Partial Least Square Regression with Variable-Bagged Model Ensemble

Abstract

A new bagging strategy, called variable-bagging, for partial least square (PLS) modelling is proposed. Variable-bagging is to bootstrap a set of spectral variables for each modelling process. Namely, Variable-bagged PLS (VBPLS) can be viewed as an ensemble of multiple PLS models built with bootstrapped molecular structure features of the analytes, for example, wavenumber. It can condense information from spectral variables broadly distributed in the wavenumber dimension and truncate the contribution of non-informative variables such as interference or simple noise. Thus, it potentially has an advantage to yield superior regression performance to conventional PLS which is based on a single model. The performance of VBPLS is demonstrated for the determination of Brix values with a set of NIR spectra collected from apples. The comparisons with other PLS techniques reveal that VBPLS yields superior regression performance, in terms of both prediction error and robustness, to conventional PLS, even with the optimal number of latent variables. Namely, VBPLS models can describe more reasonable relationships between the NIR spectra and corresponding Brix values.