Feasible Adaptation Criteria for Hybrid Wavelet – Large Margin Classifiers

Abstract

Hybrid wavelet – large margin classifiers have recently proven to solve difficult signal classification problems in cases where merely using a large margin classifier like, e.g., the Support Vector Machine may fail. The features for our hybrid classifier are selected from the outputs of all orthonormal filter banks of fixed length with respect to criteria measuring class separability and generalisation error.

In this paper, we evaluate a range of such adaptation criteria to perform feature selection for hybrid wavelet – large margin classifiers. The two main points we focus on are (i) approximation of the radius – margin error bound as the ultimate criterion for the target classifier, and (ii) computational costs of the approximating criterion for feature selection relative to those for the classifier design.

We show that by virtue of the adaptivity of the filter bank, criteria which are more efficient than computing the radius – margin are sufficient for wavelet adaptation and, hence, feature selection. Our results are relevant for image– and arbitrary–dimensional signal classification by utilising the standard tensor product design of wavelets.