基于最小k近邻错分率编码确定方法及其在多类分类中的应用  

Designing of output codes based on minimal k nearest neighbor classifying error and its application in multi-class classification

作　　者：周进登[1,2] 王晓丹[1] 

机构地区：[1]空军工程大学导弹学院 [2]94719部队

出　　处：《控制与决策》2011年第9期1295-1302,共8页Control and Decision

基　　金：国家自然科学基金项目(60975026);博士学位论文创新基金项目(DY11102)

摘　　要：构造输出编码矩阵是将多类分类问题分解为多个两类分类问题的有效方法之一,如何判断一个编码阵的好坏是此类问题的关键.提出以最小k近邻错分率作为评价标准,把构造问题简化为一个搜索问题.在M类的所有二类划分空间中,通过行交换规则和有限启发式搜索策略搜索出k近邻错分率最小的个二类划分,并依据编码规则得到最终输出编码矩阵.实验中用人工数据集和UCI数据集分别测试,通过与几种经典的编码方法比较,结果表明该编码方法能在编码长度较小情况下得到更好的分类效果.Generating the matrix of output codes is one of the most effective methods to reduce multiclass to binary. How to establish the effect of coding matrix is the key to solve this problem. Therefore, the k-nearest neighbor classifying error rateis taken as a metewand, which can translate construction issue into searching problem. The matrix of standard output codes is generated by using l binary partitions based on row-exchange rule and limited heuristic search strategy in the space of M classes. Experimental results on artificial data and UCI with logistic linear classifier（logic） and support vector machine（SVM） as the binary learner show that the proposed method has better performance of classification with shorter length of coding matrix.

关 键 词：多类分类 K近邻 输出编码 启发式搜索 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]