GMNI: Achieve good data augmentation in unsupervised graph contrastive learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Graph contrastive learning (GCL) is a powerful unsupervised approach for learning graph representations.
• Data augmentation (DA) is crucial for GCL, but optimal DA strategies are task-dependent and hard to determine in unsupervised settings.
• Existing GCL methods may suffer from insufficient or redundant information due to suboptimal DA, impacting performance.

Purpose of the Study:

• To propose a novel method, Minimal Noteworthy Information for unsupervised Graph contrastive learning (GMNI), for automated data augmentation in GCL.
• To address the challenge of selecting task-relevant information for DA in unsupervised GCL.
• To improve the quality of augmented views by balancing missing and excessive information.

Main Methods:

• GMNI employs an adversarial training strategy to generate views with minimal noteworthy information (MNI).
• Minimization optimization reduces nuisance information, while emphasis on noteworthy information ensures sufficient data.
• Randomness based on MNI is introduced to enhance view diversity and model stability.

Main Results:

• GMNI demonstrates superior performance over existing GCL methods across 14 datasets in unsupervised and semi-supervised settings.
• Achieved up to 1.64% improvement in unsupervised node classification.
• Achieved up to 1.97% improvement in unsupervised graph classification and up to 3.57% in semi-supervised graph classification.

Conclusions:

• GMNI effectively balances information in data augmentation for unsupervised GCL.
• The proposed method significantly enhances graph representation learning performance.
• GMNI offers a robust and superior alternative to existing DA strategies in GCL.