Self-supervised contrastive graph representation with node and graph augmentation.

Area of Science:

• Graph representation learning
• Knowledge engineering
• Machine learning

Background:

• Graph representation is crucial for knowledge engineering, with contrastive learning advancing self-supervised methods.
• Existing methods modify graph structures for augmentation, harming performance on structure-sensitive graphs.
• Current contrastive learning has imbalanced positive-negative node pairs, potentially misclassifying similar samples.

Purpose of the Study:

• To propose a novel Virtual Masking Augmentation (VMA) method for self-supervised graph representation.
• To generate augmentation graphs without altering original graph structures.
• To improve contrastive learning by augmenting positive node pairs and balancing sample distribution.

Main Methods:

• Virtual Masking Augmentation (VMA) generates an equivalent augmentation graph by preserving original structures.
• A node augmentation technique identifies similar nodes within the same graph to enhance positive pairs.
• The proposed method utilizes two distinct augmentation graphs within a contrastive learning framework.

Main Results:

• VMA successfully generates augmentation graphs without structural modifications.
• The node augmentation method effectively enriches positive node pairs.
• Extensive experiments on large datasets confirm state-of-the-art performance.

Conclusions:

• The proposed VMA method offers a significant advancement in self-supervised graph representation learning.
• Preserving original graph structures is vital for structure-sensitive graph data.
• The approach achieves superior results, setting a new benchmark in the field.