二类分类器的ROC曲线生成算法

:2008-09-24;:2008-12-27:(KJ2007A051);(2006KJ061B):(1982-),,,,;,,;,,ROC,,,(,243002):ROC,ROC,,ROCROCAUC,ROCROC,MBNC,MATLAB,,ROC,:;ROC;MATLAB:TP18:A:1673-629X(2009)06-0109-04AlgorithmforGeneratingROCCurveofTwo-ClassifierZOUHong2xia,QINFeng,CHENGZe2kai,WANGXiao2yu(SchoolofComputerScience,AnhuiUniversityofTechnology,Maanshan243002,China)Abstract:TheROCcurveanalysisisappliedmoreandmoreinthemachinelearningandthedataminingdomain,whichisusedtomeasureclassifiersperformancecomprehensively.TheROCcurveanalysisisatwo-dimensionaldescriptionforclassifiersperformance.Itisin2sensitivetotheclassdistributionandthedifferentmisclassificationcosts,direct-viewingaswellasbeingunderstoodandsoon,thesechar2acteristicsmakeitmoreandmoreimportantinthedistributionunknowndomainandthecostsensitivelearning.YoumustdrawROCcurveefficientlyandaccuratelysothatyoucouldmeasuretheperformanceoftheclassifierusingtheROCcurveanalysistechnologyanditsAUCapproach,anditisalsothekeytosensitivelearning.OndrawingROCcurve,thereisnotyetatoolonhandatpresent.Itwillelabo2ratethealgorithmofgeneratingtwo-classifiersROCcurveandtheconcreteprocessofdrawingROCcurvefromthetheoryandtheex2perimentseparatelyindetail,andconstructthealgorithmusingtheMATLABlanguagebasedonMBNCexperimentplatform,thencom2paredifferentclassifiersperformanceunderdifferentkindofdistribution.Byobservingtheresulitsoftheexperiment,canseethattheal2gorithmgeneratingROCcurveinthispaperisaccurate,feasibleandrealistic.Keywords:classifierappraisal;ROCcurve;MATLAB0ROC(ReceiverOperatingCharacteristic,)AUC2,ROC,,ROC:(1)(2),,,,(3),,(4)ROC,(5)[1](DiscreteClassifier)(ProbabilisticClassifier),,19620096COMPUTERTECHNOLOGYANDDEVELOPMENTVol.19No.6June2009,ROC,,ROCROC1ROC(1)(ConfusionMatrix),{p,n},p,ni;{T,F},i{T,F},4:,,TP(TruePosi2tive)1;,,FP(FalsePositive)1;,,FN(FalseNegative)1;,,TN(TrueNegative)111PnTTPFPFFNTN,M,22PNTTPRFPRFFNRTNR,:TPR=1-FNR,FPR=1-TNR[2](2)ROCROCFPR,TPR,,1,,(FPR,TPR),,12ROC,,t:it,i,,,ROCt,,,t-+,01,-+,,if(i)tROC1ROC,ROCROC:,,f;2,O(n2)(n)[1],3ROCROC,:iT,tT,iROC:if(i,+)f(i,-),,FPTP,24ROC,11ROC:L,f(i,+)01119f(i,-),p,n2ROC:ROC:,f(i):f(i)=f(i,+)/f(i,-)f(i)(0,0)forf(i)ifthen1/pelse1/nendifendfor1,3,1010,313ROC2020,,(0,0)(1,1)ROC,ROC,,ROC51,MBNC(BayesianNetworksClas2sifierusingMatlab)[3]53201+0.9011+0.402+0.8012-0.393-0.7013+0.384+0.6014-0.375+0.5515-0.366+0.5416-0.357-0.5317+0.348-0.5218-0.339+0.5119+0.3010-0.5020-0.10313ROC[4],MATLAB[5]NBC(NaiveBayesClassifier)TANC(TreeAugmentedNaiveBayesClassifier)ROC,UCI(UniversityofCaliforniainIrvine)car[6]13,4NBCROC5TANCROCROC,,ROC;67101NBCTANCROCROC[1]4NBC5TANCROC,TANCNBC;,67ROC,11116:ROC6MATLABROC,,,,ROCROC,AUC[7],,AUC,ROC,;,[7,8],ROC,ROCROC,ROC,ROC:[1]FawcettT1RocGraphs:NotesandPracticalConsiderationsforResearchers[R].PaloAlto,CA:HPLaboratories,2004.[2]HanJ,KamberM1[M].,.:,2001.[3],,.MatlabMBNC[J].,2004,43(5):729-732.[4]BradleyAP.TheuseoftheareaundertheROCcurveintheevaluationofmachinelearningalgorithms[J].PatternRecog2nitionSociety,1997,30:1145-1159.[5],,,.MATLAB[M].:,2003.[6]BohanecM1UCI[DB/OL].1997-06-01.[7],,.AUC[J].,2007,26(2):275-279.[8]ProvostF,FawcettT.AnalysisandVisualizationofClassifierPerformance:ComparisonUnderImpreciseClassandCostDistributions[C]//InProc.ThirdIntl.Conf.KnowledgeDiscoveryandDataMining(KDD-97).MenloPark,CA:AAAIPress,1997:43-48121119