Beyond Similarity: Personalized Federated Recommendation with Composite Aggregation

TL;DR

This work tackles personalized federated recommendation by addressing embedding skew inherent in one-to-one item embeddings. It introduces FedCA, a composite aggregation framework that jointly learns similarities among clients and data complementarity to update both trained and non-trained embeddings, implemented through a unified optimization and a server-side quadratic program. Empirical results across multiple real-world datasets demonstrate that FedCA outperforms strong baselines and maintains robustness under data sparsity, with ablations confirming the benefits of incorporating both similarity and complementarity. The approach is model-agnostic, privacy-preserving, and adaptable to existing FR architectures, offering a practical path toward more personalized, future-item accurate federated recommendations.

Abstract

Federated recommendation aims to collect global knowledge by aggregating local models from massive devices, to provide recommendations while ensuring privacy. Current methods mainly leverage aggregation functions invented by federated vision community to aggregate parameters from similar clients, e.g., clustering aggregation. Despite considerable performance, we argue that it is suboptimal to apply them to federated recommendation directly. This is mainly reflected in the disparate model architectures. Different from structured parameters like convolutional neural networks in federated vision, federated recommender models usually distinguish itself by employing one-to-one item embedding table. Such a discrepancy induces the challenging embedding skew issue, which continually updates the trained embeddings but ignores the non-trained ones during aggregation, thus failing to predict future items accurately. To this end, we propose a personalized Federated recommendation model with Composite Aggregation (FedCA), which not only aggregates similar clients to enhance trained embeddings, but also aggregates complementary clients to update non-trained embeddings. Besides, we formulate the overall learning process into a unified optimization algorithm to jointly learn the similarity and complementarity. Extensive experiments on several real-world datasets substantiate the effectiveness of our proposed model. The source codes are available at https://github.com/hongleizhang/FedCA.

Figures (15)

• Figure 1: Taking two clients as an ideal example, $\oplus$ denotes aggregation operators. Previous work can only update trained items repeatedly via similarity aggregation, while our composite aggregation method can both enhance trained items and update non-trained items.
• Figure 2: Empirical results regarding HR@10 and NDCG@10 on train and test sets, respectively.
• Figure 3: HR@10 results comparing our FedCA with FR baselines under varying train data ratios.
• Figure 4: NDCG@10 results comparing FedCA with FR baselines under varying train data ratios.
• Figure 5: Visualization results regarding similarity, complementarity, and composite aggregation weights.