An Empirical Comparison of Hierarchical vs. Two-Level Approaches to Multiclass Problems

Abstract

The Error Correcting Output Codes (ECOC) framework provides a powerful and popular method for solving multiclass problems using a multitude of binary classifiers. We had recently introduced [10] the Binary Hierarchical Classifier (BHC) architecture that addresses multiclass classification problems using a set of binary classifiers organized in the form of a hierarchy. Unlike ECOCs, the BHC groups classes according to their natural affinities in order to make each binary problem easier. However, it cannot exploit the powerful error correcting properties of an ECOC ensemble, which can provide good results even when the individual classifiers are weak. In this paper, we provide an empirical comparison of these two approaches on a variety of datasets, using well-tuned SVMs as the base classifiers. The results show that while there is no clear advantage to either technique in terms of classification accuracy, BHCs typically achieve this performance using fewer classifiers, and have the added advantage of automatically generating a hierarchy of classes. Such hierarchies often provide a valuable tool for extracting domain knowledge, and achieve better results when coarser granularity of the output space is acceptable.