Joint selection of wavenumber regions for MidIR and RAMAN spectra and variables in PLS regression using Genetic Algorithms

Résumé

Many methods exist for feature selection in PLS regression when there are too many variables. Less methods are available for selecting wavenumber regions for MidIR or RAMAN spectra. In this work, PLS has been coupled with genetic algorithms to allow for the selection of intervals in spectra. This work was motivated by a regression issue about transformation of cassava. Those data consist of three tables: RAMAN spectra, MidIR spectra and physico-chemical variables. The purpose is to adapt to this regression context a strategy previously designed to select intervals in NIR spectra in classification. A new algorithm is proposed to fit such multiblock data in PLS1 regression context. Illustrations on simulated data are performed before application to the real dataset.