Clustered Graph Federated Personalized Learning
Clustered Graph Federated Personalized Learning
复制标题
DOI：
10.1109/ieeeconf56349.2022.10051979
复制
发表时间：
2022-10
期刊：
影响因子：
--
通讯作者：
François Gauthier;Vinay Chakravarthi Gogineni;Stefan Werner;Yih-Fang Huang;A. Kuh
中科院分区：
文献类型：
--
作者：
François Gauthier;Vinay Chakravarthi Gogineni;Stefan Werner;Yih-Fang Huang;A. Kuh
本文提出了一种联合学习方法的图形，其中多个服务器协作以增强对群集客户的个性化学习，从本质上执行相关的学习任务。与较早的方法相反，依靠群集使用的服务器拓扑结构，提出的图形联合多任务学习（GFEDMT）框架采用了更一般的设置，其中同一群集的客户端在服务器之间分布。为了解决服务器和群集之间不平衡客户端分布的问题，以及隔离客户端的数据短缺，服务器通过与相邻服务器的本地交互进行了集群内和集群间学习。客户使用乘数的交替方向方法（ADMM）来学习其本地模型。数值模拟证明了所提出的方法在数据稀缺时确保快速准确收敛的能力。
This paper proposes a graph federated learning approach wherein multiple servers collaborate to enhance personalized learning over clustered clients, essentially performing correlated learning tasks. In contrast to earlier approaches, relying on a cluster-dedicated server topology, the proposed graph federated multitask learning (GFedMt) framework adopts a more general setting wherein clients of the same cluster are distributed across servers. In order to address problems with unbalanced client distributions among servers and clusters as well as data shortage of isolated clients, servers perform intra-cluster and inter-cluster learning collaboratively through local interaction with neighboring servers. Clients use the alternating direction method of multipliers (ADMM) to learn their local models. Numerical simulations demonstrate the ability of the proposed method to ensure fast and accurate convergence when data is scarce.