A graph neural network explainability strategy driven by key subgraph connectivity

Area of Science:

• Graph Neural Networks (GNNs)
• Machine Learning Explainability
• Subgraph Analysis

Background:

• Current GNN explainability methods often overlook critical subgraphs, leading to fragmented insights.
• This limitation hinders reliable explanations for complex GNN decision-making processes.

Purpose of the Study:

• To propose and evaluate a novel key subgraph retrieval method for GNN explainability.
• To enhance the reliability and focus of GNN decision-making interpretations.

Main Methods:

• Utilized Euclidean distance for key subgraph retrieval.
• Employed node representations from GNNs trained on BA3 and Mutagenicity datasets.
• Conducted comparative performance experiments and visualization analyses.

Main Results:

• Achieved high accuracy rates: 99.25% on BA3 and 82.40% on Mutagenicity datasets.
• Demonstrated superior effectiveness and robustness compared to existing explainability strategies.
• Visualizations confirmed the method's ability to identify significant explanatory subgraphs.

Conclusions:

• The proposed key subgraph retrieval method offers a more effective approach to GNN explainability.
• Focusing on subgraphs provides more coherent and reliable insights into GNN decisions.
• This technique advances the interpretability of complex graph-based machine learning models.