Federated Graph Neural Network for Cross-graph Node Classification

Abstract

In this paper, we propose a novel distributed scalable federated graph neural network (FGNN) to solve the cross-graph node classification problem. In the existing cross-graph node classification methods, the source graph and target graph need to share their graph data and label, for the nodes in the source graph and target graph are in the same semantic space. However, source graphs cannot share graph data and label without encryption due to regulations and interests. In order to satisfy the privacy of all parties, the universal classification rules of cross-graph nodes are learned. We add PATE mechanism into the domain adversarial neural network (DANN) to construct a cross-network node classification model, and extract effective information from node features of source and target graphs for encryption and spatial alignment. Moreover, we use a one-to-one approach to construct cross-graph node classification models for multiple source graphs and the target graph. Federated learning is used to train the model jointly through multi-party cooperation to complete the target graph node classification task. Finally, we carry out extensive experiments on five datasets to demonstrate the effectiveness of the proposed method.