ArieL: Adversarial Graph Contrastive Learning

Area of Science:

• Graph Representation Learning
• Unsupervised Learning
• Machine Learning

Background:

• Contrastive learning is key for unsupervised graph representation learning, relying on positive and negative sample construction.
• Existing methods often use node proximity, while data augmentation from computer vision shows promise but is challenging for graphs.
• Generating high-quality contrastive samples for graph data augmentation remains an open area for improvement.

Purpose of the Study:

• To propose a simple yet effective method, Adversarial Graph Contrastive Learning (ArieL), for extracting informative contrastive samples in graph representation learning.
• To address the challenges of data augmentation in graph domains by introducing an adversarial graph view.
• To generalize the proposed method for both node-level and graph-level contrastive learning.

Main Methods:

• Introduced an adversarial graph view for data augmentation to generate informative contrastive samples.
• Developed information regularization for stable training and subgraph sampling for scalability.
• Generalized the approach from node-level to graph-level contrastive learning by treating graphs as super-nodes.

Main Results:

• ArieL consistently outperformed current graph contrastive learning methods on real-world datasets for both node-level and graph-level classification.
• Demonstrated enhanced robustness of ArieL against adversarial attacks.
• Successfully extracted high-quality contrastive samples within reasonable constraints.

Conclusions:

• Adversarial Graph Contrastive Learning (ArieL) offers a powerful new approach for unsupervised graph representation learning.
• The method effectively addresses limitations in graph data augmentation for contrastive learning.
• ArieL shows significant potential for improving graph classification tasks and robustness.