Ron Vatenstein

“That decision trees guy”

Decision Tree ensembles such as Random Forest and XGBoost are powerful models widely used for classification and regression across many domains. Beyond accurate prediction, it is often essential to understand why a model reached its conclusion. We want to explain the model and understand which features played the most significant role in its decision. This goal is the main topic of my research. 

Shapley values are the most popular feature importance approach for decision trees. As part of my PhD, I developed a new and faster algorithm for computing these values. The work treats the tree, the data we want to explain its predictions and other inputs as a pseudo-Boolean function in Weighted Disjunctive Normal Form (WDNF) and uses this new representation to compute Shapley values efficiently. We call this new algorithm WOODELF. 

WOODELF is fast due to improved complexity and heavy use of vectorized operations, GPU-friendly, and implemented in pure Python (with most of the heavy lifting done by NumPy and SciPy). It’s also generic: in addition to Shapley values, we also compute interaction values and Banzhaf values.

Our code is available in the open source package ‘woodelf’: https://github.com/ron-wettenstein/woodelf

Our paper introducing WOODELF, From Decision Trees to Boolean Logic: A Fast and Unified SHAP Algorithm was accepted for an oral presentation at AAAI 2026.

More papers are coming soon, stay tuned!