An improved boosting partial least squares method for near-infrared spectroscopic quantitative analysis

Abstract

Boosting partial least squares (PLS) has been used for regression to improve the predictive accuracy of PLS models, however, there are still problems when the outliers exist in the calibration dataset. To make the method robust and enhance its prediction ability, an improved boosting PLS is proposed and applied in quantitative analysis of near-infrared (NIR) spectral datasets. In the method, a robust step is added to weaken the effect of the outliers on the model. On the other hand, the loss function defined with relative errors is suggested for updating the sampling weight during the boosting procedure. In addition, the ensemble prediction by the weighted mean of the models in the boosting series is found to be more effective than the commonly used weighted median. The performance of the improved method is tested with two large NIR datasets of industrial production. The method was found to have a marked superiority in robustness and prediction ability, particularly when outliers exist.