Improved Multi Label Classification in Hierarchical Taxonomies

Hierarchical taxonomies are used to organize and retrieve information in many domains, especially those dealing with large and rapidly growing amounts of information. In many of these domains data also tends to be multi-label in nature. In this paper, we consider the problem of automated text classification in these scenarios. We present a post-processing based approach that performs smoothing on the output of an underlying one-vs-all ensemble. In order to do this we formulate a Regularized Unimodal Regression problem and give an exact algorithm to solve it. We evaluate the performance of our approach on several real-world large-scale multi-label hierarchical taxonomies and demonstrate that our proposed method provides significant gains over other related approaches.