Taming Data Heterogeneity in Federated Graph Learning 32 views

Graph learning has recently gained significant attention in both academia and industry. Traditionally, most graph learning models, such as Graph Neural Networks (GNNs), are trained on massive graph data. However, in many real-world scenarios, graphs are usually stored at multiple data owners and cannot be directly accessed by other parties due to privacy concerns and regulation restrictions. Federated Graph Learning (FGL) is a promising solution to learn graph learning models over graph data distributed in multiple data owners. Inherited from generic Federated Learning (FL), FGL also faces the issue of data heterogeneity where the data distribution may vary significantly across clients with distributed graph data. In response, this dissertation focuses on developing powerful FGL frameworks to address the issues caused by data heterogeneity. Specifically, this dissertation contributes to the advancement of FGL through three research themes. The first theme aims to tackle node-level heterogeneity in FGL. It introduces frameworks such as FedSpray to overcome diverse neighborhood information across clients and FedTAG to deal with text-attributed graphs with different pre-trained language models. The second theme aims to tackle subgraph-level heterogeneity in FGL. It addresses the distribution shift issue by introducing FedVN, which enables clients to learn a set of shared virtual nodes jointly. The third theme aims to tackle client-level heterogeneity in FGL. It develops FedGLS to handle graphless clients only owning node features by equipping each graphless client with a local graph learner that learns its local graph structure with the structure knowledge transferred from other clients and FedHERO to promote FGL with heterophilic graphs. Through these contributions, this dissertation advances the study of FGL from three different perspectives. It enhances the understanding of the data heterogeneity issue in the realm of FGL and paves the way for future advances.

PHD (Doctor of Philosophy)

Federated Graph Learning; Federated Learning; Graph Learning

English

2025-12-11