Learning spatial patterns with variational Gaussian processes: Regression

• Published in: Computers & geosciences Vol. 161; p. 105056
• Main Authors: Gonçalves, Ítalo Gomes, Guadagnin, Felipe, Cordova, Diogo Peixoto
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.04.2022
• Subjects: Gaussian process; Kriging; Machine learning; Variational inference; Gaussian process; Variational inference; Kriging; Machine learning
• ISSN: 0098-3004; 1873-7803
• DOI: 10.1016/j.cageo.2022.105056

A variational Gaussian process (VGP) model specialized in spatial data is introduced, capitalizing on recent advances in the machine learning field. The model is modular and customizable, being capable of handling different assumptions about the data. This work focusses on multivariate robust regression, using an adaptation of the ε-insensitive loss function. The VGP possibilitates end-to-end modeling: normal score transform, detection of the spatial pattern, and prediction. A methodology to deal with large datasets is presented. An open-source implementation is available. •A variational Gaussian process model specialized in spatial data.•Stochastic variational inference is capable of handling large datasets.•Different models available as combinations of latent variables and likelihoods.•Robust multivariate regression with non-Gaussian marginals.