FedRSFL: Federated recommender systems with feature-level user data sharing

Federated recommender systems have made great progress in recent years, and they can protect the privacy of user data while maintaining recommendation performance. However, most existing federated recommender systems require every user to keep all data strictly on the local device, and they overlook the users who care less about privacy and are willing to share data with the server for better recommendation. To address this, a few recent studies explore federated recommender systems with user data sharing, which allow users to share their data with the server. However, these methods mainly focus on interaction-level user-item data, which is often sparse, and they overlook the finer-grained feature-level data. To address this problem, this paper explores a federated recommender system framework with feature-level user data sharing, termed FedRSFL, which allows users to share part of their features with the server. However, training on the shared data is non-trivial: the same shared interaction produces two inconsistent training signals, because a client learns it from the complete local data with its private user embedding while the server learns it from the shared part alone with a separate user embedding, and directly aggregating the two models mixes unaligned parameters and hurts the global model. FedRSFL therefore couples the two sides through a bidirectional preference distillation, in which the clients teach the server the complete preference that the shared data misses, while the server teaches the clients the cross-user pattern that a single client cannot see, so that the shared data improves recommendation and every user embedding stays on its device. Experiments on three public datasets demonstrate that feature-level user data sharing improves the recommendation accuracy, with an average gain of about 3.8% in HR@10 over the strongest baseline on each dataset.