Optimal Structure of Decision-tree-based Pairwise Support Vector Machines

Abstract

Decision Directed Acyclic Graph (DDAG) and Adaptive Directed Acyclic Graph (ADAG) are the decision-tree-based support vector machines for multiclass problems. These methods show high generalization abilities but their abilities depend on the structures. In this paper, we determine the structures so that the unclassifiable regions caused by voting are resolved by the decision boundaries for class pairs with low generalization ability. Namely, at the higher level of the tree, we select a pair of classes with higher generalization ability that is estimated by the error bounds proposed for SVMs. We demonstrate the effectiveness of our method using benchmark data sets.