PAGE: Prototype-Based Model-Level Explanations for Graph Neural Networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Representation Learning

Background:

• Graph neural networks (GNNs) are powerful for graph representation learning but require explainability.
• Existing GNN explanation methods primarily focus on instance-level explanations.
• There is a growing need for model-level explanations that reveal what GNNs learn overall.

Purpose of the Study:

• To propose Prototype-bAsed GNN-Explainer (PGE), a novel model-level GNN explanation method.
• To generate human-interpretable prototype graphs that explain GNN learning for graph classification.
• To provide more concise and comprehensive explanations compared to instance-level methods.

Main Methods:

• PGE clusters class-discriminative input graph embeddings to select representative graphs.
• It iteratively searches for high-matching node tuples using node embeddings and a prototype scoring function.
• The method discovers common subgraph patterns, yielding a prototype graph as the explanation.

Main Results:

• PGE qualitatively and quantitatively outperforms state-of-the-art model-level explanation methods on six graph classification datasets.
• Experimental studies demonstrate PGE's relationship with instance-level methods.
• Robustness in data-scarce environments and computational efficiency of the prototype scoring function are shown.

Conclusions:

• PGE offers a novel and effective approach for model-level GNN explanations.
• The discovered prototype graphs provide human-interpretable insights into GNN decision-making.
• PGE advances the field of explainable AI for graph-based machine learning.