Year: 2025
Type: article
Abstract: Federated learning (FL) enables distributed model training across multiple clients without sharing raw data, preserving privacy and reducing communication overhead. A key challenge in FL is sequentially scheduling clients for training and transmission to balance generalization performance with latency constraints. In this work, we address this challenge through a multi-armed bandit (MAB) framework that models client scheduling as a sequential decision-making problem. We propose a novel MAB approach for sequential client scheduling in FL, designed to minimize training latency without compromising model generalization, the ability to make accurate predictions on unseen data. Building on this framework, we develop a low-complexity algorithm that efficiently balances this trade-off. We prove that the algorithm achieves logarithmic regret over time, relative to an oracle with full knowledge of client latency means. Experimental results on both synthetic and real-world datasets demonstrate that our method consistently outperforms existing client selection strategies in terms of convergence speed and generalization performance.
Authors: Ben-Ami, Dan, Cohen, Kobi, Zhao, Qing
Institutions: Ben-Gurion University of the Negev, Cornell University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
Funder: National Science Foundation
OpenAlex ID: W7220798634
