Random Forests in Spatial Mixed Models

• Activity Number: 106 - AI and Deep Models for Spatial and Spatio-Temporal Data
• Type: Invited
• Date/Time: Monday, August 3, 2020 : 1:00 PM to 2:50 PM
• Sponsor: Section on Statistics and the Environment
• Author(s): Abhi Datta* and Arkajyoti Saha and Sumanta Basu
• Companies: Johns Hopkins University and Johns Hopkins University and Cornell University
• Keywords: Random forest; Geostatistics; Gaussian Process; Spatial statistics

Abstract

Machine learning approaches like CART and random forests have become increasingly popular. However, current applications of random forests in geo-statistics either naively ignores the spatial correlation among the observations, or uses pairwise distances between locations as additional covariates. We discuss some limitations of these approaches and propose a new random forest approach that can replace the fixed-effects regression component in a spatial linear mixed model. This new random forest algorithm adjusts for spatial correlation among the observations and can also be used for other correlated data like time-series. In absence of spatial correlation in the data, our algorithm becomes identical to the classical random forest algorithm. We offer theoretical results to understand the asymptotic properties of our algorithm and present data analyses demonstrating the benefits of our method over the existing random-forest based approaches for spatial data.