基于节点选择优化的DAG-SVM多类别分类  

DAG-SVM Multi-class Classification Based on Nodes Selection Optimization

作　　者：沈健[1] 蒋芸[1] 邹丽[1] 陈娜[1] 胡学伟[1] 

机构地区：[1]西北师范大学计算机科学与工程学院,兰州730070

出　　处：《计算机工程》2015年第6期143-146,共4页Computer Engineering

基　　金：国家自然科学基金资助项目(61163036;61263036);甘肃省高等学校研究生导师科研基金资助项目(1201-16);西北师范大学三期"知识与科技创新工程"科研骨干培育基金资助项目(nwnu-kjcxgc-03-67)

摘　　要：有向无环图支持向量机(DAG-SVM)对于N类别分类问题,会构造N×(N-1)/2个支持向量机分类器(为每2个类构造一个支持向量机),DAG-SVM可能出现由于节点选择不佳而导致整个分类器分类结果较差的情况。为此,提出一种改进的DAG-SVM。通过为每一层建立备选节点集合进行节点选择,选取下层备选节点集合中训练分类精度最高的一个节点组合作为当前层节点的下层节点,从而优化DAG-SVM的拓扑结构。实验结果表明,与已有的DAG-SVM,1-vs-1SVM,1-vs-a SVM方法相比,该方法的分类精度较高。Directed Acyclic Graph Support Vector Machine （ DAG-SVM ） is a novel algorithm of multi-class classification. For an N-class classification problem, DAG-SVM can construct N ＆#215; （ N -1 ）/2 SVM classifiers （ one classifier for a pair of classes ） but DAG-SVM may behave poor due to the poor selection of nodes, concerning the situation raised before,the new method is proposed and the nodes selection is to establish alternative sets of nodes for every layer,and it chooses the nodes group which gets the highest training classification accuracy as the lower layer of current layer form the alternative sets of nodes, so as to optimize the topology structure of DAG-SVM. Experimental results show that compared with other methods like DAG-SVM,1-vs-1 SVM and 1-vs-a SVM,the classification accuracy of this method is high.

关 键 词：有向无环图支持向量机 分类器 多类别分类 节点选择优化 备选节点 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]