-off between privacy protection and data utility. In this study, we establish a strong privacy guarantee for matrix factorization algorithms under shuffle differential privacy. Experimental results on two real-world datasets demonstrate that the proposed guarantee can enhance the performance of privacy-protected matrix factorization algorithms.