From Local Explanations to Global Understanding with Explainable AI for Trees.

Area of Science:

• Machine Learning
• Data Science
• Medical Informatics

Background:

• Tree-based models (random forests, decision trees, gradient boosted trees) are widely used for non-linear predictions.
• Interpretability of these powerful models remains a significant challenge.

Purpose of the Study:

• To enhance the interpretability of tree-based machine learning models.
• To introduce novel methods for explaining model predictions and understanding model structure.

Main Methods:

• Developed a polynomial time algorithm for optimal explanations using game theory.
• Introduced a new explanation type to measure local feature interaction effects.
• Created tools to understand global model structure by aggregating local explanations.

Main Results:

• Applied the tools to three medical machine learning problems.
• Demonstrated that combining local explanations reveals global structure while maintaining local faithfulness.
• Identified non-linear mortality risk factors, distinct population subgroups, and chronic kidney disease risk interactions.

Conclusions:

• The developed tools significantly improve the interpretability of tree-based models.
• These methods enable identification of complex patterns and monitoring of deployed models in healthcare.
• Enhanced interpretability has broad implications across various domains utilizing tree-based models.