GNN-EGG: Graph neural network explanations via graph generation

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Theory

Background:

• Graph Neural Networks (GNNs) excel at modeling network data but suffer from opaque predictions, limiting their application.
• Existing explanation methods for GNNs face challenges with consistency and random baseline issues.

Purpose of the Study:

• To introduce GNN-EGG, an architecture-agnostic post-hoc explanation algorithm for GNN classifiers.
• To address the 'black box' nature of GNNs by generating class-specific graph distributions.
• To improve the reliability and interpretability of GNN predictions.

Main Methods:

• GNN-EGG utilizes a differentiable approximation of Graph Edit Distance (GED) within its loss function.
• This approach ensures consistency between the graph and embedding spaces for representative examples.
• The method learns the data-generating distribution for each graph class.

Main Results:

• GNN-EGG demonstrates improved performance compared to state-of-the-art methods on the MUTAG dataset.
• The algorithm effectively mitigates issues related to random graph baselines.
• Successful application to large-scale GNNs for malignancy detection in digital pathology.

Conclusions:

• GNN-EGG provides a robust and effective solution for explaining GNN classifiers.
• The method enhances GNN interpretability and reduces reliance on opaque models.
• GNN-EGG shows promise for applications in sensitive domains like medical diagnostics.