FedVRU: Variational representation decomposition with uncertainty-aware refinement

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Federated Learning (FL) enables privacy-preserving collaborative training. However, data heterogeneity across clients degrades the performance of a single global model. Existing Personalized Federated Learning (PFL) methods partially mitigate this issue via client-specific modeling. However, most rely on deterministic point estimation, ignoring the implicit uncertainty within personalized features and leaving the joint representation space vulnerable to local noise. We propose FedVRU , a unified framework for Federated V ariational R epresentation Decomposition with U ncertainty-aware Refinement. Specifically, FedVRU models client-specific representations probabilistically through variational learning to characterize feature uncertainty. An Uncertainty-Aware Gating (UAG) mechanism then adaptively regulates personalized features according to the estimated uncertainty. Concurrently, a Multi-Contrastive Distillation (MCD) objective promotes temporal consistency of global representations and separation between global and personalized representations. Extensive experiments on multiple heterogeneous benchmarks demonstrate the effectiveness of FedVRU across diverse federated settings without additional communication overhead. The code is available at https://github.com/Baohua-Judy/FedVRU .