Personalized semi-decentralized federated recommender

Abstract

The recently proposed federated recommender system can alleviate privacy concerns; however, existing methods either rely on third-party servers to access other isolated graphs or restrict local training to isolated graphs. A key challenge in federated learning (FL) is statistical heterogeneity, which can undermine the generalization ability of the global model across clients. To address these issues, we propose a novel semi-decentralized federated recommender framework with adaptive local aggregation, named pFedSG. This framework improves scalability through device-to-device collaboration and enhances local subgraphs by connecting isolated graphs with predicted item-node connections, thereby preserving high-order user-item collaboration information. Furthermore, we introduce a fine-grained personalization (FGP) module, which adaptively aggregates the downloaded global model and local model for each client based on their local objectives, enabling effective learning of fine-grained personalization for users and items. To evaluate the effectiveness of the proposed pFedSG, we conducted extensive experiments on four public datasets. pFedSG significantly outperformed ten benchmark models. Specifically, compared to the best baseline, pFedSG improved HR and NDCG evaluation metrics by 7.37% and 6.51%, respectively. Additionally, pFedSG is applicable to existing graph neural network-based federated recommender methods. Further experiments also validate the superiority of pFedSG from multiple analytical perspectives.