A unified approach to hierarchical classification

This paper examines a novel approach to hierarchical classification by coupling feature optimization and classifier design with feature pruning at each classification stage. We denote this method as a hierarchical sequential pruning classification (HiSPruC) strategy. The backbone of this algorithm is a recursive application of feature ranking and designing a classifier that matches the underlying good feature distribution derived from the surviving tokens. At each stage, we prune training tokens that fall into easily separable regions. We illustrate the advantage of this approach with a number of examples derived from an active sonar echo classification problem