Boosting Graph Contrastive Learning via Adaptive Sampling.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Representation Learning

Background:

• Contrastive learning (CL) is a key self-supervised method for learning representations by contrasting positive and negative sample pairs.
• Graph CL excels at learning node representations without supervision, but uniform negative sampling limits its effectiveness.
• Uniform sampling can include uninformative nodes and incorrectly repel semantically similar nodes, hindering performance.

Purpose of the Study:

• To introduce an adaptive sampling strategy (AdaS) for graph contrastive learning.
• To enhance the learning from informative negative nodes and suppress negative impacts of false negatives.
• To improve the overall performance and discrimination ability of graph CL models.

Main Methods:

• Developed an adaptive sampling strategy (AdaS) to dynamically encode the importance of negative nodes.
• Introduced an auxiliary polarization regularizer to mitigate false negatives and boost discrimination.
• Evaluated AdaS on diverse real-world graph datasets.

Main Results:

• AdaS significantly improves the performance of graph contrastive learning models.
• The adaptive strategy effectively identifies and learns from the most informative negative nodes.
• The polarization regularizer successfully suppresses false negatives, enhancing model discrimination.

Conclusions:

• AdaS offers a more effective approach to negative sampling in graph CL.
• The proposed method enhances representation learning by focusing on informative contrasts.
• AdaS demonstrates superior performance across various real-world graph datasets.