LASSO Based Extreme Learning Machine for Spectral Multivariate Calibration of Complex Samples

Abstract

AbstractExtreme learning machine (ELM) has received increasing attention in multivariate calibration of complex samples due to its advantages of fast learning speed and good generalization ability. However, irrelevant variables in spectral matrix to target can interfere the quality of ELM modeling. Therefore, variable selection is required before multivariate calibration. In this study, least absolute shrinkage and selection operator (LASSO) combined with ELM (LASSO-ELM) is used for spectral quantitative analysis of complex samples. In the method, LASSO is firstly used to selected variables by shrinking regression coefficients of unselected variables to zero. The optimal model position s of LASSO is determined by Sp criterion. Then ELM model is built between the selected variables and analyzed target with the optimal activation function and hidden node number determined by the ratio of mean to standard deviation of correlation coefficients (MSR). Near infrared (NIR) spectra of tobacco lamina and ultraviolet (UV) spectra of fuel oil samples are used to evaluate the prediction performance of LASSO-ELM. Results show that only with tens of variables, LASSO-ELM achieves the lowest root mean square error of prediction (RMSEP) and highest correlation coefficient (R) compared with full-spectrum partial least squares (PLS) and ELM. Thus, LASSO-ELM is an effictive variable selection and multivariate calibration method for quanatitive analysis of complex samples.KeywordsLeast absolute shrinkage and selection operatorExtreme learning machineVariable selectionSpectral analysisQuantification