A Decentralized Federated Learning model based on population mobility networks: A case study of the COVID-19 pandemic

Abstract

This study proposes a Decentralized Federated Learning(DFL) framework based on population mobility networks to address the prediction challenges in multi-node dynamic systems. By incorporating an adjacency matrix constructed from population mobility networks, the framework ensures node independence while enabling dynamic information sharing and global collaborative optimization across nodes. Using the cumulative COVID-19 case predictions for 29 provinces, municipalities, and by using various provinces, autonomous regions, and municipalities in China as case studies, the performance of the column-randomized DFL model(CR-DFL) is systematically compared with that of the Isolated models(IS). Experimental results demonstrate that the CR-DFL model achieves significantly lower root mean square error (RMSE) than IS models in most nodes, with only three exceptions (Hebei, Tianjin, and Hubei) where the RMSE is slightly higher. Further analysis reveals that the CR-DFL model effectively captures complex dynamic relationships among nodes while accurately modeling global system dynamics. However, for certain nodes with unique dynamic characteristics, the prediction accuracy may be affected, indicating the need for further optimization to enhance robustness and adaptability. Additionally, this study explores potential directions for improvement, including optimizing communication efficiency, addressing data heterogeneity, and integrating multi-modal datasets.