Evaluating explainability for graph neural networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Evaluating graph neural network (GNN) explanations is critical for understanding their behavior.
• Current graph datasets lack reliable ground-truth explanations, hindering quality assessment.

Purpose of the Study:

• To introduce SHAPEGGEN, a synthetic graph data generator for creating benchmark datasets with ground-truth explanations.
• To develop GRAPHXAI, a comprehensive library for benchmarking GNN explainability methods.

Main Methods:

• SHAPEGGEN generates diverse synthetic graph datasets (varying sizes, degree distributions, homophily/heterophily).
• Datasets include ground-truth explanations to facilitate evaluation.
• The GRAPHXAI library integrates SHAPEGGEN, real-world datasets, GNN models, and evaluation tools.

Main Results:

• SHAPEGGEN provides flexible synthetic data generation mimicking real-world scenarios.
• GRAPHXAI offers a unified platform for benchmarking GNN explainability techniques.
• The availability of ground-truth explanations enables robust evaluation of explanation quality.

Conclusions:

• SHAPEGGEN and GRAPHXAI address the challenge of evaluating GNN explanations by providing reliable ground-truth data.
• This facilitates the development and assessment of more accurate and trustworthy GNN explainability methods.