类相关性影响可变选择性贝叶斯分类器  

A Selective Bayesian Classifier Based on Change of Class Relevance Influence

作　　者：程玉虎[1] 仝瑶瑶[1] 王雪松[1] 

机构地区：[1]中国矿业大学信息与电气工程学院,江苏徐州221116

出　　处：《电子学报》2011年第7期1628-1633,共6页Acta Electronica Sinica

基　　金：国家自然科学基金(No.60804022;No.60974050;No.61072094);教育部新世纪优秀人才支持计划基金(No.NCET-08-0836);霍英东教育基金会青年教师基金(No.121066);江苏省自然科学基金(No.BK2008126)

摘　　要：在最大相关最小冗余(mRMR)属性选择方法的基础上,通过设置一个调节因子来改变类别相关性在属性选择中的影响程度,解决mRMR方法易于引入冗余属性的问题,提出一种类相关性影响可变选择性贝叶斯分类器(CCRI SBC).为克服人为指定属性个数易于导致的分类结果随意性,采用贝叶斯信息准则来自动确定最优属性个数.为使CCRI SBC能够处理含有连续变量的数据集,提出等频类别依赖最大化离散化方法,具有分类准确率高和离散化时间短的优点.UCI数据集的实验结果表明,本文方法能够有效处理离散和连续高维数据的分类问题.A selective Bayesian classifier based on change of class relevance influence（CCRI SBC） was proposed by introducing a regulator factor into an attribute selection method,namely maximum relevance and minimum redundancy（mRMR）.The regulator factor was used to change the influence degree of class relevance on the attribute selection,which can avoid the existence of redundant attributes in mRMR.In addition,a Bayesian information criterion was used to determine the optimal number of attributes automatically,which can overcome the randomness of classification results that easily caused by the setting number of attributes manually.In order to further make the CCRI SBC is applicable for continuous data,a discretization method,i.e.,equal frequency class attribute interdependent maximization was proposed,which has advantages of high classification correct rate and short discretization time.Experimental results on UCI datasets show that the proposed method can deal with the classification problem for discrete or continuous and high-dimensional data effectively.

关 键 词：选择性贝叶斯分类器 属性选择 最大相关最小冗余 贝叶斯信息准则 离散化 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]