Why did AI get this one wrong? - Tree-based explanations of machine learning model predictions.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Explainable AI (XAI)

Background:

• Complex machine learning (ML) models, including boosting, bagging, and deep learning, achieve high accuracy but suffer from poor interpretability, creating a performance-intelligibility trade-off.
• This trade-off is particularly critical in high-stakes fields like medicine, where understanding model predictions is essential for trust and adoption.

Purpose of the Study:

• To propose a novel methodological approach for generating post-hoc explanations for predictions made by generic machine learning models.
• Introduce AraucanaXAI, a new explainable AI (XAI) method designed to enhance the intelligibility of black-box ML models.

Main Methods:

• Developed AraucanaXAI, a method utilizing surrogate, locally-fitted classification and regression trees.
• Applied the method to provide post-hoc explanations for specific ML model predictions on given instances.
• Evaluated the method's behavior in medical AI scenarios, including model-physician disagreement and data scarcity.

Main Results:

• AraucanaXAI demonstrates superior fidelity to the original ML model.
• The method effectively handles non-linear decision boundaries.
• Provides native support for both classification and regression tasks.
• Successfully evaluated in medical AI contexts, addressing challenges like prediction uncertainty.

Conclusions:

• AraucanaXAI offers a robust solution to the interpretability challenge in complex ML models.
• The method enhances trust and utility of AI in critical applications like medicine.
• An open-source implementation is available for broader adoption and research.