Interpretable Perturbator for Variable Selection in near-Infrared Spectral Analysis

A perturbator was developed for variable selection in
near-infrared
(NIR) spectral analysis based on the perturbation strategy in deep
learning for developing interpretation methods. A deep learning predictor
was first constructed to predict the targets from the spectra in the
training set. Then, taking the output of the predictor as a reference,
the perturbator was trained to derive the perturbation-positive (P+) and perturbation-negative (P–) features
from the spectra. Therefore, the weight (σ) of the perturbator
layer can be a criterion to evaluate the importance of the variables
in the spectra. Ranking the spectral variables by the criterion, the
number of the variables used in the quantitative model can be obtained
through cross-validation. Three NIR data sets were used to evaluate
the proposed method. The root mean squared error was found to be comparable
with or superior to that obtained by the commonly used methods. Moreover,
the selected spectral variables are interpretable in identifying the
key spectral features related to the prediction target. Therefore,
the proposed method provides not only an effective tool for optimizing
quantitative model, but also an efficient way for explaining spectra
of multicomponent samples.