Hierarchical classification using SVM

Abstract

In hierarchical classi cations classes are arranged in a hierar- chy represented by a tree forest, and each example is labeled with a set of classes located in paths from roots to leaves or internal nodes. In other words, both multiple and partial paths are allowed. A straightforward approach to learn these classi ers consists in learning one binary classi- er per node of the hierarchy; the hierarchical classi er is then obtained using a top-down evaluation procedure. In this paper, we present a new approach where node classi ers are learned by binary SVMs weighted according to the hierarchy structure and the loss function used to mea- sure the goodness of the classi ers. The result is a collection of modular algorithms that are competitive with state-of-the-art approaches. More- over, the bene ts of the modularity include the possibility of parallel implementations, and the use of all available and well-known techniques to tune binary classi cation SVMs