Binary-decision-tree-based multiclass Support Vector Machines

In this paper, we propose Support Vector Machine classifiers utilizing binary decision tree to solve multiclass problems. In training process, we determine the hyperplane that separates the classes into two categories at the top node and this procedure is repeated until only one class remains. In order to obtain higher accuracy, easier separable class should be separated in the upper node. Hence a decision tree is constructed according to the degree of difficulty to classify classes which is measured by Euclidean distance and standard deviation. The proposed binary decision tree SVM is designed to obtain superior performance in classification accuracy and speed. Its performance is measured on the dataset of Sogou news corpus which has eight classes of news. The method is compared with some traditional methods like one-against-one and one-against-all. The results of the experiments indicate that it has comparative accuracy with other traditional methods and much faster recognition speed.