一种基于支持向量机决策树多类分类器

A MULTI-CLASS CLASSIFIER BASED ON SVM DECISION TREE

提出一种基于支持向量机决策树的多类分类器SVMDT(Support Vector Machines based Decision Tree)。训练时,SVMDT采用样本类间最小距离原则进行决策树分叉,综合考虑局部类簇,生成一棵平衡的分类二叉树。分类时,SVMDT采用最大距离原则匹配决策。SVMDT训练时采用的距离为等效距离,综合考虑特征空间中样本类的中心距离以及样本类自身的分布特点,使得训练过程中确定各个SVM的优先级别更加合理,由此生成的决策树将特征空间严格划分开,避免了拒识区域的出现。UCI样本数据集实验结果表明,和传统的1对多SVM分类器相比,SVMDT具有训练速度快、分类速度快,分类精度高的特点。

A multi-class classifier based on support vector machine decision tree, named as SVMDT, is proposed in this paper. At the training phase, the minimum distance measurement between sample class is used by SVMDT as the forking principle of decision tree. The grown decision tree is a balanced binary tree because local class cluster are comprehensively considered. At classifying phase, the maximum distance measurement is used by SVMDT as the matching decision principle. The distance adopted by SVMTD in training is the equivalent distance, two factors, including distance between centres of sample class in feature space and the distribution feature of each cluster in sample class, are synthetically employed, which leads to determine a more reasonable priority grade for each SVM during the training process. Moreover the obtained decision tree divides the mapped space into many sub-spaces. There are no unclassifiable regions in the sub-spaces. The experiment based on UCI data set shows that the SVDMT classifier has more rapid training speed, classifying speed and better generalization performance than traditional one-against-the-rest classifier.