关键词: 支持向量机, 决策树, 空间重叠度, 类间可分性, 多类分类

Abstract: For the traditional DTBSVM algorithm may have the misjudgement of the difficulty to separate the classes, a method which is based on the space overlapping region DTSVM multi-classification is proposed. The method will compute the space overlapping region by the labeled samples, then merges those classes who have the overlapped regions, and combines them into a new class, then uses an effective extra-class measurement of separability to devide the new labeled classes, to make sure that those classes who are easy to be partitioned will be separated from the root of the decision tree. It separates those classes who have overlapping regions, if do like this, the algorithm will avoid the risk of the cumulative of error successfully, and constructs better classification results decision tree. The experiment shows that the method will enhance the accuracy of DTBSVM multi-classification.

Key words: support vector machines, decision tree, space overlapping region, extra-class measurement of separability, multi-classification