Data Distribution-Aware Online Client Selection Algorithm for Federated Learning in Heterogeneous Networks
Jaewook Lee, Haneul Ko, Sangwon Seo, Sangheon Pack
Abstract
Federated learning (FL) has received significant attention as a practical alternative to traditional cloud-centric machine learning (ML). The performance (e.g., accuracy and convergence time) of FL is hampered by the selection of clients having non-independent and identically distributed (non-IID) data. In addition, a long convergence time is inevitable if clients with poor computation or communication capabilities participate in the FL procedure (i.e., the straggler problem). To minimize convergence time while guaranteeing high learning accuracy, we first formulate an optimization problem on client selection. As a practical solution, we devise a data distribution-aware online client selection (DOCS) algorithm. In DOCS, the FL server finds several clusters having near IID data and then uses a multi-armed bandit (MAB) technique to select the cluster with the lowest convergence time. The evaluation results demonstrate that DOCS can reduce the convergence time by up to <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$10\% \!\sim\! 41\%$</tex-math></inline-formula> and improve the learning accuracy by up to <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$4\% \!\sim\! 13\%$</tex-math></inline-formula> compared to the traditional client selection schemes.