Progressive Collaboration Pruning for Federated Learning

Abstract

Collaboration-based personalized federated learning (Co-PFL) mitigates data heterogeneity by enabling clients to selectively exchange knowledge through a learned collaboration graph. Despite their success, existing Co-PFL methods often update collaboration weights mainly from pairwise similarity, which can produce graphs that are either overly sparse and hinder knowledge propagation or overly dense and weaken personalization, leaving graph connectivity largely uncontrolled. In this paper, we propose FedPGP, a Co-PFL framework that explicitly schedules collaboration from dense to sparse over training rounds, enabling clients to first exploit broadly shared knowledge and then refine specialized knowledge aligned with their local distributions. FedPGP builds on a connectivity-aware similarity regularization that jointly accounts for global graph connectivity and client-model similarity, together with an efficient greedy pruning rule that removes edges with the least contribution to the regularized objective. We further present FedPGP+, a module-specific extension that learns different collaboration graphs for different model modules, allowing collaboration intensity to vary along model depth. Extensive experiments across diverse non-IID settings show that FedPGP and FedPGP+ consistently outperform state-of-the-art Co-PFL baselines, achieving up to 5.52% absolute accuracy improvement, while producing interpretable dense-to-sparse collaboration schedules.