Statistics Seminar

Abstract: This talk is about using some new capabilities in the GUIDE regression tree algorithm to solve one new and one old problem. The first one is "Explainable AI".  We start with questions such as what "explainability" means or should mean, and what qualifies as an explainable model. Then we show how to construct a GUIDE regression tree model to approximate any black-box model and to provide an approximate explanation of the latter.  It turns out that the regression tree model often has higher prediction accuracy than the model it approximates. Consequently, it serves as a fully transparent, explainable, and superior replacement for the black-box model.

The second problem is the age-old one of fitting regression models to data with missing values in the predictor variables. "Regression" is considered generally here and includes traditional linear and logistic regression as well as modern machine learning methods. For more than fifty years, the usual practice has been imputation, which requires assumptions that are often neither theoretically justifiable nor practically verifiable. More importantly, missingness itself typically contains useful information that imputation destroys. We show how the GUIDE algorithm can build prediction models without resorting to imputation.

Finally, to connect the two problems, we show how to explain and improve upon AI models built from incomplete data.  The GUIDE software is available from https://pages.stat.wisc.edu/~loh/guide.html.