Interpret Gaussian Process Models by Using Integrated Gradients.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Statistical Modeling

Background:

• Gaussian Process Regression (GPR) provides predictions and confidence intervals but lacks interpretability, especially for its uncertainty estimates.
• Existing Explainable AI (XAI) methods struggle to interpret the predicted standard deviation in GPR models.
• Deep learning integrations with GPR show promise for accuracy but exacerbate interpretability challenges.

Purpose of the Study:

• To develop a novel method for interpreting GPR predictions, focusing on the uncertainty component.
• To enhance the explainability of GPR models by quantifying feature contributions to predictive uncertainty.
• To improve trust and understanding of GPR models in critical applications through interpretable uncertainty.

Main Methods:

• Incorporation of the Integrated Gradients (IG) method with Gaussian Process Regression (GPR).
• Evaluation of feature importance by assessing the contribution of each explanatory variable to the prediction.
• Analysis of the posterior distribution's standard deviation to decompose predictive uncertainty.

Main Results:

• The proposed IG-based method successfully interprets GPR predictions by attributing uncertainty to individual feature contributions.
• This approach provides a detailed decomposition of predictive uncertainty, highlighting influential variables.
• The methodology quantifies feature-specific uncertainty, offering insights into model reliability.

Conclusions:

• The novel IG-GPR interpretation method enhances understanding of GPR model behavior and predictive uncertainty.
• This technique increases trust in GPR predictions by making the model's decision-making process more transparent.
• The approach is particularly valuable in domains requiring high interpretability alongside predictive accuracy.