Robert E. McCulloch
University of
Texas at Austin
BART: Extensions and Interpretations
Authors: Hugh A. Chipman (Acadia University), Edward I. George
(University of Pennsylvania), and Robert E. McCulloch (University of
Texas at Austin)
In BART: Bayesian Additive Regression Trees, we developed a general
approach for fitting the model y = f(x) + e with minimimal assumptions
about the structure of f. The function f is modeled as a sum of
simple tree models. This sum is made meaningful by a Bayesian analysis
with a simple yet powerful prior specification that provides for both
overall regularization of the fit and the limiting of the contribution
of each individual tree to the sum. An MCMC algorithm provides
inference. The BART model is shown to have excellent out of sample
predictive performance.
Users however, often want more than just good predictions. In this
talk we discuss ways to interpret the BART fit as well as ways to
inject non-generic prior information. For example, users often want to
know which variables are most important. How can this information be
gleaned from the BART inference?