Online Program Home
My Program

Abstract Details

Activity Number: 308 - Recent Advancements in Spatial and Spatio-Temporal Modeling
Type: Contributed
Date/Time: Tuesday, July 30, 2019 : 8:30 AM to 10:20 AM
Sponsor: Section on Statistics and the Environment
Abstract #305345 Presentation 1 Presentation 2
Title: Bayesian Selection of Tuning Parameters
Author(s): Nathan Wikle* and Ephraim Hanks
Companies: Pennsylvania State University and Pennsylvania State University
Keywords: Bayesian regularization; importance sampling; stochastic optimization; spatially varying coefficient models; lake nutrient prediction

Predictive models have long been known to benefit from regularization, where a regularization term is chosen to penalize model complexity and prevent overfitting (e.g., ridge regression, LASSO, etc.). In a Bayesian setting, such penalty terms correspond to the statistician’s choice of prior and/or hyperparameters. However, fitting Bayesian models often requires nontrivial computational resources, making both cross-validation and exhaustive grid search approaches for choosing hyperparameters undesirable. In this project, we explore how sequential importance sampling and stochastic optimization methodologies can be combined to efficiently identify hyperparameters that minimize out-of-sample prediction error for large Bayesian models. The success of this approach is demonstrated with a spatially varying coefficient (SVC) model for lake nutrient data, fitted to 10,000 lake observations from the LAGOS Northeast data set. We show that hyperparameter choice can have a dramatic impact on the predictive accuracy of SVC models, and that our methodology can be utilized to reduce computational time while simultaneously improving out-of-sample prediction.

Authors who are presenting talks have a * after their name.

Back to the full JSM 2019 program