Federated Active Learning via Class-adaptive Local–Global Balancing

ICLR 2026 Conference Submission21389 Authors

Keywords: active learning, federated learning, decentralized learning

Abstract:

Active learning has emerged as a pivotal approach for addressing data scarcity and annotation cost constraints in machine learning systems. However, its implementation in federated learning settings introduces unique challenges, particularly concerning data heterogeneity across clients. Our comprehensive analysis of existing centralized and decentralized methodologies reveals that state-of-the-art federated active learning techniques do not always outperform simpler baselines where centralized techniques are applied independently to clients. We identify a critical trade-off in performance: decentralized approaches excel when inter-client data heterogeneity is minimal, while centralized methods demonstrate superior performance under high heterogeneity conditions. Moreover, we observe a class-dependent variance phenomenon where the efficacy of each approach strongly correlates with the distribution variance of class samples across federated clients, highlighting critical bounds that limit existing methods. To address these limitations, we propose Adaptive Hybrid Federated Active Learning (AHFAL), a novel framework that dynamically integrates centralized and decentralized paradigms based on class-specific distribution characteristics. AHFAL combines enhanced entropy-based sampling with heterogeneity mitigation strategies, adaptively selecting the optimal paradigm per class based on cross-client variance metrics. Experiments across diverse datasets demonstrate that AHFAL outperforms state-of-the-art methods by prioritizing heterogeneity management over traditional uncertainty
sampling, particularly in low-resource and high heterogeneity scenarios.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 21389