بررسي چالش داده‌هاي نامتوازن در يادگيري فدرالي شخصي‌سازي شده

In federated learning, the distribution of data across clients is often heterogeneous an‎d imbalanced, which can lead to significant degradation in the accuracy of the global model an‎d reduced training stability. In this study, we propose a novel approach to enhance federated learning performance under conditions of imbalanced data an‎d limited computational resources. Our method leverages an intelligent probabilistic sampling mechanism to simultaneously pursue two key objectives: selec‎ting clients with a more balanced data distribution an‎d higher data quality. Experimental results across multiple benchmark scenarios demonstrate that the proposed approach not only substantially reduces runtime compared to baseline algorithms such as Fed-AVG, Fed-CBS, an‎d CLIMB, but also achieves higher prediction accuracy, particularly for minority classes, while maintaining superior stability. Moreover, its convergence speed is markedly faster than that of alternative methods, making it an efficient an‎d practical solution for training deep learning models in federated learning settings.

5