A strategy of small sample modeling for multivariate regression based on improved Boosting PLS

Abstract

In multivariate calibration technology, the problem of small sample modeling existed due to the kinds of actual situation restrictions. Traditional regression methods like PLS could not extract enough information from the limited samples. In order to try and resolve the problem, a strategy of small sample modeling for multivariate regression based on improved Boosting PLS (Im-BPLS) was proposed. In Im-BPLS, the deterministic selection method of initial calibration set and new sample weight optimization criterion were proposed. These made the information extracted easier for samples with small size and provided simple, stable and accurate regression models simultaneously. The performance of Im-BPLS was tested with three groups of small sample spectral data. The results indicated that Im-BPLS is an effective method for small calibration dataset and could give a better and more stable predictive accuracy compared with ordinary PLS and Boosting PLS.