A0687
Title: Bayesian analysis for functional ANOVA model
Authors: Yongdai Kim - Seoul National University (Korea, South) [presenting]
Abstract: The functional ANOVA model is a useful tool for constructing an interpretable prediction model. While there are several frequentist procedures to estimate the components in the functional ANOVA model (i.e. MARS, Splines), Bayesian procedures focus mainly on the generalized additive model (GAM), which is the simplest one among functional ANOVA models due to computational difficulties. A computationally efficient Bayesian procedure is proposed to infer components in the functional ANOVA model. The algorithm, called ANOVA-BART, is a modification of BART (Bayesian Additive Regression Tree), a well-known Bayesian procedure for estimating high-dimensional regression models. BART combines many baseline trees (simple trees) to infer the final prediction model. Even though BART is very good at prediction, its interpretation is not easy. To improve the interpretability of BART, the sets of baseline trees are constructed corresponding to each component of the functional ANOVA model, put prior to each set of baseline trees, and an MCMC algorithm is developed to search good linear combinations of baseline trees for each component.
