Visualizable and Interpretable Regression Tree Models

Abstract

There are many regression methods that can fit models with high prediction accuracy. But, few of these models can be interpreted or visualized. In this talk, we present a tree-structured method that fits a relatively simple linear model to each partition of the variable space. This permits the fitted regression surface to be visualized with graphs or contour plots. Our tree models have negligible variable selection bias and possess good prediction accuracy, as demonstrated in an empirical study comparing the performance of twenty-four methods on fifty-two real datasets.