BI-FedGNN: Federated graph neural networks framework based on Bayesian inference

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• The Industrial Internet of Things (IIoT) generates vast data, presenting opportunities for enhanced machine learning through data sharing.
• Graph Neural Networks (GNNs) excel at learning from graph-structured data but face challenges with privacy and data concentration in IIoT.
• Existing methods struggle with privacy constraints and noisy or incomplete graph data in industrial settings.

Purpose of the Study:

• To propose a privacy-preserving federal graph learning framework for the Industrial Internet of Things.
• To enhance the accuracy and generalization of GNNs in IIoT environments despite data privacy restrictions.
• To develop a robust method that handles noisy graph structures and missing relational edges.

Main Methods:

• Introduced a federal graph learning framework based on Bayesian Inference (BI-FedGNN).
• Extended Bayesian Inference (BI) to Federal Graph Learning (FGL) by incorporating weighted random samples in client-side training.
• Improved the similarity of training graph data to real-world data structures.

Main Results:

• BI-FedGNN achieved a 0.5%-5.0% accuracy improvement over existing federal graph learning baselines.
• Experiments on heterogeneous graph datasets demonstrated at least a 1.4% improvement in classification accuracy.
• The framework effectively handles noisy graph structures and missing relational edges.

Conclusions:

• BI-FedGNN offers a robust and accurate solution for privacy-preserving graph learning in the IIoT.
• The framework enhances GNN performance by improving data representation and generalization capabilities.
• BI-FedGNN shows significant potential for applications involving sensitive and complex graph data in industrial settings.