Core Concepts
SosicFL proposes a novel method, Solution Simplex Clustered Federated Learning, to address the challenge of achieving good performance in federated learning with highly heterogeneous client distributions by assigning subregions of the solution simplex to each client.
Abstract
"SosicFL introduces a novel approach to federated learning, aiming to resolve the trade-off between global and local model performance. By assigning subregions of the solution simplex to clients based on their label distributions, SosicFL allows for personalized models within a common global model. This method improves both global and local performance while minimizing computational overhead. The experiments conducted demonstrate the effectiveness of SosicFL in accelerating training processes and enhancing accuracy for both global and personalized federated learning scenarios."
Stats
θk = 0.16 θ1 + 0.7 θ2 + 0.14 θ3
local accuracy improved accuracy on local client data clients