Scout: An Efficient Federated Learning Client Selection Algorithm Driven by Heterogeneous Data and Resource

Federated Learning is a novel distributed machine learning paradigm that leverages the computing power of numerous decentralized data sources for jointly training machine learning models while ensuring user privacy. In the most commonly used cross-device scenarios, the client cluster typically cover a vast number of heterogeneous end devices. Due to physical limitations such as bandwidth, only a few clients can participate in each round of training. The core issue of the client selection is to determine an appropriate client set for each training round. However, existing selection algorithms, especially the widely adopted random selection, suffer from a number of issues that prevent them from achieving a good balance between training efficiency and speed. Therefore, we propose Scout, which utilizes the heterogeneity features of clients’ data and resources to jointly model the utility function, and enhances the utilization of correlation among clients and the diversity among selected clients to achieve better training efficiency and speed. Furthermore, Scout maintains the scalability and fairness. Our experiments demonstrate that in large-scale heterogeneous clients scenarios, Scout outperforms three baseline algorithms and the state-of-the-art dual-feature dimension algorithm Oort in evaluation metrics.