Large-Scale Gaussian Processes for Spatiotemporal Modeling of Disease Incidence

• Activity Number: 351
• Type: Topic Contributed
• Date/Time: Tuesday, August 11, 2015 : 10:30 AM to 12:20 PM
• Sponsor: Government Statistics Section
• Author(s): Seth Flaxman*
• Keywords: spatial ; time series ; machine learning ; Gaussian Processes ; non-parametric regression

Abstract

Scaling up small area estimation using Gaussian process-based models faces a challenging computational bottleneck. I will review recent statistical machine learning work on efficient, and in some cases exact, inference for large-scale Gaussian process regression. I will show an application to jointly modeling a century of weekly reports of disease incidence by US state compiled by Project Tycho (www.tycho.pitt.edu), comparing exact Kronecker-based inference with a Gaussian observation model to Kronecker-based inference with a negative binomial observation model and the Laplace approximation. I focus on learning novel covariance functions, uncovering spatiotemporal patterns, and enabling accurate spatiotemporal prediction and forecasting. I will briefly discuss the implementation of these methods in GPML (gaussianprocess.org/gpml/code), a popular machine learning Gaussian process toolkit.