Contrastive message passing for robust graph neural networks with sparse labels.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Network Science

Background:

• Graph Neural Networks (GNNs) excel in semi-supervised learning but struggle with limited labels and structural noise.
• Traditional message passing in GNNs is sensitive to perturbations, degrading performance on noisy graphs.
• Overfitting is a significant issue in GNNs when training data is scarce.

Purpose of the Study:

• To develop a robust framework for GNNs that addresses challenges of sparse labels and structural noise.
• To improve the classification accuracy and resilience of GNNs in real-world scenarios.
• To introduce novel supervision signals beyond limited node labels.

Main Methods:

• Proposed a noise-resistant framework utilizing contrastive message passing.
• Introduced contrastive graph likelihood, defined as the product of edge likelihoods for connected node pairs.
• Implemented two unfolding update steps: feature updating with edge probability initialization and binary edge application for homophily/heterophily views.

Main Results:

• Demonstrated superior performance in semi-supervised node classification tasks with sparse labels.
• Achieved excellent robustness against structural perturbations in graph data.
• The contrastive approach effectively mitigates overfitting and enhances generalization.

Conclusions:

• The proposed contrastive message passing framework significantly improves GNN performance under challenging conditions.
• The method offers a promising direction for developing more reliable GNNs for real-world applications.
• Effective integration of topological structure as supervision enhances model stability and accuracy.