UbiQTree: Uncertainty quantification in XAI with tree ensembles.

Area of Science:

• Artificial Intelligence
• Machine Learning Interpretability
• Computational Statistics

Background:

• Explainable AI (XAI) is vital for understanding complex models, especially ensemble tree-based methods.
• Shapley additive explanations (SHAP) values are commonly used but often lack uncertainty quantification.
• Existing SHAP interpretations treat values as point estimates, ignoring aleatoric and epistemic uncertainties.

Purpose of the Study:

• To develop a method for decomposing SHAP value uncertainty into aleatoric, epistemic, and entanglement components.
• To enhance the reliability and interpretability of SHAP attributions in critical applications like healthcare.
• To provide insights into managing and reducing uncertainty in AI model explanations.

Main Methods:

• Utilized Dempster-Shafer evidence theory for uncertainty decomposition.
• Employed Dirichlet process (DP) hypothesis sampling over tree ensembles.
• Validated the approach through use-case analysis to reveal epistemic uncertainty insights.

Main Results:

• Successfully decomposed SHAP value uncertainty into distinct components.
• Demonstrated the presence and impact of epistemic uncertainty in SHAP explanations.
• Showcased how tree-based models, particularly bagging, can quantify these uncertainties.

Conclusions:

• Quantifying SHAP value uncertainty is essential for robust AI decision-making.
• Reducing epistemic uncertainty necessitates improvements in data quality and model development.
• The proposed method enhances trust and reliability in AI explanations for critical domains.