结合特征选择的二叉树SVM多分类算法  

Binary tree SVM multi-class classification algorithm with feature selection

作　　者：范柏超[1] 王建宇[1] 薄煜明[1] 

机构地区：[1]南京理工大学自动化学院,江苏南京210094

出　　处：《计算机工程与设计》2010年第12期2823-2825,共3页Computer Engineering and Design

摘　　要：为解决现有二叉树SVM多分类算法采用固定的特征集和结构存在分类精度较低的问题,提出了一种结合特征选择的二叉树SVM多类分类算法。采用自上而下分裂的方式构造整个二叉树结构,首先计算各节点的所有可能分割,并以分离度和相似度作为依据为各分割选择有效的分类特征子集,再以相应的特征子集计算各分割的类间距,最后选择类间距最大的分割生成子节点。实验结果表明,该算法分类精度较高且计算复杂度低。To deal with the low accuracy in existing binary tree SVM algorithm,which adopts a fixed feature set and binary tree structure,a novel binary tree SVM multi-class classification algorithm with feature selection is proposed.The entire binary tree is structured in a top-down split way.For each node,all possible partitions are calculated firstly,and then the inter-class distance of each partition is obtained through its corresponding effective classification feature sub-set selected by separability and similarity,finally the partition with maximum inter-class distance is chosen to generate the child nodes.Experimental result demonstrates the algorithms’ high classification accuracy and low computational complexity.

关 键 词：支持向量机 二叉树 多类分类 分离度 相似度 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]