Node transfer with graph contrastive learning for class-imbalanced node classification

Area of Science:

• Graph Representation Learning
• Machine Learning
• Data Science

Background:

• Class imbalance is a major challenge in graph representation learning.
• Existing methods struggle with both node quantity and feature space imbalance.
• Minority classes are often disadvantaged by majority class dominance.

Purpose of the Study:

• To introduce a novel framework, node transfer with graph contrastive learning (NT-GCL).
• To enhance graph neural network (GNN) representation for minority classes.
• To balance node quantity and feature space distributions in imbalanced datasets.

Main Methods:

• A node transfer algorithm redistributes misclassified nodes to balance quantity and feature space.
• This prevents majority classes from compressing minority class feature spaces.
• Self-supervised contrastive learning trains the model without labels, reducing bias.

Main Results:

• NT-GCL effectively balances node quantity and feature space distributions.
• The framework prevents feature space compression for minority classes.
• Experiments show NT-GCL's strong competitiveness in class-imbalanced node classification.

Conclusions:

• NT-GCL offers a robust solution for class-imbalanced node classification.
• The proposed methods significantly improve GNN performance on imbalanced graph data.
• This framework advances the field of graph representation learning for imbalanced datasets.