Subspace Gaussian process regression model for ensemble nonlinear multivariate spectroscopic calibration

In this paper, we have proposed a set of ensemble Gaussian process regression (GPR) models for nonlinear spectroscopic calibration. Based upon the random subspace modeling method, the new subspace GPR model constructs several subspaces along those directions determined by principal component analysis of the spectral data. Unlike the random subspace method in which the subspaces are constructed through a random manner, the new method determines the subspaces through uncorrelated directions, which could improve both robustness and diversity for the ensemble model. Several comparative studies are carried out among the basic GPR model, the random subspace GPR and the new developed subspace GPR model. It is demonstrated that both of the prediction accuracy and robustness have been improved by the new subspace GPR model.

Automated summary

This paper proposes a set of ensemble Gaussian process regression (GPR) models for nonlinear spectroscopic calibration. The new subspace GPR model constructs multiple subspaces along uncorrelated directions to improve the robustness and diversity of the ensemble model. Comparative studies show that the new subspace GPR model improves both prediction accuracy and robustness.