Abstract
Federated graph learning (FGL) is a combination of graph representation learning and federated learning that utilizes graph neural networks (GNNs) to process complex graph-structured data while addressing data silo issues. However, during the local training of GNNs, each client only has access to a subgraph, significantly deteriorating performance. To address this issue, recent solutions propose completing the subgraph with pseudo graph nodes generated by a generator trained using the local subgraph. Despite their effectiveness, such methods may introduce biases as the local pseudo graph nodes cannot accurately represent the global graph distribution. To overcome this problem, we introduce MN-FGAGN, which mitigates the impact of missing neighbor information by generating pseudo graph nodes that follow the global distribution. The main idea of our approach is to partition the generative adversarial neural network into a client-side discriminator and a server-side generator. In this way, the generator can receive supervised information from all clients and can thus generate graph nodes that contain global information. Experiments on four real-world graph datasets show that it outperforms the state-of-the-art methods.