关键词: 子图联邦学习, 多尺度信息融合卷积, 图卷积网络, 注意力机制

Abstract:  Subgraph federated learning processes subgraphs of the global graph locally using a graph convolutional network and updates these parameters via a server, enabling the protection of user privacy. Existing approaches lack attention to nodes that are more important for certain tasks or graph structures, which may reduce the efficiency of node embedding. This research proposes FedMFG, a framework for graph federation learning, which introduces multi-scale fusion convolution to integrate node features with neighbor information, thus enhancing node feature representation. The framework passes parameters in pre-training rounds to reduce the communication cost and applies an attention mechanism on the server side to dynamically adjust the weights for better aggregation of global parameters. Experimental results on a standard reference dataset show that FedMFG has higher accuracy, stability, and lower communication cost than previous studies.

Key words: subgraph federated learning, multi-scale information fusion convolution, GCN, attention mechanism