Contrastive learning of graphs under label noise

Area of Science:

• Machine Learning
• Graph Neural Networks
• Data Mining

Background:

• Semi-supervised node classification is crucial for graph-structured data, but performance degrades with label noise and sparsity.
• Existing Graph Neural Networks (GNNs) are vulnerable to label noise, especially when labeled data is scarce.

Purpose of the Study:

• Propose a novel approach, Contrastive Robust Graph Neural Network (CR-GNN), to enhance robustness against label noise in semi-supervised node classification.
• Improve GNN performance in scenarios with sparse and noisy labels.

Main Methods:

• Employ unsupervised contrastive loss and neighbor contrastive loss incorporating graph homophily.
• Introduce a dynamic cross-entropy loss to mitigate overfitting to label noise by selecting reliable nodes.
• Incorporate cross-space consistency to bridge the semantic gap between contrastive and classification tasks.

Main Results:

• CR-GNN demonstrates superior performance in resisting label noise compared to existing methods.
• Extensive experiments on multiple public datasets validate the effectiveness of the proposed approach.

Conclusions:

• CR-GNN offers a robust solution for semi-supervised node classification in the presence of label noise.
• The method effectively handles sparse and noisy labels, outperforming current state-of-the-art techniques.