Ensemble Partial Least Squares Algorithm in Mutual Information-Induced Subspace for Near-infrared Quantitative Calibration / 分析化学

ABSTRACT

In the framework of ensemble, a partial least squares(PLS) regression ensemble algorithm in subspace(MIESPLS), which is the combination of bootstrap and variable selection based on mutual information(MI), was proposed. The key of the proposed algorithm is to introduce the diversity of member models by bootstrap re-sampling on the training set and the subsequent MI calculation. Each time, those variables whose MI are lower than a defined threshold are first eliminated;then, a member model can be trained on a smaller subspace of original spectral variables. Two kinds of model fusion strategies, i.e., simple average fusion(SAF) and weighted average fusion(WAF), were adopted and compared. By two experiments concerning quantitative application of near-infrared(NIR) spectroscopy, MISEPLS is confirmed to be superior to the full-spectrum PLS and MIPLS method, i.e., PLS combined with MI-induced variable selection. The proposed MISEPLS can produce a more accurate and robust calibration model, but without increasing the complexity.