Information-based dichotomization: A method for multiclass Support Vector Machines

Approaches for solving a multiclass classification problem by support vector machines (SVMs) are typically to consider the problem as combination of two-class classification problems. Previous approaches have some limitations in classification accuracy and evaluation time. This paper proposes a novel method that employs information-based dichotomization for constructing a binary classification tree. Each node of the tree is a binary SVM with the minimum entropy. Our method can reduce the number of binary SVMs used in the classification to the logarithm of the number of classes which is lower than previous methods. The experimental results show that the proposed method takes lower evaluation time while it maintains accuracy compared to other methods.