Visualization of BART and BARP using SHAP.
bartXViz: Visualization of BART and BARP using SHAP
The bartXViz package provides SHAP-based model explanation tools for Bayesian Additive Regression Trees (BART) and Bayesian Additive Regression Trees with Post-Stratification (BARP).
The version uploaded on January 26, 2026 corresponds to v1.0.11, the same version currently released on CRAN. (https://CRAN.R-project.org/package=bartXViz)
Complex machine learning models are often difficult to interpret.
Shapley values provide a principled framework for understanding why a model makes specific predictions by quantifying each variable's contribution.
This package implements permutation-based Shapley values for BART and BARP models, enabling users to evaluate variable importance and contribution across Bayesian posterior samples obtained through MCMC.
The SHAP approach follows the method proposed by Strumbel and Kononenko (2014)doi:10.1007/s10115-013-0679-x, adapted to the Bayesian tree ensemble framework introduced by Chipman, George, and McCulloch (2010)doi:10.1214/09-AOAS285.
bartXViz is compatible with several popular R implementations of BART, including:
- BARTdoi:10.18637/jss.v097.i01
- bartMachinedoi:10.18637/jss.v070.i04
- dbarts (https://CRAN.R-project.org/package=dbarts)
For gradient boosting and baseline comparisons, the package also considers the SHAP framework proposed by Lundberg et al. (2020)doi:10.1038/s42256-019-0138-9.
The BARP model, originally proposed by Bisbee (2019)doi:10.1017/S0003055419000480, is implemented with reference to jbisbee1/BARP.
BARP extends post-stratification to compute variable contributions within each stratum defined by stratifying variables, improving small-area estimation interpretability.
The resulting Shapley values can be visualized through both global and local explanation methods, allowing users to explore model interpretability in Bayesian tree ensembles with intuitive visualizations.