Statistics Seminar

Abstract: Machine learning regression techniques like deep learning and ensembles of trees are the best currently known for out-of-sample predictive performance.  However, these methods can be viewed as black-box models, i.e., a vast number of parameters and details so complex that their meaning can only be gleaned from predictions.  This has sparked a lot of interest in attempts to explain these predictions via marginal effects.  In this talk, I will focus on two popular marginal effect methods: Friedman's partial dependence function and Shapley values.  These approaches are widely applicable to machine learning/nonparametric regression.  Here, I will focus on one particular method, Bayesian Additive Regression Trees, but these results likely hold in the wider genre.