Bayesian Additive Regression Trees using the bartMachine package

By: John Snyder

Bayesian Additive Regression Trees, or BART, is a highly flexible machine learning approach which is gaining popularity in recent years. In additional to being high performing, it has numerous advantages as a result of being a fully developed Bayesian technique such as obtaining credible intervals along with estimates, performing model free variable selection, and others. The bartMachine R package is a relatively recent(2016) implementation which adds many computational and practical quality of life improvements over the BART author's release.

This month, I will first provide a brief and high level overview of the BART structure and how the model is fit. I will then thoroughly discuss and demonstrate the practical usage of the package applied to publicly available datasets. Finally, I will demonstrate and show the results of a simulation demonstrating BART's consistently superior predictive performance over many popular machine learning techniques such as random forests and gradient boosting.

The meetup will be in The Showroom in the CIC building at 20 South Sarah Street, St. Louis, MO 63108. The building entrance is at the corner of Forest Park Avenue and Sarah Street. You can find directions at http://stl.cic.us/directions/ (CIC@CET).

We will meet for snacks, set-up, and conversation at 6:00PM. The presentation will start at 6:30PM and will be about 60 minutes long.