A hybrid variable selection method based on uninformative variable elimination and discretized wild horse...

Spectral analysis technology has become an effective method in the field of modern analytical chemistry. However, spectral data has high-dimensional characteristics, and using single variable selection method has limitations. In this research, a hybrid strategy based on uninformative variable elimination (UVE) and discretized wild horse optimizer (DWHO) is proposed to achieve efficient and accurate spectral variable selection. Noise variables were firstly eliminated by UVE and the remaining variables were further refined by DWHO. Partial least squares (PLS) model was built with the selected variable subsets by UVE-DWHO. Four datasets of orange juice, diesel, wine and blood samples were used to validate this method. Compared with full-spectrum PLS, UVE-PLS, Monte Carlo-UVE-PLS (MC-UVE-PLS) and randomization test-PLS (RT-PLS), results showed that the number of selected variables was reduced over 85% by UVE-DWHO-PLS. The lower root mean squared error of prediction (RMSEP) was achieved and the R values were all above 0.95 for the four datasets, indicating that stable and accurate prediction model was successfully established with fewer variables. Therefore, UVE-DWHO provides a valuable new variable selection method for analyzing the spectra of complex samples.