A data-driven functional projection approach for the selection of feature ranges in spectra with ICA or cluster analysis

摘要

Prediction problems from spectra are largely encountered in chemometry. Inaddition to accurate predictions, it is often needed to extract informationabout which wavelengths in the spectra contribute in an effective way to thequality of the prediction. This implies to select wavelengths (or wavelengthintervals), a problem associated to variable selection. In this paper, it isshown how this problem may be tackled in the specific case of smooth (forexample infrared) spectra. The functional character of the spectra (theirsmoothness) is taken into account through a functional variable projectionprocedure. Contrarily to standard approaches, the projection is performed on abasis that is driven by the spectra themselves, in order to best fit theircharacteristics. The methodology is illustrated by two examples of functionalprojection, using Independent Component Analysis and functional variableclustering, respectively. The performances on two standard infrared spectrabenchmarks are illustrated.