基于模糊最大散度差判别准则的聚类方法  

Fuzzy Maximum Scatter Difference Discriminant Criterion Based Clustering Algorithm

作　　者：皋军[1,2,3] 王士同[1,3] 

机构地区：[1]江南大学信息工程学院,江苏无锡214122 [2]盐城工学院信息工程学院,江苏盐城224001 [3]浙江大学CAD&CG国家重点实验室,浙江杭州310027

出　　处：《软件学报》2009年第11期2939-2949,共11页Journal of Software

基　　金：国家自然科学基金Nos.60773206; 60903100; 90820002;国家高技术研 究发展计划(863)Nos.2007AA1Z158; 2006AA10Z313;国防应用基础研究基金No.A1420461266;浙江大学CAD&CG国家重点实验室开放课题No.CX09B-175Z;江苏省普通高校研究生科研创新计划No.A0802~~

摘　　要：基于最大散度差判别准则提出了一种模糊最大散度差准则,并根据模糊最大散度差准则提出一种聚类方法(fuzzy maximum scatter difference discriminant criterion based clustering algorithm,简称FMSDC).该方法通过迭代优化方法实现聚类的同时还可以实现特征降维.该方法首先在最大散度差判别准则中引入模糊概念;然后通过具体原则设定模糊最大散度差判别准则中的参数η,从而在一定程度上降低了由参数η引起的敏感性;最后分别根据模糊隶属度μik、最优鉴别矢量ω进行聚类和特征降维.实验结果表明,FMSDC方法不但具有基本的聚类功能,而且具有较好的鲁棒性和较强的特征降维能力.In this paper, a fuzzy scatter difference discrimininant criterion is presented. Based on this criterion, fuzzy clustering algorithm FMSDC （fuzzy maximum scatter difference discriminant criterion based clustering algorithm） is also presented. The proposed algorithm reduces dimensionality while clustering by iterative optimizing procedure. First, it introduces the fuzzy concept into maximum scatter difference discriminant criterion; then the parameter ηin the fuzzy criterion is appropriately determined based on specific principles so that the sensibility aroused by parameter η can be decreased to some extent; At last clustering and reducing dimensionality are realized according to fuzzy membership ηik and optional discriminant vector co, respectively. Experimental results demonstrate the proposed method FMSDC is not only capable of clustering but also robust and capable of reducing dimensionality.

关 键 词：模糊最大散度差判别准则 鉴别矢量 降维 模糊聚类 鲁棒性 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]