Parallel cross-validation: a scalable fitting method for Gaussian process models

TL;DR

This paper introduces a parallel cross-validation method for Gaussian process models that significantly reduces computation time, enabling analysis of large spatial datasets with high-performance computing resources.

Contribution

The paper presents a novel parallelizable cross-validation approach for Gaussian process parameter estimation and prediction, improving scalability and efficiency over existing methods.

Findings

Achieves accurate covariance parameter estimates and predictions.

Demonstrates excellent weak and strong parallel scaling properties.

Fitted a 5-million-observation canopy height dataset in under 1.5 minutes using 512 cores.

Abstract

Gaussian process (GP) models are widely used to analyze spatially referenced data and to predict values at locations without observations. In contrast to many algorithmic procedures, GP models are based on a statistical framework, which enables uncertainty quantification of the model structure and predictions. Both the evaluation of the likelihood and the prediction involve solving linear systems. Hence, the computational costs are large and limit the amount of data that can be handled. While there are many approximation strategies that lower the computational cost of GP models, they often provide only sub-optimal support for the parallel computing capabilities of current (high-performance) computing environments. We aim at bridging this gap with a parameter estimation and prediction method that is designed to be parallelizable. More precisely, we divide the spatial domain into overlapping subsets and use cross-validation (CV) to estimate the covariance parameters in parallel. We present simulation studies, which assess the accuracy of the parameter estimates and predictions. Moreover, we show that our implementation has good weak and strong parallel scaling properties. For illustration, we fit an exponential covariance model to a scientifically relevant canopy height dataset with 5 million observations. Using 512 processor cores in parallel brings the evaluation time of one covariance parameter configuration to less than 1.5 minutes. The parallel CV method can be easily extended to include approximate likelihood methods, multivariate and spatio-temporal data, as well as non-stationary covariance models.