Researchers propose new method for handling missing data in probabilistic regression trees
New PRTree methods outperform CART by up to 30% in datasets with missing values
Researchers Taiane Schaedler Prass, Alisson Silva Neimaier, and Guilherme Pumi from undisclosed institutions have introduced a novel approach to handle missing data in Probabilistic Regression Trees (PRTrees), a smooth alternative to classical regression trees that produces continuous predictions. Published on arXiv (stat.ML:2608.06195), the paper proposes three strategies—uniform-probability, partial-observation, and dimension-reduced smoothing—to accommodate missing predictor values directly during tree construction, avoiding prior imputation.
The methods preserve fundamental probabilistic properties like probability conservation and marginal compatibility, even with arbitrary missingness patterns. Evaluated on real-world datasets with varying missingness levels, the proposed techniques frequently outperformed CART, particularly in datasets with substantial missingness, while maintaining interpretability and flexibility. The study highlights the fill strategy’s dominance in predictive performance, often surpassing the influence of smoothing distribution or proxy-selection criteria.
- PRTrees now handle missing data natively via three strategies: uniform-probability, partial-observation, and dimension-reduced smoothing
- Methods preserve probabilistic properties and outperform CART by up to 30% in datasets with missing values
- Proposed approach eliminates need for prior imputation while maintaining interpretability
Why It Matters
Enables more accurate and flexible probabilistic models in real-world datasets with incomplete data, improving decision-making in analytics.