Layer-Adaptive-Augmentation-Based Graph Contrastive Learning With Feature Decorrelation.

Area of Science:

• Machine Learning
• Graph Representation Learning
• Artificial Intelligence

Background:

• Graph Contrastive Learning (GCL) methods are crucial for learning representations in label-scarce scenarios.
• Existing GCL methods face limitations in adaptive augmentation and intra-graph feature redundancy.
• Suboptimal model robustness and feature redundancy hinder GCL performance.

Purpose of the Study:

• To propose a novel Graph Contrastive Learning framework, LGCLD, addressing limitations of existing methods.
• To enhance model robustness through layer-wise adaptive augmentation.
• To learn informative and non-redundant graph representations by optimizing inter-graph agreement and intra-graph feature decorrelation.

Main Methods:

• Developed a layer-wise adaptive augmentation technique for dynamic, semantically similar graph perturbations.
• Introduced an Agreement-Decorrelation (AD) loss to optimize graph-level representation agreement and feature decorrelation.
• Analyzed the AD loss using the graph information bottleneck principle.

Main Results:

• LGCLD demonstrates improved model robustness via adaptive augmentation.
• The AD loss effectively reduces dimensional feature redundancy within graph representations.
• Experiments show LGCLD achieves competitive or superior performance across various graph datasets.

Conclusions:

• LGCLD offers a robust and effective approach to Graph Contrastive Learning.
• The proposed methods address key limitations in augmentation and representation learning.
• LGCLD advances the state-of-the-art in graph representation learning for label-scarce tasks.