Federated Graph Learning with Cross-subgraph Missing Links Recovery

Federated Graph Learning (FGL) has been developed to enable multiple parties to collaboratively train a graph neural network while maintaining their local graph data. Despite its advantages, FGL is not immune to the issue of cross-subgraph link missing, which arises when there are edges between nodes in different subgraphs that are not observed by any of the parties. This can have a substantial negative impact on the performance of the Graph neural networks, particularly in cases where the missing edges are of crucial importance. This paper proposes a novel Federated Graph Learning Framework with Cross-subgraph Missing Links Recovery (FGL-CMLR) to address this issue by capturing the global graph structure and identifying the missing links. FGL-CMLR reduces the search space of nodes with potential link loss from the entire subgraph to a target set comprising clustered central and edge nodes. Next, a federated GCN-based link prediction model is trained to repair the lost links between central-edge nodes pairs belonging to different subgraphs. By aggregating representations of neighboring nodes in cross-subgraphs, the previous embedding of nodes with missing links can be corrected. Numerical experiments demonstrate that FGL-CMLR successfully alleviates the issue of cross-subgraph link missing and achieves improved accuracy.