Bias Mitigation for Federated Learning with Spatially Correlated Participation
Oussama Harrak, Malcolm Egan, Claire Goursaud, M. Morel, Conte, Alberto · HAL (Le Centre pour la Communication Scientifique Directe) · 2025
Federated learning is a key strategy to exploit data distributed throughout edge networks. However, clients may participate intermittently only when they observe relevant data. In spatially-distributed sensing, the activity of nearby clients can be correlated. In this paper, we study the impact of correlation and variable-size active client sets for FedSGD. We analyze the convergence for smooth nonconvex learning objectives and show that bias due to unequal participation can be effectively mitigated via importance weighting. We validate our analysis with experiments using the MNIST dataset and show the impact of constraints on the availability of communication links.