Dual-channel hard negative sample generation for graph contrastive learning

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Representation Learning

Background:

• Graph Contrastive Learning (GCL) excels at learning representations from graph data.
• Existing GCL methods suffer from low-quality negative samples, leading to performance degradation.
• Issues include semantically invalid negative samples and false negatives from indiscriminate processing.

Purpose of the Study:

• To propose a novel method, Dual-Channel Hard Negative Sample Generation for Graph Contrastive Learning (DCGCL).
• To address the limitations of negative sample quality in GCL.
• To enhance the generalization and performance of GCL models.

Main Methods:

• Employs a dual-channel graph generator for controlled perturbations to adjacency and feature matrices.
• Introduces a mechanism to maximize distributional divergence between original and perturbed graphs while constraining node probability distributions.
• Utilizes a two-stage training strategy to dynamically inject hard negative samples.

Main Results:

• DCGCL generates high structural and feature similarity between original and perturbed graphs.
• The method effectively differentiates graph semantics.
• Experimental results show significant performance enhancement on diverse downstream tasks.

Conclusions:

• DCGCL significantly improves GCL model performance by overcoming negative sample quality issues.
• The proposed method outperforms state-of-the-art baselines.
• DCGCL enables learning more generalized feature representations.