Towards Inductive and Efficient Explanations for Graph Neural Networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Explaining Graph Neural Network (GNN) predictions is crucial but challenging, with current methods focusing on instance-level explanations.
• Existing local explanation methods lack generalizability and are inefficient for large datasets, hindering inductive learning.
• The need for global understanding and efficient, generalizable GNN explanation methods is unmet.

Purpose of the Study:

• To address the limitations of current GNN explanation techniques.
• To propose PGExplainer, a parameterized explainer for generating multi-instance explanations.
• To enhance the generalizability and efficiency of GNN interpretability.

Main Methods:

• Developed PGExplainer, a deep neural network-based parameterized explainer for GNNs.
• Enabled multi-instance explanation generation through a parameterized approach.
• Utilized explanation networks as regularizers for improved GNN generalization.

Main Results:

• PGExplainer demonstrates superior generalization ability and supports inductive settings without retraining.
• Achieved significant speed-up compared to existing leading explanation methods.
• Showcased highly competitive performance, with up to 24.7% relative improvement in AUC for graph classification.

Conclusions:

• PGExplainer provides an efficient and generalizable solution for GNN interpretability.
• The method facilitates inductive learning and improves overall GNN performance.
• PGExplainer represents a significant advancement in explaining complex graph-based models.