A Meta-Learning Approach for Training Explainable Graph Neural Networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Existing methods for explaining graph neural networks (GNNs) are applied post-training.
• These methods often identify subgraphs to interpret GNN predictions.
• There is a need for methods that improve explainability during the GNN training process.

Purpose of the Study:

• To propose a novel meta-training framework, MATE (MetA-Train to Explain), to enhance GNN explainability.
• To integrate explainability directly into the GNN training procedure.
• To improve the human-interpretability of GNN decisions without sacrificing accuracy.

Main Methods:

• MATE jointly trains a GNN for its primary task (e.g., node classification) and for generating human-friendly outputs.
• It employs meta-training to optimize GNN parameters towards minima that facilitate post-hoc explainers.
• The framework uses an on-the-fly GNNExplainer to guide the optimization of internal representations.

Main Results:

• MATE significantly improves the explainability of GNNs across various architectures and datasets.
• The proposed approach enhances the effectiveness of different explanation algorithms.
• Achieved improvements in explainability did not lead to a decrease in model accuracy.

Conclusions:

• MATE offers a model-agnostic method to boost GNN explainability at training time.
• The framework provides a way to generate more interpretable GNNs without performance trade-offs.
• MATE represents a significant advancement in developing trustworthy and understandable graph neural networks.