Federated learning is a novel framework for training artificial intelligence models on distributed user data. In this framework, instead of transmitting raw data, only model weights are sent from local user devices to a central server an‎d aggregated to create a shared global model. For the next learning round, this aggregated model is sent to users as a starting point, implicitly sharing information an‎d progressively improving model quality. However, a major challenge in this field is data heterogeneity, which leads to reduced model accuracy, slower convergence, an‎d unfair performance among users. This research aims to solve the challenge of severe data heterogeneity by proposing a new approach based on dynamic user grouping an‎d intelligent user selec‎tion. A spectral clustering algorithm is used, based on the model weights transmitted by users, to improve the separation of users with different data distributions. Additionally, the proposed sampling subsystem selec‎ts users from within their clusters in each learning round using a multi-armed ban‎dit algorithm based on their past performance. This simultaneously improves both the modelʹs generalization an‎d fairness. Test results show that the proposed approach achieves complete dominance in accuracy over the baseline method in 83% of learning rounds in scenarios with severe data heterogeneity. It also improves fairness among users by an average of 21%. Furthermore, the networkʹs convergence speed increases significantly. In 500-round simulations, the networkʹs accuracy, on average, surpasses the mean of the top 10% of the baselineʹs best rounds from round 66 onwards. Keywords Federated Learning, Data Heterogeneity, Clustering, Fairness an‎d Inequality.

امير خورسندي كوهانستاني

محمدرضا احمدزاده , سمانه حسيني