UbiQTree: Uncertainty Quantification in XAI with Tree Ensembles

SHAP explanations need uncertainty estimates: Current explainability tools like SHAP provide single-number explanations without showing how confident we should be in those explanations. This paper demonstrates that adding uncertainty quantification makes these explanations more reliable for high-stakes decisions in healthcare and similar domains.

High importance doesn't mean high reliability: A key finding from the research is that "features with the highest SHAP values are not necessarily the most stable," meaning an AI system might identify a feature as very important for a decision, but that importance rating could be unreliable due to data limitations or noise.

Two types of uncertainty matter differently: The paper distinguishes between aleatoric uncertainty (irreducible noise in the data itself) and epistemic uncertainty (gaps from insufficient data). Understanding which type affects an explanation helps organizations know whether they need better data or should simply accept inherent unpredictability.

Responsible AI requires transparency about confidence: By decomposing uncertainty in AI explanations, this work supports more honest communication with stakeholders about what models actually know versus what they're uncertain about—critical for building appropriate trust in high-stakes applications like healthcare analytics.

Better explanations enable better decisions: The research suggests that comprehensive uncertainty information in model explanations can "guide the development of robust decision-making processes," helping organizations make safer choices about when and how to rely on AI recommendations.