SemiGraphFL: Semi-supervised Graph Federated Learning for Graph Classification

GNNs have achieved remarkable performance on graph classification tasks. It can be attributed to the accessibility of abundant graph data, which are usually isolated by different data owners. Graph Federated Learning (GraphFL) allows multiple clients to collaboratively build GNN models without explicitly sharing data. However, all existing works assume that all clients have fully labeled data, which is impractical in reality. This work focuses on the graph classification task with partially labeled data. (1) Enhancing the collaboration processes: We propose a new personalized FL framework to deal with Non-HD data. Clients with more similar data have greater mutual influence, where the similarities can be evaluated via unlabeled data. (2) Enhancing the local training process: We introduce auxiliary loss for unlabeled data that restrict the training process. We propose a new pseudo-label strategy for our SemiGraphFL framework to make more effective predictions. Extensive experimental results prove the effectiveness of our design.