On learning hierarchical classifications

Many significant real-world classification tasks involve a large number of categories which are arranged in a hierarchical structure; for example, classifying documents into subject categories under the library of congress scheme, or classifying world-wide-web documents into topic hierarchies. We investigate the potential benefits of using a given hierarchy over base classes to learn accurate multi-category classifiers for these domains. First, we consider the possibility of exploiting a class hierarchy as prior knowledge that can help one learn a more accurate classifier. We explore the benefits of learning categorydiscriminants in a “hard” top-down fashion and compare this to a “soft” approach which shares training data among sibling categories. In doing so, we verify that hierarchies have the potential to improve prediction accuracy. But we argue that the reasons for this can be subtle. Sometimes, the improvement is only because using a hierarchy happens to constrain the expressiveness of a hypothesis class in an appropriate manner. However, various controlled experiments show that in other cases the performance advantage associated with using a hierarchy really does seem to be due to the “prior knowledge” it encodes.