Application of a hybrid variable selection method for the classification of rapeseed oils based on 1H NMR spectral analysis

Abstract

1H nuclear magnetic resonance (NMR) spectroscopy was utilized to distinguish the brands of rapeseed oils. As there are more than four hundreds of NMR variables which can cause the discrimination model redundancy, it is necessary to do effective variable selection. Successive projections algorithm (SPA) executed on the full spectrum only improved a few correct answer rate (CAR) and Cohen’s kappa coefficient (K) compared to full spectrum-least-square support vector machine (LS-SVM) model. The better results of uninformative variable elimination (UVE)-based SPA calculation show that it is necessary to do UVE before SPA. Because the cutoff threshold selection in UVE algorithm using an artificial random noise cannot obtain the optimal results, we applied simulated annealing (SA) algorithm to estimate the optimal cutoff threshold. The discrimination results show that UVE-SA did better works than conventional UVE. Only 13 variables were obtained by UVE-SA-SPA while the conventional UVE-based SPA selected 77 variables. The best 97.5% CAR and K of 0.967 result of UVE-SPA-LS-SVM model show that it is feasible to distinguish different brands of rapeseed oils using 1H NMR spectra. It shows that a combination of SA, UVE, and SPA is effective method for the classification of rapeseed oils. Final result shows that all acyl chains, linolenyl and linoleyl chains, and triglycerides were most important for the classification.