StabFedVAE: Stabilizing Global-Local Latent Representations for Federated Recommendation

Abstract

: Federated recommendation leverages distributed client data without centralizing interaction histories. However, sparse and non-identically distributed data can cause scale mismatches between global and local latent representations. We propose StabFedVAE, a global-local variational autoencoder (VAE) that disentangles shared and client-specific preference patterns. A sigmoid-gated multiplicative fusion regulates local latent contributions with provable contraction properties and bounded variance. While the global encoder-decoder are shared, local encoders remain private. We introduce a round-dependent local Kullback–Leibler (KL) penalty to stabilize the local posterior. Theoretical analysis proves that for fixed global gates, the fusion is contractive and bounds local latent variance. Empirical validation across four datasets confirms the observable variance inequality. Experiments show that moderate local KL regularization improves normalized discounted cumulative gain at cutoff 20 (NDCG@20) by 3.0%–10.3% over no regularization, while stronger penalties reduce posterior dispersion at the cost of ranking quality. Our results demonstrate that balancing latent stability and client personalization is essential rather than optimizing them independently.