資源描述:
《模糊聚類在特征選取中的應(yīng)用.pdf》由會(huì)員上傳分享,免費(fèi)在線閱讀,更多相關(guān)內(nèi)容在行業(yè)資料-天天文庫(kù)。
1、第44卷第6期2012年12月南京航空航天大學(xué)學(xué)報(bào)JournalofNanjingUniversityofAeronautics&Astronautics模糊聚類在特征選取中的應(yīng)用劉全金1’2趙志敏1李穎新V01.44No.6Dec.2012(1-南京航空航天大學(xué)理學(xué)院,南京,210016;2.安慶師范學(xué)院物理與電氣工程學(xué)院,安慶,246011;3.北京經(jīng)緯紡機(jī)新技術(shù)有限公司機(jī)器視覺(jué)與智能研究所,北京,100176)摘要:提出了一種基于模糊聚類算法的高維特征選取方法。首先,利用Bhattacharyya距離過(guò)濾樣本類別無(wú)關(guān)的特征;然后,基于遞歸特征別除過(guò)程,提出了基于模糊迭代自組織數(shù)據(jù)分析技術(shù)
2、(Interactiveself—organizingdataanalysistechnique,ISODATA)聚類方法,以樣本與聚類中心的加權(quán)距離作為可分性指標(biāo),產(chǎn)生候選特征子集;最后,以候選特征子集分類和聚類的接受者操作特征曲線下面積(Areaunderthereceiveroperatingcharacteristiccurve,AUC)值和正確率作為目標(biāo)函數(shù),確定最佳特征子集。將該方法用于選取5個(gè)基因表達(dá)譜數(shù)據(jù)集的特征基因,結(jié)果顯示該方法所選特征具有較好的分類和聚類能力,說(shuō)明了提出的特征選取方法的有效性。關(guān)鍵詞:特征選取}模糊迭代自組織數(shù)捂分析技術(shù);層次聚類;支持向量機(jī);K近鄰中圈分
3、類號(hào):TP391;Q812文獻(xiàn)標(biāo)識(shí)碼;A文章編號(hào):1005—2615(2012)06—0881—07ApplicationofFuzzyClusteringAlgorithmonFeatureSelectionLiuQuanjinl“,ZhaoZhiminl,LiYingxin3(1.CollegeofScience,NanjingUniversityofAeronautics&Astronautics,Nanjing,210016,China;2.SchoolofPhysicsandElectronic,AnQingNormalUniversity,Anqing,246011,China;3
4、.InstituteofMachineVisionandMachineIntelligence,BeijingJingweiTextileMachineryNewTechnologyCo.,Ltd.,Beijing,100176,China)Abstract:Anewfeatureselectionmethodbasedonclusteringalgorithmisproposedtoselecteinforma~tirefeatures.First,category—unrelatedfeaturesarekickedoutaccordingtOBhattacharyyadistance.T
5、hen,basedontheprocessofrecursivefeatureelimination,flweighteddistancebetweensampleandtheclustercentergeneratedbythefuzzyinteractiveself—organizingdataalgorithm(ISODATA)isusedastheindexoffeatureforseparatingdifferentclasses.Finally,thecandidatefeaturesubsetwiththemaxi—mumareaunderthereceiveroperating
6、characteristiccurve(AUC)valueandaccuracyratebothinclassi~ficationandclusteringtestsisselectedastheoptimalfeaturesubset.Theproposedfeaturesubsetselec~tionmethodisappliedtofivegeneexpressionprofiledatasetsandexperimentresultsshowthatthese—lectedfeatureshavegoodperformanceintermsofbothclassificationand
7、clusteringmeasurements.Re—sultsdemonstratethattheproposedmethodiseffectiveforselectinginformativefeaturesfromhigh—·di—.mensionaldataset.Keywords:featureselection;fuzzyiteractiveself—organizingdataanal