HarmoFGL: Harmonizing GNN Latent Factors for Federated Graph Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Federated graph learning (FGL) aims to train models on distributed graph data while preserving privacy.
• Existing FGL methods struggle with graph heterogeneity (feature and structural deviations) and effective aggregation strategies.

Purpose of the Study:

• To propose a novel framework, HarmoFGL, to address key challenges in federated graph learning.
• To harmonize graph neural network (GNN) latent factors for improved cross-client federated training.

Main Methods:

• Introduced the HarmoFGL framework utilizing a latent factor perspective.
• Developed an implicit feature crossing (IFC) approach to mitigate feature deviation.
• Established a cross-client symbiotic parameter space for GNN latent factors to handle structural deviation.
• Implemented client relevance-driven personalized parameter aggregation and graph Laplacian regularization.

Main Results:

• The HarmoFGL framework effectively harmonizes GNN latent factors for federated graph learning.
• Demonstrated significant improvements in mitigating feature and structural deviations across clients.
• Validated the framework's effectiveness on diverse public and medical graph datasets.

Conclusions:

• HarmoFGL offers a robust solution for privacy-preserving federated graph learning.
• The proposed methods effectively address graph heterogeneity and enhance collaborative training benefits.
• The framework shows strong performance across various real-world graph data scenarios.