Federated Subgraph Learning via Global-Knowledge-Guided Node Generation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Federated graph learning (FGL) combines graph representation learning and federated learning using graph neural networks (GNNs).
• Local training in FGL uses only subgraphs, leading to performance degradation due to missing information.
• Existing methods for subgraph completion may introduce biases by not representing the global graph distribution.

Purpose of the Study:

• To introduce a novel federated graph learning approach, MN-FGAGN, to mitigate performance issues caused by missing neighbor information.
• To generate pseudo graph nodes that accurately represent the global graph distribution, overcoming local biases.
• To enhance the accuracy and robustness of GNNs in decentralized settings.

Main Methods:

• Proposed MN-FGAGN partitions a generative adversarial neural network (GAN) into a client-side discriminator and a server-side generator.
• The server-side generator receives supervised information from all clients to learn the global data distribution.
• Pseudo graph nodes are generated to incorporate global information, compensating for locally missing data.

Main Results:

• MN-FGAGN effectively mitigates the impact of missing neighbor information by generating globally representative pseudo graph nodes.
• The proposed client-discriminator and server-generator architecture enables the generator to learn from distributed data.
• Experiments on four real-world datasets demonstrate superior performance compared to state-of-the-art methods.

Conclusions:

• MN-FGAGN offers a robust solution for federated graph learning by addressing the challenge of local subgraph limitations.
• The approach successfully generates pseudo graph nodes that capture global graph distributions, improving model performance.
• This work advances federated graph learning by enabling more accurate and unbiased GNN training in decentralized environments.