Experiments with hierarchical text classification

This paper applies Boosting to hierarchical text classification where the hierarchical structure is given as directed acyclic graph and compares the results to Support Vector Machines. Hierarchical classification is performed top-down and in each node a flat classifier decides if a document should be further propagated or not. As flat classifiers BoosTexter, CentroidBooster and Support Vector Machines are used, were CentroidBooster is an AdaBoost.MH based alternative similar to BoosTexter. Experiments on the Reuters Corpus Volume 1 and the OHSUMED data set show that the F-1-measure increases if the hierarchal structure of a data set is taken into account. Regarding time complexity we show, that depending on the structure of a hierarchy, learning and classification time can be reduced. Besides these hard classification approaches we also investigate the ranking performance of hierarchical classifiers. Ranking, which can be achieved by providing a meaningful score for each classification decision, is important in most practical settings. We investigate an approach based on using a sigmoid function for calculating a meaningful score, where parameter estimation is based on error bounds from computational learning theory.