Beyond Point Prediction: Artificial Representative Trees with Uncertainty

Data and Models for Reproducibility This repository contains the models and results from the paper: Kronziel et al. (2026) “Beyond Point Prediction: Artificial Representative Trees with Uncertainty” The upload includes three zipped data collections corresponding to: the simulation study the benchmark experiment the application to the NHANES dataset Together with the R code provided in the associated GitHub repository (https://github.com/imbs-hl/ART_uncertainty_paper) all figures from the publication can be fully reproduced. In addition, users can simulate their own data and run custom experiments. Contents of the Zipped Files Each of the three zipped archives contains the following components: Regression modelsFiles with the prefix regression_trees contain the fitted regression prediction models, including Artificial Representative Trees (ARTs) and Decision Trees (DTs) Probability modelsFiles with the prefix probability_trees contain the fitted probability prediction models (ARTs and DTs). For the simulation study and the benchmark experiment, probabilities are predicted with respect to a threshold of 0.5 (mean value). For the NHANES application, two separate files are provided: one for predicting probabilities above the prediabetes threshold (glycohemoglobin ≥ 5.7) one for predicting probabilities above the diabetes threshold (glycohemoglobin ≥ 6.5) ResultsFiles with the prefix results contain data frames with the computed performance metrics. Predicted probabilitiesFiles with the prefix pred_probabilities contain the predicted probabilities per terminal node for ARTs and DTs using conformal predictive systems. Variable usage of random forest (RF)For the NHANES application, an additional CSV file is provided reporting, for each variable in the dataset, how often it was used as a split variable in the corresponding RF (expressed as relative frequency). Reproducibility All results presented in the paper can be reproduced using: the data and model objects provided here, and the R code from the GitHub repository. The setup also enables users to: generate new simulated datasets apply the methods to custom data conduct additional experiments