Knowledge and Information Systems Human-Computer Interaction

GP-FedRS: trainable global prototype learning for federated recommendation

Bao, Jiayu, Zhang, Rui, Fan, Jiansong, Liu, Yuan
Springer Nature · 2026-11-15 · DOI 10.1007/s10115-026-02882-3

Abstract

Federated recommendation suffers from representation shift under heterogeneous client data, and existing prototype-based methods commonly construct global representations by directly averaging local prototypes, which can shrink inter-prototype margins and limit semantic diversity. We propose GP-FedRS, a federated recommendation framework that learns trainable global prototypes on the server, introduces dynamic-margin contrastive learning to enhance inter-prototype separability, and employs discrepancy-aware client selection to improve semantic coverage during training. We evaluate GP-FedRS against eight strong baselines on four public benchmark datasets containing 943–8072 users, 1682–12,454 items, and 63,836–1,000,209 interactions. Across HR and NDCG metrics, GP-FedRS consistently achieves the best performance, with relative improvements of 4.82–11.36% over the strongest baseline. These results demonstrate that explicitly optimizing trainable global prototypes is an effective way to improve robustness and ranking quality in heterogeneous federated recommendation.

Read at publisher →