GNNExplainer: Generating Explanations for Graph Neural Networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Graph Neural Networks (GNNs) are powerful for machine learning on graphs, integrating node features and graph structure.
• Explaining GNN predictions remains a challenge due to model complexity.

Purpose of the Study:

• Introduce GnnExplainer, a general, model-agnostic method for interpretable GNN explanations.
• Identify crucial subgraph structures and node features impacting GNN predictions.

Main Methods:

• Formulate GnnExplainer as an optimization task maximizing mutual information.
• Identify compact subgraphs and feature subsets for instance-level explanations.
• Generate consistent explanations for classes of instances.

Main Results:

• GnnExplainer successfully identifies important graph structures and node features.
• Outperforms baseline approaches by up to 43.0% in explanation accuracy.
• Demonstrated effectiveness on synthetic and real-world graph datasets.

Conclusions:

• GnnExplainer provides interpretable insights into GNN predictions.
• Enables visualization of relevant structures and aids in debugging GNNs.
• Offers a general solution for GNN interpretability across tasks.