G2CL: Gradient-guided graph contrastive learning for eliminating the message contrastive conflict.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Representation Learning

Background:

• Graph contrastive learning methods using InfoNCE loss have advanced graph representation learning.
• Existing methods struggle with message contrastive conflict (MCC) due to InfoNCE loss and graph neural networks' message-passing, hindering negative sample similarity minimization.

Purpose of the Study:

• To address the message contrastive conflict (MCC) in graph contrastive learning.
• To resolve issues of false negative samples and long-tail conflict effect (LCE) under MCC.
• To propose a novel gradient-guided graph contrastive learning (G2CL) method.

Main Methods:

• Theoretically demonstrated the existence and impact of MCC, false negative samples, and LCE.
• Introduced a gradient-guided dynamic capturer to eliminate MCC.
• Developed a false negative strategy leveraging semantic and topological graph information.
• Proposed a pheromone-based message-passing mechanism to address LCE.

Main Results:

• The proposed G2CL method effectively eliminates MCC.
• The false negative strategy and pheromone-based mechanism address key challenges.
• Extensive experiments on 11 datasets validated G2CL's superiority.

Conclusions:

• G2CL significantly improves graph contrastive learning by resolving MCC.
• The novel components of G2CL offer effective solutions for false negatives and LCE.
• G2CL sets a new state-of-the-art performance benchmark in graph representation learning.