Reproduce “A Survey on Explainability of Graph Neural Networks” on OpenResearch

About this paper

Jaykumar Kakkad, Jaspal Jannu, Kartik Sharma, Charu Aggarwal, Sourav Medya

A detailed survey categorizes the expanding landscape of Graph Neural Network (GNN) explainability methods, proposing a new taxonomy that integrates both factual and counterfactual approaches. It systematically reviews various techniques, evaluation metrics, and datasets to foster a clearer understanding and guide future research in interpretable GNNs.

Key insights

• GNN explainability methods can be broadly classified into factual explanations (identifying contributing input features/substructures) and counterfactual explanations (determining minimal input changes to alter predictions).
• Factual methods are further categorized into self-interpretable (built-in explainability) and post-hoc methods (applied after training), with various sub-approaches like gradient-based, perturbation-based, and generation-based techniques.
• Counterfactual methods primarily generate explanations through perturbation, neural frameworks, or search-based techniques, each with distinct mechanisms for altering graph inputs to achieve desired prediction changes.

Summary by alphaXiv.