FA-GCL: Feature-augmented graph contrastive learning method.

Area of Science:

• Graph Representation Learning
• Machine Learning
• Data Science

Background:

• Existing graph contrastive learning methods often rely on complete node attributes or structural information.
• Incomplete node attributes and false positives from structure-enhancement hinder performance in real-world graph data.
• There is a need for robust graph representation learning techniques that are less sensitive to data completeness.

Purpose of the Study:

• To propose a novel Feature Augmentation-based Graph Contrastive Learning (FA-GCL) method.
• To enhance the accuracy and robustness of graph representations.
• To address limitations of existing methods in handling incomplete node attributes and structural noise.

Main Methods:

• Employs a dynamic dropout-based feature augmentation technique with a triangular wave function for adaptive dropout rates.
• Introduces two singular value decomposition (SVD) based feature augmentation methods: full SVD and randomized projection SVD.
• The SVD methods add controlled noise to singular values and reconstruct features for high-quality augmented samples, with randomized SVD offering linear complexity.

Main Results:

• FA-GCL demonstrates consistent superior performance across twelve graph datasets.
• The method significantly outperforms baseline approaches in node classification, node clustering, and graph classification tasks.
• Feature augmentation proves effective in improving the quality and robustness of learned graph representations.

Conclusions:

• FA-GCL offers a robust and effective approach to graph representation learning, particularly when node attributes are incomplete.
• The proposed feature augmentation strategies enhance model performance and generalizability.
• This work advances graph contrastive learning by introducing flexible and powerful data augmentation techniques.