New Algorithm Enumerates Decision Tree Rashomon Sets Efficiently

The "Rashomon effect" highlights that multiple models can achieve similar performance on a given task, impacting robustness and interpretability. For decision trees, enumerating these "Rashomon sets" – collections of nearly optimal models – has been challenging, often requiring data binarization which can limit the scope of analysis. A new algorithm, ArborEnum, addresses this by directly enumerating decision-tree Rashomon sets for continuous features. This innovation removes the need for data binarization, which previously restricted the types of splits trees could make or increased computational complexity. The algorithm demonstrates orders-of-magnitude speedups over existing methods. It also includes a relaxation for approximate enumeration and an anytime algorithm that refines candidate thresholds, providing increasingly detailed approximations that converge to the true continuous-feature Rashomon set. This advancement helps uncover more diverse and important features that coarse binarization might miss.