Exploring personalized federated learning from a distribution-based perspective

Tianhao Yu

doi:10.1016/j.patcog.2026.113774

Exploring personalized federated learning from a distribution-based perspective

2026-04-18

0

OA

AI

DOI：10.1016/j.patcog.2026.113774

Original

Save

View PDF

Abstract

En 中文

• Introduces a distribution-based Bayesian PFL framework beyond Gaussian assumptions. • Employs rank-1 parameterization to reduce computational and memory overhead. • Models group-specific posterior families to better capture heterogeneous client patterns. • Provides theoretical proof showing rank-1 inference preserves local variance structure. • Achieves superior accuracy and calibration across multiple federated benchmarks.

Keywords:

Federated learning

Bayesian inference

Uncertainty quantification

Exploring personalized federated learning from a distribution-based perspective

Abstract

AI Summary

Journal

Organization