分类数据挖掘中若干基本问题的研究

《分类数据挖掘中若干基本问题的研究》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、天津大学博士学位论文分类数据挖掘中若干基本问题的研究姓名：李仁璞申请学位级别：博士专业：管理科学与工程指导教师：王正欧20030601奎鲞查兰堕主堂堡笙塞ABSTRACTFacingthemassivevolumeandhighdimensionaldatahowtobuildeffectiveandscalablealgorithmfordatamiI】ingisoneofresearchdirectionsofdatamining．Aimingataboveissues，somebasicproblemsofdatam协ingforclassificationhaveb

2、eenstudiedsubstantially船follows：Astructure-adaptiveapproachforneural-network-basedfeatureselectionisproposedinthispaper,Bypruningtheredundantinputfeannesandhiddenunitsalternatively,networkm'chitectureiskeptreasonable．啪entsshowthatthismethodcaneffectivelyselectfeatureswhileimprovetheger础ta／

3、izafionabilityofnetwork．Ahybridmethodforminingclassificationrulesisproposed．Firstlyattributereductionisdonetwicerespectivelybyroughsettheoryandbyneuralnetwork,andthenrulesare由(臼Ⅻ吐edfiomreduceddecisiontablesbyroughsettheory．Experimentalresln也showthatthisalgorithmcanproducenloIeeffectiveand出

4、nplerrulesquicklyandposseS∞$goodrobuslness．Localdiserctizafionmethodsaresimplebuthave1ms砒i甜hctofyeffect,whileglobaldiserefizalionmethodscangetbetterresultsbuthavecostlycomputation．WepresentanappropriatecompromisebetweentwokindmeIatlodsofdiscrefizafion．Throughaddinganinconsistencycheckillgt

5、oana出血培engopy-bascdlocalapproach,ouralgorithmpossessesa酉0．balprop嘶．Experimentsindicatethatwiththe鞠n碡ruleggncratorC4．5，OUrmethodcanproducestrongerrulesthanexistingmethods．Severalwidelyuseduncertaintyn璩a母Ⅱesbasedonroughsettheoryandinfonnalionentropyarecempm-edandanalyzed．WeprovethattheseⅡ墟as

6、u燃existinconsistencyinevaluatinguncertaintyofrulesundgiveanecessaryconditionofneeurringtheinconsistency．Thefurtherdirectionofbuildingmoreefficientuncertaintymea_qUl=eisalsoproposed．Analgorithmbasedonroughsettheoryfor∞血孤血培rulesfromdataclassbyclassisproposed．Firstlyareduetisderivedforeachcla

7、ssofd粗andthenforeachclassadiscernibilitymaUixandamergermatrixar℃consmaetedandrulesforthisclassaleexamctedbasedonthetwomaUices．Ex-pofme吣OnUCIdatasetsshowthatcompareAwithtraditionalmethodsouralgorithmcangetnlo鹏accuraterulesinashortertime．keywords：natamin堍classif