Graph Contrastive Learning With Adaptive Proximity-Based Graph Augmentation.

Area of Science:

• Graph Neural Networks (GNNs)
• Machine Learning
• Artificial Intelligence

Background:

• Graph neural networks (GNNs) excel in graph-based tasks.
• Improving GNN performance often involves capturing long-range node relationships, primarily in supervised settings.
• Unsupervised learning offers a way to leverage graph data without labels.

Purpose of the Study:

• To propose an unsupervised learning pipeline for Graph Neural Networks (GNNs) inspired by contrastive learning (CL).
• To efficiently inject long-range similarity information into GNNs.
• To address the diminishing utility of augmented views during unsupervised training.

Main Methods:

• Reconstructing the original graph in feature and topology spaces to create three augmented views.
• Employing a contrastive learning approach to maximize agreement between augmented views and the original graph representations.
• Introducing a novel view update scheme to adaptively adjust augmented views for sustained information gain.
• Optimizing an efficient channel-level contrastive objective to train a shared GNN encoder.

Main Results:

• The proposed method effectively captures long-range relationships in GNNs using an unsupervised contrastive learning framework.
• The adaptive view update scheme ensures augmented views provide continuous learning signals.
• Experiments on diverse assortative and disassortative graphs demonstrate the method's effectiveness.

Conclusions:

• The developed unsupervised pipeline enhances GNNs by effectively incorporating long-range dependencies.
• The adaptive view update mechanism is crucial for maintaining the utility of augmented data in contrastive learning.
• This approach provides a powerful tool for unsupervised representation learning on graphs.