User-Centric Federated Matrix Factorization Based on Differential Privacy

Matrix factorization is a popular recommendation method used in many fields, but it often lacks sufficient privacy protection. To address privacy concerns in connected living, we propose a user-centric federated matrix factorization model based on differential privacy (FMF-DP). First, we design a federated matrix factorization framework where each user is a participant and only uploads their gradients after completing the training task locally. Second, we introduce the user-average-rating bias into the matrix factorization method to improve accuracy. Finally, we use DP to enhance privacy protection without compromising computational efficiency. We have implemented the prototype of an FMF-DP and tested it on several datasets, showing promising results. Compared with the baseline model, the root-mean-square error is reduced by 0.25 on average, and the computational efficiency is also improved.