Open bonStats opened 3 years ago
bartMachine currently has built in covariate and interaction importance: https://github.com/kapelner/bartMachine/blob/master/bartMachine/R/bart_package_variable_selection.R based on https://arxiv.org/abs/1310.4887
https://arxiv.org/abs/2108.04310 interaction visualisations
https://arxiv.org/abs/2210.11391 more variable importance work
See also Alan's new {bartMan} package (especially the variable importance section a the the end of the vignette) as well as the 2021 paper from the same author "Visualizations for Bayesian Additive Regression Trees" that also covers variable importance metrics.
Could potentially use Sobol indices: https://arxiv.org/abs/2005.13622
There is an implementation of this method here: https://bitbucket.org/mpratola/openbt/src/master/
https://bitbucket.org/mpratola/openbt/src/master/src/sobol.cpp
This method does not require refitting the BART mdoels