Explainability in Graph Neural Networks: A Taxonomic Survey

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Data Analysis

Background:

• Deep learning models excel in AI tasks but lack interpretability.
• Explainability techniques are crucial for understanding model predictions.
• Graph Neural Networks (GNNs) are rapidly developing, but GNN explainability lacks standardization.

Purpose of the Study:

• To provide a unified and taxonomic view of current GNN explainability methods.
• To establish a standardized benchmark and testbed for evaluating GNN explainability.
• To facilitate further methodological developments in the field.

Main Methods:

• A comprehensive survey and classification of existing GNN explainability techniques.
• Development of a testbed including datasets, algorithms, and evaluation metrics.
• Conducting comparative experiments to analyze method performance.

Main Results:

• A unified taxonomy highlighting commonalities and differences in GNN explainability methods.
• A functional testbed for reproducible evaluation of GNN explainability.
• Empirical analysis comparing the performance of various GNN explainability techniques.

Conclusions:

• The survey offers a structured overview of GNN explainability.
• The provided testbed enables standardized and rigorous evaluations.
• This work lays the foundation for advancing GNN explainability research and development.