Exploring Model Diversity in Decision-Tree Rashomon Sets Beyond Discretization

The Rashomon effect---the existence of many models with nearly equivalent predictive performance---has important implications for robustness, multi-objective optimization, feature importance, and model customizability. Studying these effects requires characterizing the Rashomon set: the collection of models whose regularized loss is near-optimal. One model class often studied in this way is sparse decision trees, for which modern methods support enumeration. However, existing methods can fully enumerate this set only after continuous features have been binarized, which can exclude high-performing models and lead to incomplete analyses of predictive multiplicity and feature importance. We introduce ArborEnum, the first framework for enumerating decision-tree Rashomon sets directly over continuous features. ArborEnum supports exact enumeration, an efficient approximate variant, and an anytime procedure that progressively refines the set of candidate cut points and converges to the full continuous-feature Rashomon set. Across datasets, our approximations recover nearly all trees at a fraction of the computational cost, while our anytime results show that coarse binarization can obscure diversity in features and predictive multiplicity relevant to robustness analysis.