CDCGAN: Class Distribution-aware Conditional GAN-based minority augmentation for imbalanced node classification.

Area of Science:

• Graph Neural Networks
• Machine Learning
• Data Science

Background:

• Node classification is crucial for Graph Neural Networks (GNNs).
• Class imbalance in GNNs degrades minority class representation and overall performance.
• Existing data augmentation methods for minority classes are limited by their inability to capture class distributions and generate distinguishable samples.

Purpose of the Study:

• To address the limitations of existing methods in handling class imbalance in GNNs.
• To propose a novel generative adversarial network for augmenting minority class nodes.
• To generate diverse and distinguishable minority nodes that accurately reflect class distribution characteristics.

Main Methods:

• Proposed a Class Distribution-aware Conditional Generative Adversarial Network (CDCGAN).
• Extracted node embeddings and class distributions, preserving graph topology and attributes.
• Utilized conditional generator with nonlinear transformations and adversarial learning for diverse node generation.
• Implemented a joint discriminator for node discrimination and classification.

Main Results:

• CDCGAN effectively generates diverse and distinguishable minority nodes.
• The proposed method significantly improves GNN performance on node classification tasks.
• Experimental results on six datasets demonstrate superior performance compared to state-of-the-art methods.

Conclusions:

• CDCGAN successfully addresses the class imbalance problem in GNNs.
• The approach enhances the representation ability of minority classes.
• The method offers a promising solution for improving GNN classification accuracy in imbalanced scenarios.