Fedadasparse: Communication-Efficient Personalized Federated Learning Via Similarity-Based Client Clustering And Adaptive Gradient Sparsification For Non-IID Edge Data

Abstract

Federated learning (FL) enables collaborative model training across distributed clients without centralizing raw data, but its practical deployment on resource-constrained edge devices is hindered by two persistent challenges: statistical heterogeneity (non-independent and identically distributed, non-IID, client data) and the high communication cost of exchanging full model updates every round. This paper proposes FedAdaSparse, a federated learning framework that jointly addresses both challenges through (i) similarity-based client clustering, which groups clients with comparable local update directions to produce personalized cluster-level models rather than forcing convergence to a single global model, and (ii) adaptive top-k gradient sparsification, which progressively reduces the fraction of transmitted parameters as training stabilizes. We evaluate FedAdaSparse against the FedAvg baseline on a controlled non-IID benchmark constructed via Dirichlet label-skew partitioning across 12 simulated clients. Averaged over three random seeds, FedAdaSparse reduces cumulative communication cost by 65.6% at convergence while achieving statistically comparable final accuracy to FedAvg (48.4% vs. 50.0%) and, notably, achieves faster and higher peak accuracy during early-to-mid training (up to +13.1% at round 14). We further analyze a limitation intrinsic to periodic re-clustering — transient accuracy dips at cluster-reassignment rounds — and discuss mitigations. These results suggest that combining personalization with adaptive sparsification is a promising, practical direction for deploying FL on communication-constrained edge networks.