Gen-GraphEx: Generative In-Distribution Graph Explanations for Time-Efficient Model-Level Interpretability of GNNs.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Graph neural networks (GNNs) are crucial for tasks involving graph data, impacting fields like recommendation systems and drug discovery.
• The increasing use of GNNs necessitates trustworthy and interpretable models, especially when they affect end-users directly.

Purpose of the Study:

• To introduce Gen-GraphEx, a novel model-agnostic, model-level explanation method for GNNs.
• To enhance the interpretability and trustworthiness of GNNs by generating explanation graphs.

Main Methods:

• Gen-GraphEx employs a graph generative model (GGM) to produce explanation graphs for a given class label.
• The method ensures explanation graphs are discriminative and distributionally consistent with real graphs belonging to the target class.
• It uniquely interpolates GGMs of different classes to explore decision boundaries.

Main Results:

• Gen-GraphEx generates explanation graphs that are faithful to the GNN's learned patterns.
• The method demonstrates computational efficiency and does not require additional deep learning modules for explanation.
• Comparative analyses on real and synthetic datasets show competitive performance against state-of-the-art explainers.

Conclusions:

• Gen-GraphEx offers a user-centric approach to GNN explanation, enhancing model transparency.
• The method provides deeper insights into GNN decision-making through generative interpolation.
• Gen-GraphEx represents a significant advancement in creating reliable and understandable graph neural network models.