Abstract
Variable selection provides useful information about the most important predictors in the dataset, information which is not always available at the beginning of an analysis. Two recent variable selection methods, backward variable selection for partial least squares (BVSPLS) and powered partial least squares (PPLS), were compared against each other and against forward stepwise selection (FSS) and full spectrum partial least squares (PLS) in terms of their ability to produce accurate prediction models in NIR spectroscopy data. All four regression methods were studied using three different NIR datasets. PPLS and BVSPLS gave good prediction results in all three datasets, even with a very limited number of calibration samples available (<40). All methods gave similar prediction results when the number of calibration samples was higher (>150). PPLS gave the best predictive performance of all methods and also gave the selections of variables that were most easily assigned to specific chemical bonds. Hence, the PPLS models were more easily interpretable than the other models. This study quantifies differences between the two recent variable selection methods as well as the differences between recent methods and more established methods. Moreover, if the number of calibration samples can be reduced through variable selection, the labour and cost associated with wet chemistry reference methods can be reduced accordingly.
© 2012 IM Publications LLP
PDF Article
More Like This
Cited By
You do not have subscription access to this journal. Cited by links are available to subscribers only. You may subscribe either as an Optica member, or as an authorized user of your institution.
Contact your librarian or system administrator
or
Login to access Optica Member Subscription