基于相似度加速的自适应聚类联邦学习OA北大核心CSTPCD

In order to solve the problem of model performance degradation caused by data heterogeneity in the federated learning process,it is necessary to consider personalizing in the federated model.A new adaptively clustering federated learning(ACFL)algorithm via similarity acceleration was proposed,achieving adaptive acceleration clustering based on geometric properties of local updates and the positive feedback mechanism during clients federated training.By dividing clients into different task clusters,clients with similar data distribution in the same cluster was cooperated to improve the performance of federated model.It did not need to determine the number of clusters in advance and iteratively divide the clients,so as to avoid the problems of high computational cost and slow convergence speed in the existing clustering fed-eration methods while ensuring the performance of models.The effectiveness of ACFL was verified by using deep con-volutional neural networks on commonly used datasets.The results show that the performance of ACFL is comparable to the clustered federated learning(CFL)algorithm,it is better than the traditional iterative federated cluster algorithm(IFCA),and has faster convergence speed.