Generalized heterophily graph data augmentation for node classification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Graph data augmentations are effective for homophilic graph neural networks (GNNs).
• These augmentations show reduced effectiveness and performance on heterophilic graphs.
• Existing methods struggle to adapt augmentation strategies for heterophilic graph learning.

Purpose of the Study:

• To propose GePHo, a unified augmentation approach for heterophilic GNNs.
• To leverage self-supervised learning and graph data augmentation for improved GNN performance.
• To develop a type-agnostic pseudo-homophily graph generation for broader applicability.

Main Methods:

• GePHo employs a regularization technique based on self-supervised learning.
• It generates a type-agnostic pseudo-homophily graph to guide model learning.
• A sharpening technique and auxiliary pseudo-labels are used to regularize neighbors and constrain node representations.

Main Results:

• GePHo demonstrates competitive effectiveness in node classification tasks across multiple datasets.
• Experiments show significant performance improvements on both homophilic and heterophilic graphs.
• Ablation studies confirm the efficacy of GePHo's graph data augmentation strategies.

Conclusions:

• GePHo offers a unified and effective approach for graph data augmentation in GNNs.
• The method successfully addresses the limitations of traditional augmentations on heterophilic graphs.
• GePHo enhances GNN performance by constraining local and global node representations through novel augmentation techniques.