-FedVAE: Detached Posterior-Confidence Gating for Dimension-Wise KL Regularization in Personalized Federated Collaborative Filtering

Abstract

: Federated Variational Autoencoders (VAEs) keep interaction data local, but existing federated VAE recommenders typically apply uniform KL regularization and do not adapt dimension-wise penalties to unreliable posteriors in sparse interaction scenarios. We propose α -FedVAE, which uses a detached, clipped normalized signal-to-noise ratio as a local confidence gate for each KL dimension of a fused user posterior, without extra communication. Across MovieLens-100K, MovieLens-1M, and Amazon Video, α -FedVAE improves mean HR@20 by 7.8%–55.3% and NDCG@20 by 8.0%–65.8% over FedDAE. These results indicate that α -FedVAE improves personalized recommendation under sparse and decentralized settings while preserving the communication footprint.