Application of recursive partial least square regression for prediction of apple juice sensory attributes from NMR spectra

Author Summary: This study demonstrates the application of a novel variable selection method here employed for the prediction of sweet and sour taste of apple juice from Nuclear Magnetic Resonance (NMR) spectra. The method is called recursive weighted Partial Least Square (rPLS). It operates by iteratively re-weighting the spectral variables using the regression coefficients calculated by PLS. The only parameter to be estimated by the operator is the number of latent factors to be used in the model. This approach provides an easier model interpretation than a regular PLS model, since it converges towards a very limited number of variables and therefore the assignment effort is drastically reduced. These properties suggest a profitable use of the rPLS for the prediction of even more complex sensory features from different types of spectroscopic data.