Adversarial Augmentation With Maximum Discrepancy for Graph Contrastive Learning

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Graph contrastive learning (GCL) relies on data augmentation but often ignores relationships between augmentations.
• Existing discrete augmentation methods limit diversity and joint optimization in graph self-supervised learning (SSL).

Purpose of the Study:

• To propose a novel adversarial augmentation method, AMD-GCL, for jointly optimizing pairwise augmentations in GCL.
• To enhance the diversity and complementarity of augmentations for improved GCL performance.

Main Methods:

• Developed an adversarial augmentation constraint module to maximize the discrepancy between pairwise augmentations.
• Utilized graph reconstruction error in a continuous space as a surrogate for minimizing mutual information.
• Designed a min-max problem involving continuous adversarial perturbations and reconstruction error maximization, followed by a unified minimization problem.

Main Results:

• AMD-GCL demonstrates superior and robust performance across 18 datasets on various downstream tasks.
• The proposed method effectively addresses limitations of existing augmentation strategies in GCL.
• Experimental results validate the effectiveness of jointly optimizing pairwise augmentations.

Conclusions:

• AMD-GCL significantly advances graph contrastive learning by introducing effective adversarial augmentation strategies.
• The method offers a promising direction for improving self-supervised learning on graph data.
• AMD-GCL provides a robust framework for enhancing graph representation learning.