数据挖掘：实用机器学习工具与技术_07变换技术ppt课件.ppt

《数据挖掘：实用机器学习工具与技术_07变换技术ppt课件.ppt》由会员上传分享，免费在线阅读，更多相关内容在教育资源-天天文库。

1、DataMiningPracticalMachineLearningToolsandTechniquesSlidesforChapter7ofDataMiningbyI.H.Witten,E.FrankandM.A.Hall2DataMining:PracticalMachineLearningToolsandTechniques(Chapter7)DatatransformationsAttributeselectionScheme-independent,scheme-specificAttrib

2、utediscretizationUnsupervised,supervised,error-vsentropy-based,converseofdiscretizationProjectionsPrincipalcomponentanalysis,randomprojections,partialleast-squares,text,timeseriesSamplingReservoirsamplingDirtydataDatacleansing,robustregression,anomalyde

3、tectionTransformingmultipleclassestobinaryonesSimpleapproaches,error-correctingcodes,ensemblesofnesteddichotomiesCalibratingclassprobabilities3DataMining:PracticalMachineLearningToolsandTechniques(Chapter7)Justapplyalearner?NO!Scheme/parameterselectiont

4、reatselectionprocessaspartofthelearningprocessModifyingtheinput:DataengineeringtomakelearningpossibleoreasierModifyingtheoutputRe-calibratingprobabilityestimates4DataMining:PracticalMachineLearningToolsandTechniques(Chapter7)AttributeselectionAddingaran

5、dom(i.e.irrelevant)attributecansignificantlydegradeC4.5’sperformanceProblem:attributeselectionbasedonsmallerandsmalleramountsofdataIBLverysusceptibletoirrelevantattributesNumberoftraininginstancesrequiredincreasesexponentiallywithnumberofirrelevantattri

6、butesNaïveBayesdoesn’thavethisproblemRelevantattributescanalsobeharmful5DataMining:PracticalMachineLearningToolsandTechniques(Chapter7)Scheme-independentattributeselectionFilterapproach:assessbasedongeneralcharacteristicsofthedataOnemethod:findsmallests

7、ubsetofattributesthatseparatesdataAnothermethod:usedifferentlearningschemee.g.useattributesselectedbyC4.5and1R,orcoefficientsoflinearmodel,possiblyappliedrecursively(recursivefeatureelimination)IBL-basedattributeweightingtechniques:can’tfindredundantatt

8、ributes(butfixhasbeensuggested)Correlation-basedFeatureSelection(CFS):correlationbetweenattributesmeasuredbysymmetricuncertainty:goodnessofsubsetofattributesmeasuredby(breakingtiesinfavorofsmallersubsets):6DataMining:PracticalMac