Hierarchical mixture of linear regressions for multivariate spectroscopic calibration

This paper investigates the use of the hierarchical mixture of linear regressions (HMLR) and variational inference for multivariate spectroscopic calibration. The performance of HMLR is compared to the classical methods: partial least squares regression (PLSR), and PLS embedded locally weighted regression (LWR) on three different NIR datasets, including a publicly accessible one. In these tests, HMLR outperformed the other two benchmark methods. Compared to LWR, HMLR is parametric, which makes it interpretable and easy to use. In addition, HMLR provides a novel calibration scheme to build a two-tier PLS regression model automatically. This is especially useful when the investigated constituent covers a large range.