Joint Client Selection & Bandwidth Allocation in Federated Learning Based on Long-Term Data-Quality

Federated Learning is a decentralized framework that facilitates the training of a universal model by co-training on edge mobile client nodes. This approach eliminates the need to upload local data while allowing the use of non-independent and identically distributed (non-IID) training data on mobile clients. The non-IID data arises from the mismatch between the distribution of local records and the overall data distribution, which has a significant impact on the accuracy and the convergence speed of the model. Although client selection can be a solution to this problem, it faces several challenges related to the selection metrics. Previously proposed metrics for assessing the quality of mobile client data were computed independently for each round. In addition, these metrics also tied client selection to the upload of training results to the central server, resulting in a waste of resources for unused training results. To address these challenges, we propose a novel long-term client selection federated learning scheme based on data balance. This scheme computes the long-term data quality of each mobile client to alleviate data imbalance based on the current training states. Due to the limited resources in wireless communication, we have formulated a joint optimization problem for bandwidth allocation and client selection. Our experimental results on various datasets demonstrate its effectiveness in mitigating performance degradation caused by non-IID issues.