FedCDR: Privacy-preserving federated cross-domain recommendation

Abstract

Cross-Domain Recommendation (CDR) aims to solve data sparsity and cold-start problems by utilizing a relatively information-rich source domain to improve the recommendation performance of the data-sparse target domain. However, most existing approaches rely on the assumption of centralized storage of user data, which undoubtedly poses a significant risk of user privacy leakage because user data are highly privacy-sensitive. To this end, we propose a privacy-preserving Federated framework for Cross-Domain Recommendation, called FedCDR. In our method, to avoid leakage of user privacy, a general recommendation model is trained on each user's personal device to obtain embeddings of users and items, and each client uploads weights to the central server. The central server then aggregates the weights and distributes them to each client for updating. Furthermore, because the weights implicitly contain private information about the user, local differential privacy is adopted for the gradients before uploading them to the server for better protection of user privacy. To distill the relationship of user embedding between two domains, an embedding transformation mechanism is used on the server side to learn the cross-domain embedding transformation model. Extensive experiments on real-world datasets demonstrate that our method achieves performance comparable with that of existing data-centralized methods and effectively protects user privacy.