Bayesian variable selection for Gaussian process regression: Application to chemometric calibration of spectrometers

Gaussian processes have received significant interest for statistical data analysis as a result of the good predictive performance and attractive analytical properties. When developing a Gaussian process regression model with a large number of covariates, the selection of the most informative variables is desired in terms of improved interpretability and prediction accuracy. This paper proposes a Bayesian method, implemented through the Markov chain Monte Carlo sampling, for variable selection. The methodology presented is applied to the chemometric calibration of near infrared spectrometers, and enhanced predictive performance and model interpretation are achieved when compared with benchmark regression method of partial least squares.