Federated Graph Neural Networks

Federated Graph Neural Networks

Survey:Federated Graph Neural Networks: Overview, Techniques and Challenges

1.Taxonomy in FedGNN

2.Data Owners Related by a Graph

FedGNNs with a Central Server
如何聚合不同client上的图结构信息是关键。
（1）通过FL server来实现GNN的跨client的聚合
（2）更新局部图结构

FedGNNs without a Central Server
（1）Weighted Summation of FL Model Parameters
（2）Graph Regularization on FL Model Parameters

3.Data Owners Related by a Graph

The assumption is that graphs from different domains are stored by different data owners.

Clients with No Overlapping Nodes
Clients with Partially Overlapping Nodes
Clients with Completely Overlapping Nodes

Mainly focus on Clients with No Overlapping Nodes.

1. Federated graph classification over non-iid graphs.
   Main question: whetherreal-world graphs from heterogeneous sources (e.g., different datasets or even divergent domains) can provide useful common information among each other?
   Solution: cluster the similar clients.

GNN pretrain
1.GPT-GNN: generate the input graph.