Parsimonious multiscale classification models

Wavelet transforms can be used to construct parsimonious multivariate class ification models of spectral data. The basic idea is to use a combination o f both wavelet data compression and variable/scale selection. The simplest approach determines the optimal resolution level with respect to prediction error and a model complexity measure. A classification model that maintain s an acceptable prediction error using only scales with relatively low freq uency content can be said to be parsimonious. In addition, it is possible t o enhance the interpretation of the classification model by identifying bro ad or narrow features in the spectral profiles that are important for the p rediction. However, for the simplest approach it is more difficult to deter mine the wavelength domain localization of these features compared to wavel ength-scale variable selection methods. The discriminant partial least squa res (DPLS) method is used to demonstrate the feasibility of these approache s to parsimonious model building. Copyright (C) 2000 John Wiley & Sons, Ltd .