A hybrid variable selection strategy based on continuous shrinkage of variable space in multivariate calibration.

When analyzing high-dimensional near-infrared (NIR) spectral datasets, variable selection is critical to improving models' predictive abilities. However, some methods have many limitations, such as a high risk of overfitting, time-intensiveness, or large computation demands, when dealing with a high number of variables. In this study, we propose a hybrid variable selection strategy based on the continuous shrinkage of variable space which is the core idea of variable combination population analysis (VCPA). The VCPA-based hybrid strategy continuously shrinks the variable space from big to small and optimizes it based on modified VCPA in the first step. It then employs iteratively retaining informative variables (IRIV) and a genetic algorithm (GA) to carry out further optimization in the second step. It takes full advantage of VCPA, GA, and IRIV, and makes up for their drawbacks in the face of high numbers of variables. Three NIR datasets and three variable selection methods including two widely-used methods (competitive adaptive reweighted sampling, CARS and genetic algorithm-interval partial least squares, GA-iPLS) and one hybrid method (variable importance in projection coupled with genetic algorithm, VIP-GA) were used to investigate the improvement of VCPA-based hybrid strategy. The results show that VCPA-GA and VCPA-IRIV significantly improve model's prediction performance when compared with other methods, indicating that the modified VCPA step is a very efficient way to filter the uninformative variables and VCPA-based hybrid strategy is a good and promising strategy for variable selection in NIR. The MATLAB source codes of VCPA-GA and VCPA-IRIV can be freely downloaded in the website: https://cn.mathworks.com/matlabcentral/profile/authors/5526470-yonghuan-yun.