Graph contrastive learning with node-level accurate difference.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Theory

Background:

• Graph contrastive learning (GCL) excels at self-supervised learning for graphs.
• Current GCL methods use predefined augmentations, potentially altering graph semantics.
• This can hinder distinguishing structurally similar but semantically different graphs.

Purpose of the Study:

• To develop a GCL framework that accurately quantifies graph dissimilarities.
• To enhance the model's ability to differentiate graphs with subtle differences.
• To improve the capture of relationships between graph samples.

Main Methods:

• Proposing Accurate Difference-based Node-Level Graph Contrastive Learning (DNGCL).
• Training a node discriminator to distinguish original and augmented nodes.
• Employing cosine dissimilarity to measure node-level differences.
• Utilizing multiple data augmentation strategies for richer local information.

Main Results:

• DNGCL effectively distinguishes between similar graphs with minor variations.
• The framework demonstrates superior performance across six benchmark datasets.
• Outperforms state-of-the-art baseline methods in graph contrastive learning tasks.

Conclusions:

• Quantifying graph differences is crucial for accurate GCL.
• DNGCL provides a robust method for learning node-level dissimilarities.
• The proposed approach advances self-supervised graph representation learning.