Variable selection in near-infrared spectroscopy: Benchmarking of feature selection methods on biodiesel data

• Published in: Analytica chimica acta 2011-04, Vol.692 (1), p.63-72
• Main Authors: Balabin, Roman M., Smirnov, Sergey V.
• Format: Article
• Language: English
• Subjects: Algorithms; Artificial neural networks; Benchmarking; Biofuel (ethanol-gasoline fuel, vegetable oil, biodiesel); Biofuels - analysis; Dietary Fats, Unsaturated - analysis; Discriminant Analysis; Least squares method; Least-Squares Analysis; Linear Models; Linear regression (projection to latent structures); Mathematical models; Neural Networks (Computer); Petroleum (crude oil); Projection; Quality control (process analytical chemistry, process analytical technology); Regression; Spectrophotometry, Infrared - methods; Spectrophotometry, Infrared - standards; Spectroscopy; Spectroscopy, Fourier Transform Infrared; Spectrum Analysis, Raman; Support vector machines; Vibration; Wavelengths
• ISSN: 0003-2670; 1873-4324
• DOI: 10.1016/j.aca.2011.03.006

During the past several years, near-infrared (near-IR/NIR) spectroscopy has increasingly been adopted as an analytical tool in various fields from petroleum to biomedical sectors. The NIR spectrum (above 4000 cm −1) of a sample is typically measured by modern instruments at a few hundred of wavelengths. Recently, considerable effort has been directed towards developing procedures to identify variables (wavelengths) that contribute useful information. Variable selection (VS) or feature selection, also called frequency selection or wavelength selection, is a critical step in data analysis for vibrational spectroscopy (infrared, Raman, or NIRS). In this paper, we compare the performance of 16 different feature selection methods for the prediction of properties of biodiesel fuel, including density, viscosity, methanol content, and water concentration. The feature selection algorithms tested include stepwise multiple linear regression (MLR-step), interval partial least squares regression (iPLS), backward iPLS (BiPLS), forward iPLS (FiPLS), moving window partial least squares regression (MWPLS), (modified) changeable size moving window partial least squares (CSMWPLS/MCSMWPLSR), searching combination moving window partial least squares (SCMWPLS), successive projections algorithm (SPA), uninformative variable elimination (UVE, including UVE-SPA), simulated annealing (SA), back-propagation artificial neural networks (BP-ANN), Kohonen artificial neural network (K-ANN), and genetic algorithms (GAs, including GA-iPLS). Two linear techniques for calibration model building, namely multiple linear regression (MLR) and partial least squares regression/projection to latent structures (PLS/PLSR), are used for the evaluation of biofuel properties. A comparison with a non-linear calibration model, artificial neural networks (ANN-MLP), is also provided. Discussion of gasoline, ethanol-gasoline (bioethanol), and diesel fuel data is presented. The results of other spectroscopic techniques application, such as Raman, ultraviolet-visible (UV-vis), or nuclear magnetic resonance (NMR) spectroscopies, can be greatly improved by an appropriate feature selection choice.