One_Class分类器研究

One2Class1,2,3,1,1(11,210007;2.,225009;3.,210016):,(),,,One2Class.,One2Class,,,.,.:;;;:TP181:A:037222112(2009)1122496208OverviewofStudyonOne2ClassClassifiersPANZhi2song1,CHENBin2,3,MIAOZhi2min1,NIGui2qiang1(11PLAUniversityofScienceandTechnology,Nanjing,Jiangsu210007,China;2.DepartmentofComputers,YangzhouUniversity,Yangzhou,Jiangsu225009,China;3.DepartmentofComputers,NanjingUniversityofAeronauticsandAstronautics,Nanjing,Jiangsu210016,China)Abstract:InTraditionalbinaryormultipleclassificationproblemofmachinelearning,eachclassofsamplesisnecessaryforclassifierdesign;however,insomecaseonlyoneclassofsamplescanbeacquired(duetothecomplexityortheexpensivecosts),sowehavetolearnfromtheonlyoneclassofsamplesandformthedatadescriptionforclassification.Theclassificationproblemisnamedasone2classclassification.Sincenowthereexistthedomain2specificandthegenericmethods,thispaperfirstpresentsanoverviewonone2classclassification,thenemphasizesontheanalysisofthekernel2basedone2classclassifiersanddividesthisclassofmethodsintotwotypes,thatis,dual2basedandkernel2induceddistancebased.Hereafter,thecharacteristicsofthesetwotypesofmethodsareanalyzed.Finally,wesummarizetheimplementationtechniquesandapplicationsofone2classclassificationinfaultanal2ysis,anomalydetection,anddiseasediagnosisandhostilityrecognition.Keywords:one/single2class;kernelmethods;classifier;anomalydetection1,.,,.,,()(,),.,,,,.,,[1].,(One2classclassification)[2].,:.;[3].,:2008211220;:2009203230:(No.60603029);(No.BK2005009;BK2007074)11200911ACTAELECTRONICASINICAVol.37No.11Nov.2009.,.,,,oneclass:.,.2,:;;;..211(Density2basedclassifiers),,(outlier)[4,5].[6],pG(x(t)):pG(x(t))=1(2)Ndet()exp(-12(x-)T-1(x-))(1),.x,:h(x)=1,ifpG(x(t))0,ifpG(x(t))(2),h(x),1(target),0(outlier)..,,Sain[7],pMoG(x(t)):pMoG(x(t))=1(2)Nj=1j1det(j)exp(-12(x-j)Tj-1(x-j))(3)j,jiiEM.,.,BishopParzen[8,9],,.Breunig[10],(LOF,LocalOutlierFactor),KNN[11].,,.,,.,,,Parzen,:,[12].212(ANN2basedclassifiers),,Japkowicz:Autoassocia2tor[13].Autoassoci2ator,MLP,MLP,.Autoassociator,,.Autoassociator,.Autoassociator,,,,.,(Auto2Encoders)[3](LVQ,LearningVectorQuantization)[3](SOM,Self2organizingMap)[14].[15]SOM,SOM,Voronoi,SOM.,,,.,,,,.213(Clustering2basedclassifiers),794211:One2Class,K[12]K[16].,,.KK,,.KK,.,,,.,K..,..KKK,.,.[17],1,.214(Domain-basedclassifiers),,,,.SVM(SupportVectorMa2chines)(Margin)[18],,One2classSVM[19]SVDD(SupportVectorDataDescrip2tion)[20].One2classSVM,();SVDD.,Scholkopf[21]SlabSVM,.,,,.215,,,SVM.Fisher,,FDA(FisherDiscrimi2nantAnalysis),,[22,23].Stewart[24],SVM,,.Bnhalmi[25],,SVM.,,,.,,[26],.One2Class,SVDD.[27]CSVDD,.[28]One2Class.3,.,.,()(),,[29],,(SVM)PCA.3112090,,.1995Scholkopf[30],1999,TaxDuinSVDD,1(b)..,SVDD,.,Wei[31]SVDD,(minimumvol2umeenclosingellipsoid,MVEE)[32]89422009(kernelminimumvolumecoveringellipsoid,KMVCE)[33],.2001,Scholkopf[19]one2class,one2classSVM,,,[16],1(a).One2classSVM,,.,Scholkopf,,.RBF,SVDD[19],,O(n3)[34].,..,.One2classSVM,,Scholkopf[21]SlabSVM,.Tao[35]SlabSVM,,,,SVM,SlabOCC.Lanckriet[36](SingleClassMinimaxProbabilityMachine,SCMPM).Chebychev,,,.,..One2classSVM(SVDD)(),(),..,.[37],,.31231211(SVDD)SVDD:(),;SVDD.f(x,w),.aR,.(),,.,SVMxii0,Pi,SVDD:struct(R,a,)=R2+Ciis.t.(xi)-a2R2+i,i0,Pi(4)C.Lagrange,,L=ii(xi)(xi)-i,jij(xi)(xj)s.t.(1)ii=1,(2)0iC,Pi(5)KKT,i=0;0iC,,i=C(Out2lier).(xi)(xj)=K(xi,xj),:L=iiK(xi,xi)-i,jijK(xi,xj)s.t.(1)ii=1,(2)0iC,Pi(6)994211:One2Class:fSVDD(z;,R)=sign((z)-(a)2R2)=signK(z,z)-2iiK(z,xi)+i,jijK(xi,xj)R2(7)1,-1.31212One2classSVMOne2classSVM,.x,f(x),f(x),.One2classSVM:minF,Rl,R122+1lii-s.t.((xi))-i,i0(8)0i,:f(x)=sign(((x))-)(9)xi,.(0,1].SVM,,Lagrange,:L(w,,,,)=12w2+1lii--ii((w(xi))-+i)-iii(10):min12i,jijk(xi,xj)s.t.0i1l,ii=1(11),(19)i,i0.,xii:=((xi))=jjk(xj,xi)(12)31213One2classSVM,;ColinCampbell[38],.One2classSVM,.Margin,,.xi(i=1,,m),xi(xi),,w=jj(xj).w(xi)+b=0.,.f(z)=0.f(z)=iiK(z,xi)+b,,,0.f(z),.,:W(,b)=mi=1mj=1jK(xi,xj)+b(13):(1)mj=1jK(xi,xj)+b0;(2)mi=1i=1,i0(14)2006[39]K2NN,,,.313SVDD,.,,,(Pre2image);,.,,,,.,,,..:x(x)F,xX,X,F.:00522009J(wj)=(X)-(wj)2(15).Mercer:J(wj)=(X)-(wj)2=K(X,X)+K(wj,wj)-2K(X,wj)(16)dD,,.[17],,1-Means:minJ(U,V)=nj=1umjd2((xj),v)+nj=1(1-uj)ms.t.0uj1,foralli.(17),One2classSVMTsang[40],,One2classSVM,,.,,,.,,,,.4,[41,42][9,43][17,36][40].:(1)91,.(2):,,.(3):,,..,[31].,().,:.,,.,,.,,.5..,.,;,..:[1]MoyaM,KochM,HostetlerL.One2classclassifiernetworksfortargetrecognitionapplications[A].WorldCongressonNeuralNetworks93[C].Portland:InternationalNeuralNet2workSociety,1993.797-801.[2]PlattJ,ScholkopfB,Shawe2TaylorJ,etal.Estimatingthesup2portofhigh2dimensionaldistribution[R].Redmond:MicrosoftResearch,1999.[3]TaxD.One2classclassificationconcept2learningintheabsenceofcounter2examples[D].Netherlands:UniversiteitDelft,2001.[4]TarassenkoL,HaytonP,BradyM.Noveltydetectionforthei2dentificationofmassesinmammograms[A].Proc.oftheFourthInternationalIEEConferenceonArtificialNeuralNet2works[C].London:IEEpress,1995.442-447.[5]ChawlaAB.Learningfromlabeledandunlabeleddatausinggraphmin2cuts[A].Proc.18thInternationalConferenceMa2chineLearning[C].SanFrancisc