Boosting-GNN: Boosting Algorithm for Graph Networks on Imbalanced Node Classification.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Data Science

Background:

• Graph Neural Networks (GNNs) are prevalent for graph data representation.
• Existing GNN research predominantly focuses on balanced datasets, neglecting imbalanced scenarios.
• Traditional imbalanced data techniques are often incompatible with GNN architectures.

Purpose of the Study:

• To propose an effective GNN-based ensemble model for imbalanced graph datasets.
• To enhance classification accuracy and reliability in GNNs facing data imbalance.
• To leverage transfer learning for improved computational efficiency and fitting ability.

Main Methods:

• Developed Boosting-GNN, an ensemble model utilizing GNNs as base classifiers.
• Implemented a boosting strategy that assigns higher weights to misclassified training samples.
• Integrated transfer learning to optimize computational cost and generalization.

Main Results:

• Boosting-GNN demonstrated superior performance compared to GCN, GraphSAGE, GAT, SGC, and N-GCN on synthetic imbalanced datasets.
• The model outperformed advanced reweighting and resampling methods.
• Achieved an average performance improvement of 4.5% over baseline methods.

Conclusions:

• Boosting-GNN offers a robust solution for imbalanced graph classification tasks.
• The proposed method significantly improves accuracy and reliability over existing GNNs and traditional techniques.
• The integration of boosting and transfer learning enhances the model's effectiveness and efficiency.