Personalized semi-decentralized federated recommender

Jiayu Bao, Yicheng Di, Song Shen, Rongsheng Hu, Yuan Liu

Abstract: Highlights•We propose a novel semi-decentralized federated recommender framework (pFedSG). The framework significantly improves scalability and privacy preservation, and can be applied to almost all GNN-based federated recommender models.•To protect privacy while utilizing high-order collaborative information, we connect isolated ego-graphs using predicted item nodes.•We propose a module called Fine-Grained Personalization (FGP). This module obtains necessary information from element-level global models and adaptively aggregates global and local models for local goals.•The efficacy of our recommender algorithm is thoroughly assessed using four authentic datasets, and the findings unequivocally indicate that our approach enhances the overall performance of recommendations.

External IDs:dblp:journals/ipm/BaoDSHL26