Beyond Personalization: Federated Recommendation with Calibration via Low-rank Decomposition

TL;DR

This paper introduces PFedCLR, a federated recommendation method that calibrates user embeddings using low-rank decomposition to improve personalization, efficiency, and privacy while addressing embedding skew.

Contribution

It proposes a novel calibration mechanism with low-rank decomposition to mitigate user embedding skew in federated recommendation systems.

Findings

PFedCLR reduces user embedding skew effectively.

It outperforms state-of-the-art methods in accuracy.

The approach balances performance, efficiency, and privacy.

Abstract

Federated recommendation (FR) is a promising paradigm to protect user privacy in recommender systems. Distinct from general federated scenarios, FR inherently needs to preserve client-specific parameters, i.e., user embeddings, for privacy and personalization. However, we empirically find that globally aggregated item embeddings can induce skew in user embeddings, resulting in suboptimal performance. To this end, we theoretically analyze the user embedding skew issue and propose Personalized Federated recommendation with Calibration via Low-Rank decomposition (PFedCLR). Specifically, PFedCLR introduces an integrated dual-function mechanism, implemented with a buffer matrix, to jointly calibrate local user embedding and personalize global item embeddings. To ensure efficiency, we employ a low-rank decomposition of the buffer matrix to reduce the model overhead. Furthermore, for privacy, we train and upload the local model before personalization, preventing the server from accessing sensitive information. Extensive experiments demonstrate that PFedCLR effectively mitigates user embedding skew and achieves a desirable trade-off among performance, efficiency, and privacy, outperforming state-of-the-art (SOTA) methods.