Applying Hierarchical Classification Techniques Without A Hierarchy

Abstract

The need to classify text documents within topic hierarchies has given rise to techniques that use the hierarchical structure to improve classification performance. We propose two methods, each utilizing information in a confusion matrix, which apply hierarchical concepts to problems where no a priori hierarchy exists. One method involves learning a hierarchy and then applying existing hierarchical techniques to it. The second method performs a second stage classification, where a second stage classifier attempts to correct common errors of the first stage classifier. Our experiments using the first technique show no improvement over our best flat classifier results, while those using the second technique produce minor improvements. Both sets of experiments suggest the possibility of more pronounced success in the future.