Learning model-level explanations of graph neural networks via subgraph order embedding space

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Model-level graph neural network (GNN) explainers identify critical graph patterns for class prediction.
• Existing methods often lack constraints, producing unreliable and atypical patterns that hinder explanation quality.

Purpose of the Study:

• To propose MOSE (MOdel-level explanations via a Subgraph order Embedding space), a novel explainer for generating reliable and typical graph patterns.
• To enhance the generalization of GNN explainability by extending MOSE to node classification tasks.

Main Methods:

• MOSE utilizes a graph encoder to create an embedding space preserving subgraph relationships.
• A score function with greedy sampling generates constrained, reliable graph pattern candidates.
• Explanation candidates are selected based on the GNN's predicted probability.

Main Results:

• MOSE demonstrates effectiveness across multiple metrics, including predictive accuracy, model utility, and efficiency.
• Experiments on synthetic and real-world datasets validate MOSE's performance.
• The extension of MOSE to node classification shows enhanced generalization.

Conclusions:

• MOSE provides a simple yet effective approach to generating reliable and typical model-level GNN explanations.
• The method addresses limitations of current explainers by incorporating subgraph order embeddings.
• MOSE offers a promising direction for improving GNN interpretability and applicability.