Hierarchically classifying documents with multiple labels

Abstract

This paper describes the evaluation of a hierarchical classifier for classifying multi-labeled documents organized in a two-level taxonomy. The hierarchical classifier consists of a tree of independent naive Bayes classifiers, with output probabilities from parent classifiers propagated to child classifiers as additional features. Each classifier uses Bi-Normal Feature Separation for word feature selection. Experiments were performed using the Weka Toolkit [7] adapted to deal with multi-labeled documents. The hierarchical classifier accuracy marginally out-performed a set of independent binary classifiers trained to classify documents for each class in the taxonomy.