Contrastive Graph Representation Learning with Adversarial Cross-View Reconstruction and Information Bottleneck.

Area of Science:

• Graph Neural Networks (GNNs)
• Machine Learning
• Data Mining

Background:

• GNNs excel at information aggregation but struggle with popularity bias and noisy labels in graph datasets.
• Existing Graph Contrastive Learning (GCL) methods often maximize mutual information, potentially introducing redundant information.
• Node classification tasks are hindered by suboptimal node representations due to these challenges.

Purpose of the Study:

• To propose an effective GCL method, CGRL, for node classification that overcomes popularity bias and noisy labels.
• To adaptively learn optimal graph structure representations by masking nodes and edges.
• To enhance node feature representation robustness and model generalization.

Main Methods:

• Developed Contrastive Graph Representation Learning (CGRL) incorporating adversarial cross-view reconstruction and information bottleneck theory.
• Implemented adaptive node and edge masking for optimal graph structure representation.
• Introduced information bottleneck theory to filter redundant information and noise perturbations for adversarial reconstruction.

Main Results:

• CGRL adaptively masks graph structures to obtain optimal representations.
• Information bottleneck theory effectively removes redundant information while preserving classification-relevant details.
• Adversarial cross-view reconstruction enhances the robustness of node feature representations.
• Theoretical analysis confirms the effectiveness of the proposed mechanisms.

Conclusions:

• CGRL significantly outperforms existing state-of-the-art algorithms on real-world datasets for node classification.
• The proposed method effectively addresses popularity bias and noisy labels in GNNs.
• CGRL offers improved robustness and generalization performance in graph representation learning.