Novel comprehensive variable selection algorithm based on multi-weight vector optimal selection and bootstrapping soft shrinkage

The dimensionality of spectral data is increasing with the advancements in spectral technology. Therefore, there is an urgent need to develop high-performance variable selection algorithms for chemometrics applications. This study proposes a novel multi-weight optimal-bootstrap soft shrinkage (MWO-BOSS) method for variable selection based on the bootstrap soft shrinkage (BOSS) algorithm, comprising three effective improvement strategies. First, the optimal weight vector of six weight vectors are used as weights of the selection variables, rather than the absolute value of the regression coefficients based only on a single weight vector. Second, in each loop, a step-by-step strategy is implemented to determine the optimal set of variables. Finally, a smoothing operation is added to the weight vector to improve the anti-noise performance of the algorithm. The performance of the MWO-BOSS algorithm was tested on the four spectral datasets corn protein, corn oil, soil, and beer and compared with six high-performance algorithms, namely interval partial least squares (iPLS), Moving Window Partial LeastSquares(MWPLS), competitive adaptive reweighted sampling (CARS), variable combinatorial population analysis (VCPA), VCPA-IRIV and BOSS. The results show that the MWO-BOSS algorithm effectively improves the predictive ability of the model, with MWO-BOSS-Step-S providing the best results among the four tested datasets.

Automated summary

The dimensionality of spectral data is increasing, so there is a need for high-performance variable selection algorithms in chemometrics. This study proposes a novel MWO-BOSS method for variable selection based on the BOSS algorithm, with three improvement strategies. The MWO-BOSS algorithm effectively improves the predictive ability of the model and performs best among the tested datasets.