Application of the Modified UVE-PLS Method for a Mid-Infrared Absorption Spectral Data Set of Water—Ethanol Mixtures

The UVE-PLS (uninformative variable elimination/ partial least-squares) method 1 originally developed by Centner et al. aims at making a better calibration model and the resultant high predictive ability in the PLS method. Independent wavelength (or wavenumber) variables that cannot signi® cantly contribute to the model are eliminated, while useful ones are retained. Centner et al. referred to the eliminated variables as uninformative variables. The essence of the UVE-PLS method is addition of noise variables. A set of arti ® cial noise variables, the number of which is the same as the original experimental (wavelength) ones, are added to the data matrix as independent variables. Then, some of the experimental variables that cannot contribute more to the model construction than the noise variables are eliminated by a criterion described later. Concerning the procedure of the UVE, most of the conventional and current approaches are to pick up the best wavelength subsets by evaluating the PLS weighting vector or the PLS b-coef® cients,2±5 the PLS predictive power,6 and the PLS calibration slopes.7 Search-based selection methods, such as the genetic algorithm, 8 simulated annealing, arti® cial neural networks, and branch-andbound, have been also developed and studied. Although all those methods must work quite well theoretically, it seems that the quality or the degree of contribution of