Title: Bridging Rashomon Theory and PRAXIS: A Streamlined Approach to Decision Tree Rashomon Sets

Abstract:

Conventional machine learning workflows frequently yield numerous models that perform nearly optimally. These collections, known as "Rashomon sets," present both significant hurdles and valuable prospects for decision-making processes that prioritize robustness and uncertainty awareness. They enable the integration of domain expertise and user preferences, which are typically challenging to encode directly into objective functions, while also measuring the variety among acceptable models for a specific dataset and goal. Despite their utility, calculating Rashomon sets—even for straightforward, interpretable structures like sparse decision trees—demands substantial computational power, including vast memory and long execution times.

To address this bottleneck, we introduce PRAXIS, an algorithm designed to approximate these Rashomon sets with dramatic gains in efficiency, reducing both runtime and memory consumption by orders of magnitude. Our validation demonstrates that PRAXIS consistently retrieves nearly the entire full Rashomon set. This capability empowers researchers and practitioners to model Rashomon sets for real-world data at scale. The source code for PRAXIS can be accessed at https://github.com/zakk-h/PRAXIS.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC