From Rashomon Theory to PRAXIS: Efficient Decision Tree Rashomon Sets

Researchers introduce PRAXIS, an algorithm that efficiently computes Rashomon sets—collections of near-optimal machine learning models—achieving orders of magnitude improvements in runtime and memory usage compared to existing methods. The breakthrough enables practitioners to scalably explore model diversity and incorporate domain knowledge into decision-making for interpretable models like decision trees.

PRAXIS addresses a fundamental computational bottleneck in modern machine learning: understanding and leveraging the multiplicity of near-optimal models that emerge from standard training procedures. Rashomon sets represent the ensemble of models performing nearly identically on validation metrics, yet diverging in their predictions and internal logic. This diversity matters because it quantifies model uncertainty and enables practitioners to select among valid alternatives based on domain-specific constraints rather than optimizing a single objective function.

The research builds on growing recognition that machine learning pipelines are underdetermined—many models fit data equally well. Previous approaches to computing Rashomon sets required prohibitive computational resources, limiting their practical application to toy problems. PRAXIS's efficiency gains stem from algorithmic innovations that approximate rather than exhaustively enumerate these sets, recovering nearly complete coverage while dramatically reducing computational demands.

For practitioners and researchers, this advancement unlocks previously inaccessible applications in high-stakes domains like healthcare, finance, and policy where interpretability and robustness matter alongside accuracy. Decision trees represent an ideal use case due to their interpretability, making Rashomon set exploration directly actionable for domain experts. The released code democratizes access to these tools, potentially accelerating adoption across industries.

Looking forward, the work positions uncertainty quantification through model diversity as a central consideration in AI development. As regulatory frameworks increasingly demand explainability and robustness, tools enabling efficient exploration of valid model alternatives become more valuable. Extensions to other model classes and deeper investigation of how to leverage Rashomon sets for improved out-of-distribution robustness represent key research directions.

• →PRAXIS achieves orders-of-magnitude improvements in computing Rashomon sets with drastically reduced memory and runtime requirements.
• →The algorithm enables practitioners to explore multiple near-optimal models and incorporate domain knowledge into decision-making processes.
• →Rashomon sets quantify model uncertainty and diversity, valuable for high-stakes applications requiring interpretability and robustness.
• →Released open-source code makes efficient Rashomon set computation accessible to researchers and industry practitioners.
• →The work addresses growing recognition that machine learning models are underdetermined, with many valid solutions fitting equally well.