Global explanation supervision for Graph Neural Networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Graph Neural Networks (GNNs) are increasingly popular for predictive tasks on graph-structured data.
• Explainability of GNNs is critical, yet current methods primarily focus on generating explanations rather than their accuracy or global applicability.
• Existing GNN Explanation Supervision (GNES) framework improves local explanation reasonableness but lacks global scope.

Purpose of the Study:

• To address the limitations of local explanations in GNNs.
• To develop a method for generating accurate and faithful global explanations for GNN predictions.
• To enhance the explainability power of the GNES framework for a global understanding of GNNs.

Main Methods:

• Proposed Global GNN Explanation Supervision (GGNES) technique.
• Utilized a trained GNN to create local explanations.
• Integrated local explanations into a Global Logic-based GNN Explainer for global explanation learning.
• Employed iterative training of GNN and explainer for reasonable global explanations.

Main Results:

• GGNES effectively improves the quality of global explanations for GNNs.
• The proposed method maintains or even enhances the predictive performance of the backbone GNN model.
• Experimental results demonstrate the effectiveness of GGNES in generating reasonable and faithful global explanations.

Conclusions:

• GGNES offers a significant advancement in GNN explainability by focusing on global explanations.
• The iterative training approach ensures both explanation reasonableness and model performance.
• GGNES provides a robust solution for applications requiring global understanding of GNN behavior.