DPGCL: Dual pass filtering based graph contrastive learning.

Area of Science:

• Machine Learning
• Graph Representation Learning
• Artificial Intelligence

Background:

• Graph Contrastive Learning (GCL) excels at learning node/graph representations from graph data.
• Existing GCL methods primarily use low-frequency filtering, limiting effectiveness on heterophilous graphs where dissimilar nodes connect.

Purpose of the Study:

• Investigate the impact of high-frequency signals on GCL performance.
• Develop a more comprehensive GCL framework to capture both low and high-frequency graph information.
• Enhance representation learning for heterophilous graphs by considering neighbor diversity.

Main Methods:

• Experimentally analyze the influence of high-frequency signals in GCL.
• Propose a GCL framework combining low-pass and high-pass signal contrasts.
• Introduce Adap-infoNCE, a novel contrastive loss considering spatial and featural neighbors for adaptive negative sampling.

Main Results:

• Incorporating high-frequency signals demonstrably improves GCL performance.
• The proposed framework captures both low and high-frequency graph information effectively.
• Adap-infoNCE enhances representation learning by accounting for neighbor diversity in heterophilous graphs.

Conclusions:

• High-frequency signals are valuable for GCL, especially on heterophilous graphs.
• The novel GCL framework and Adap-infoNCE loss yield significant improvements over existing methods.
• The approach achieves state-of-the-art results on unsupervised benchmarks, even outperforming supervised methods.