Bootstrapping soft shrinkage variable selection method based on the combination of frequency and regression coefficient

Aiming at the problems that the spectral lines obtained using Fourier transform infrared spectrometer are enormous, and directly using all the spectral lines to perform multiple linear regression easily leads to over-fitting, poor stability and long analysis period. In this paper, a bootstrap soft shrinkage variable selection method based on the combination of frequency and regression coefficient is proposed. This method selects the variables based on the weight of the variables; in each iterative process, the new weight of the variable is calculated according to the regression coefficient and frequency of the variable, and the soft shrinkage of the variables is realized through weighted bootstrap sampling technology. The method was verified using the infrared spectrum datasets of corn. On the corn oil dataset, the root mean square error of prediction (RMSEP) and correlation coefficients (Rp) are 0.020 2 and 0.976 5, respectively, the number of variables is reduced from the original 700 to 13. On the corn protein dataset, the RMSEP and Rp are 0.027 9 and 0.996 8, respectively, the number of variables is reduced from the original 700 to 16. The result shows that the proposed variable selection algorithm can select fewer and more precise variables, and has practical application value. © 2020, Science Press. All right reserved.