Selecting wavelet transform scales for multivariate classification

Abstract

The wavelet transform is a relatively new method that partitions a signal into components which differ in the frequency of their features. Chemical data can be thought of as a signal being composed of several different frequency components, some more analytically significant than others. Therefore, a wavelet transform can be used to transform chemical data into components, called scales, which differ in the underlying frequency of the signal. These scales possess different signal to noise ratios, or alternatively contain different amounts of information, than the original data. In this manuscript, two classification problems are used to demonstrate how this property can be exploited to improve classification performance. We demonstrate that the classification performance of the individual scales varies, and we show that selectively combining scales results in a classifier with better performance than that of either a classifier trained on the original data or developed on any individual scale. This analysis demonstrates the applicability of the wavelet transform as a simple preprocessing step that can improve classification performance. Copyright © 2007 John Wiley & Sons, Ltd.