Accurate Shapley Values for explaining tree-based models

Shapley Values, a popular tool for explaining AI decisions, can produce inaccurate results depending on how data is encoded — the paper demonstrates that categorical variables (like "red/blue" or "yes/no") are particularly sensitive to encoding choices, meaning the same information can yield different explanations based on technical implementation rather than actual model behavior.

Current methods for computing Shapley Values in tree-based models (like random forests) are inefficient and less accurate than necessary — the researchers developed two new estimators that better exploit the structure of these models, providing more reliable explanations with practical improvements over existing approaches.

Even improved Shapley Value methods have inherent limitations for local explanations — the paper acknowledges that while their approach is more accurate, Shapley Values themselves have fundamental constraints as tools for explaining individual predictions, suggesting users should be cautious about over-relying on them as complete explanations.

Technical choices in AI explanation tools can significantly impact the explanations users receive — this research highlights that how developers implement explanation methods matters greatly; two systems using "Shapley Values" might give different answers to the same question, potentially misleading stakeholders about why an AI made a particular decision.