Application of latent projective graph in variable selection for near infrared spectral analysis

Latent projective graph (LPG) is a technique developed in chemical factor analysis (CFA) for investigating the nature of hyphenated data. Selective variables can be found because collinear variables present a straight line in an LPG. Variable selection in near infrared (NIR) spectral analysis has been a notoriously difficult task for improving the quality of the models, the aim of which is to find informative variables specific to the target component. In this work, based on the assumption that collinear wavelengths in the calibration spectra may have the same contribution to the modeling, LPG was adopted for variable selection in NIR spectral analysis. The variables located at the inflections of an LPG are found to be informative for the quantitative models. With three NIR datasets of pharmaceutical tablets, blood and plant samples, it was proved that a very parsimonious model can be built by using only several selected variables. Compared with the previous work, the method provides a simple way for variable selection.

► Latent projective graph (LPG) is used as a tool for variable selection.
► Informative variables in near infrared (NIR) spectra were found by using LPG.
► Parsimonious models were achieved for three NIR datasets.