Enhancing GNN learning with node augmentation

Area of Science:

• Machine Learning
• Graph Neural Networks
• Data Augmentation

Background:

• Graph Neural Networks (GNNs) excel in various applications but struggle with overfitting and poor generalization on limited or low-diversity graph datasets.
• Existing graph data augmentation techniques often use static transformations and do not generate new training samples, limiting their effectiveness.

Purpose of the Study:

• To introduce Node-based Augmentation (NAug), a novel framework for synthesizing new data samples to augment training sets for GNNs.
• To address the limitations of existing augmentation methods by employing a learning-based generation process that produces semantically and structurally consistent data.

Main Methods:

• NAug utilizes two core components: an augmented node feature generator and a link placement predictor.
• These components are jointly trained to create new node features and edges, ensuring consistency with the local graph context.
• The framework is designed to be versatile, supporting multiple downstream graph learning tasks.

Main Results:

• NAug was evaluated on eleven benchmark datasets across node classification, link prediction, and graph classification tasks.
• The framework consistently outperformed state-of-the-art augmentation methods.
• NAug also demonstrated improvements in machine learning safety measures, including calibration, robustness to noise, and adversarial attacks.

Conclusions:

• Node-based Augmentation (NAug) is an effective and versatile framework for general-purpose data augmentation in GNNs.
• The learning-based synthesis approach significantly enhances GNN performance and robustness.
• NAug offers a promising solution for improving GNNs trained on limited or challenging graph datasets.