Quantifying uncertainty in graph neural network explanations.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Graph Neural Networks (GNNs) are increasingly used for complex data analysis.
• Existing GNN explanation methods often overlook uncertainties from data and model parameters, leading to unreliable explanations.
• Quantifying uncertainty in post-hoc, model-agnostic GNN explanations is challenging.

Purpose of the Study:

• To develop a novel framework for uncertainty quantification in GNN explanations.
• To address the limitations of existing methods by considering data and parameter uncertainties.
• To improve the reliability and trustworthiness of GNN predictions.

Main Methods:

• A new uncertainty quantification framework for GNN explanations is proposed.
• The framework accounts for two distinct data uncertainties to assess explanation uncertainty.
• It learns parameter distributions directly from data, quantifying explanation uncertainty without distribution assumptions.

Main Results:

• The proposed framework successfully quantifies uncertainties stemming from graph data and model parameters.
• It integrates seamlessly with existing post-hoc GNN explanation methods.
• Empirical results demonstrate superior GNN explanation performance on real-world benchmarks.

Conclusions:

• The novel framework provides a robust solution for uncertainty quantification in GNN explanations.
• This approach enhances the reliability of GNN predictions by accounting for inherent uncertainties.
• The method sets a new standard for GNN explanation performance and trustworthiness.