GP-SVM: Tree Structured Multiclass SVM with Greedy Partitioning

Abstract

In this paper, we propose a hierarchical SVM framework for multiclass classification problems. Use of multiple SVMs in a hierarchical structure has been a popular approach to handle multiclass classification by Support Vector Machines which are otherwise known to two-class classifiers. Among commonly-used hierarchical structures, binary tree structured SVM has computational advantages over other techniques. In order to devise an effective tree structured hierarchy of multiple SVMs, it is important to devise a process of recursive subdivision of classes, known as binarization process. We propose here a greedy heuristic as binarization strategy with partition function as the separability measure. To the best of our knowledge, no attempt has been made in this direction and the proposed algorithm takes advantage of partition function, binary structure and class membership information. We show empirically that our method provides higher accuracy with less computational overhead compared to most of the major multiclass SVM classifiers. Our method is also useful in taxonomy learning for multiclass problems.