The article discusses the challenges of using Gaussian processes for spatial data analysis due to their high computational demands. It introduces a novel stochastic gradient Markov chain Monte Carlo (SGMCMC) framework that leverages the Vecchia approximation to enhance computational efficiency. By subsampling locations at each iteration, the method achieves linear time complexity, allowing for scalable estimation. Simulation studies show that SGMCMC competes with leading scalable GP methods on computational speed and accuracy. The approach is further validated through its application to global ocean temperature data, emphasizing its utility in large spatial datasets.
Gaussian processes (GPs) face challenges in scalability due to cubic time complexity in spatial datasets. Our SGMCMC framework provides a solution for efficient computation.
We develop an SGMCMC framework that utilizes a Vecchia-approximated GP likelihood, achieving linear time complexity for scalable estimation of Gaussian processes.
Simulation results indicate that our proposed SGMCMC method competes effectively with existing scalable Gaussian process algorithms in terms of both computational efficiency and parameter accuracy.
Application of the SGMCMC framework to global ocean temperature data illustrates its practicality and effectiveness in handling large-scale spatial analyses.
Collection
[
|
...
]