Subspace partial least squares model for multivariate spectroscopic calibration

• A subspace partial least squares model is developed for spectroscopic calibration.
• Various subspaces are constructed through different latent variable directions.
• A contribution index is defined for variable selection in each subspace.
• The efficiency of the new method is evaluated through a benchmark case study.

As a typical multivariate calibration method, the partial least squares (PLS) regression model has been widely used in the past years. To improve the calibration performance, an ensemble form of the PLS model, namely subspace PLS is proposed in the present paper. Based on the orthogonal characteristic of latent variables of the PLS model, various subspaces are constructed through different latent variable directions. Meanwhile, by defining a contribution index, the most important variables in each subspace are selected for modeling. For performance evaluation, an experimental cast study is carried out on a benchmark spectra dataset. According to the obtained results, it can be found that both of the construction and variable selection procedures in each subspace are important for ensemble modeling.