FedSUM Family: Efficient Federated Learning Methods under Arbitrary Client Participation

2025-12-20

AI Summary

Existing federated learning methods often rely on restrictive assumptions—such as fixed client participation patterns or homogeneous data distributions—rendering them fragile under highly dynamic, arbitrary client availability. Method: This paper proposes FedSUM, a general family of algorithms that requires no assumptions on data distribution. It models participation variability via two complementary metrics—maximum delay τ_max and average delay τ_avg—and introduces three variants: FedSUM-B, FedSUM, and FedSUM-CR. These integrate delay-aware aggregation, decentralized gradient correction, communication compression, and asynchronous updates. Contribution/Results: We provide rigorous convergence analysis showing that FedSUM achieves an O(1/√T) rate under arbitrary client participation and non-IID data—without requiring statistical or system-level homogeneity. Empirical evaluation demonstrates superior robustness and broader applicability in realistic deployment scenarios, significantly improving practical feasibility over prior approaches.

Technology Category

• Machine Learning: Distributed Machine Learning & Federated Learning
• Natural Language Processing: Summarization
• Constraint Satisfaction and Optimization: Distributed CSP/Optimization

Application Category

• User Modeling, Personalization and Recommendation: Federated recommendation systems and personalization
• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• Systems and Infrastructure for Web, Mobile and WoT: Federated Web and WoT systems, including distributed, federated and edge-based data processing

Abstract

Federated Learning (FL) methods are often designed for specific client participation patterns, limiting their applicability in practical deployments. We introduce the FedSUM family of algorithms, which supports arbitrary client participation without additional assumptions on data heterogeneity. Our framework models participation variability with two delay metrics, the maximum delay $τ_{max}$ and the average delay $τ_{ ext{avg}}$. The FedSUM family comprises three variants: FedSUM-B (basic version), FedSUM (standard version), and FedSUM-CR (communication-reduced version). We provide unified convergence guarantees demonstrating the effectiveness of our approach across diverse participation patterns, thereby broadening the applicability of FL in real-world scenarios.

Problem

• Efficient federated learning under arbitrary client participation
• Modeling participation variability with delay metrics
• Providing unified convergence guarantees for diverse patterns

Innovation

• Supports arbitrary client participation patterns
• Models variability with maximum and average delay metrics
• Provides unified convergence guarantees across patterns