... An Incremental Learning Algorithm Based on Support Vector Domain Classifier YinggangZhao,QinmingHeCollegeofComputerScience,ZhejiangUniversity,Hangzhou310027,ChinaEmail:ygzl29g163.comSVDD algorithm givesus an enlightenment:whenweclassifyabinary-classdataset,ifweonlyknowpartofsample'sAbstractcategory(forexample,sampleswithcategorylabelyi=1),yettheotherpartofsample'scategoryisunknown,thenwe Incremental learning techniqueisusuallyusedtosolvecandesignnewtypeof classifier based on SVDDnamedlarge-scaleproblem.Wefirstlygaveamodifed support vector support vector domain classifier (SVDC).Thisnew classifier machine(SVM)classificationmethod ... support vector onlyneedtodescribethedatawithknowncategory,then domain classifer(SVDC),then an incremental learning obtainingthedescriptionboundaryofthisclassofdata. algorithm based on SVDCwasproposed.ThebasicideaofFinally,wecanclassifytheunknownbinary-classdatathis incremental algorithm istoobtaintheinitialtargetaccordingtotheobtainedboundary.conceptsusingSVDCduringthetrainingprocedureandthenInthispaperour incremental learning algorithm is based updatethesetargetconceptsby an updatingmodel.Difierent on SVDC,andthis algorithm ismotivatedbythefromtheexisted incremental learning approaches,inourperson -learning procedure.When learning acomplicated algorithm, themodelupdatingprocedureequalstosolveaconcept,peopleusuallyobtainainitialconceptbyusingpartquadraticprogramming(QP)problem,andtheupdatedmodelofusefulinformation,thenupdatetheobtainedconceptbystillownsthepropertyofsparssolution.Comparedwithotherutilizingnewinformation.Intermofour incremental existed incremental learningalgorithms,theinverseprocedure algorithm based on SVDC,itfirstlyutilizepartofdataofour algorithm (i.e.decreasing learning) iseasytoconduct(memoryspacepermitting),thenobtainaconcept(namelythewithoutextracomputation.Experimentresultsshowourparameterofobtaineddecisionhypersurface)bySVDC algorithm iseffectiveandfeasible. learning algorithm, finallyaccordingtotheinformationofdecisionhypersurfaceacquiredinlaststep,updatetheparameterofdecisionhypersurfacegainedinlaststeputilizingKeywords: Support Vector Machines, Support Vector Domain specializedupdatingmodelintheprocessof incremental Classifier, Incremental learning, Classification. learning, namelyupdatingtheknownconcept.Our algorithm ownsthefollowingcharacters:1.INTRODUCTION1)The incremental updatingmodelinthis algorithm Withlargeamountsofdataavailabletomachine learning hasasimilarmathematicsformcomparedwithcommunity,theneedtodesigntechniquesthatscalewellisstandardSVDC algorithm, andany algorithm usedmorecriticalthanbefore.AssomedatamaybecollectedovertoobtainthestandardSVDCcanalsobeusedtolongperiods,thereisalsoacontinuousneedtoincorporatetheobtaintheupdatingmodelofour algorithm; newdataintothepreviouslylearnedconcept. Incremental 2)Theinverseprocedureofthis algorithm, i.e.the learning techniquescansatisfytheneedforboththescalabilitydecreasing learning procedureiseasytoand incremental update.implement,thatistosaywhenweperceivethe Support vector machine(SVM)is based on statisticalgeneralizationperformancedroppedinthe learning theory,whichhasdevelopedoverlastthreedecades incremental process,wecaneasilyreturnlaststep[1,2].Ithasbeenprovenverysuccessfulinmanyapplicationswithoutextracomputation;[3,4,5,6].SVMisasupervisedbinary-class classifier, whenTheexperimentalresultsshowthe learning performancewetrainsamplesusingSVM,thecategoriesofthesamplesareofthis algorithm approachesthatofbatchtraining,andneededtobeknown.However,inmanycases,itisrarethatperformancewellinlarge-scaledatasetcomparedtootherwecanobtainthedatawiththeircategorybeknown,inotherSVDC incremental learning algorithm. words,mostoftheobtaineddata'scategoriesareunknown.InTherestofthispaperisorganizedasfollows.Insection2thissituation,traditionalSVMisn'tappropriate.wegive an introductionofSVDC,andinsection3wepresentTAXetalproposedamethodfordata domain descriptionour incremental algorithm. Experimentalandresultscalled support vector domain description(SVDD)[7],anditisconcerningtheproposed algorithm areofferedinSection4.usedtodescribedata domain anddeleteoutliers.ThekeyideaSection5collectsthemainconclusions.OfSVDDistodescribeoneclassofdatabyfindingaspherewithminimumvolume,whichcontainsthisclassofdata.Proc.5thIEEEInt.Conf. on CognitiveInformatics(ICCI'06)Y.Y.Yao,Z.Z.Shi,Y.Wang,andW.Kinsner(Eds.)801-4244-0475-4/06/$20.OO@2006IEEE802. Support Vector Domain Classifier withconstrains,==1,and0<a,<C.Wherethe2.1 Support Vector Domain Description[7]innerproducthasbeenreplacedwithkernelfunctionK(.,.),andK(.,.)isadefinitekernelsatisfyingmercerOfadatasetcontaiingNdataobjcondition,forexampleapopularchoiceistheGaussianOfadatasetcontainingNdataobjects,enl(,)=ep-xz2/2),>0fx,Z=1, ... akYkXk(I10)(13)informula(10),xkrepresents support vector, andkisFinallyweobtainthefollowingdecisionfunction:thenumberof support vector. fk(x)=sgntRk-{K(x,x)+2Ea,y,K(x,X)-ZEa,ayjy,yjK(x,ix)}Iff(x)>0,thetestedsampleiscontainedinsphere,,ESV,ESVandwelookthesamplesenclosedIspherethesame-classsgn{R21+2RklEaoy1xi+(EaciyiXi)2}objects.Otherwiseitisrejected,andwelookitastheXi,SVkxi,SVkoppositeobjects.-{K(x,x)+2Ea1yiK(x,xi)-Eaa1jy1yjK(x,xj)}xiESVxiESV3.SVDC Incremental Learning Algorithm Accordingformula(6),wesupposetheobtainedinitialsgn{ffk(x)+2RkLEaiy,x,+(aciyixi)2}parameter(sphereradius) learning withinitialtrainingsetisxicsVkxicsVkRO,andthesetof support vectorsisSVO.Theparameter(14)Fromequation(14)wecanseeitiseasytoreturnthebecomesRkinthekth incremental learning, andthesetlaststepof incremental earningwithoutextracomputation.of support vectorsbecomesSVk,andthenewdatasetinFromtheaboveanalysiswecanseeonlyconductatriflingmodification on thestandardSVDC,canitbeusedklhstepbecomesDk={(xkyk)j}l-tosolvetheupdatedmodelin incremental learning procedure.Our incremental algorithm canbedescribedasNowwesummarizeouralgorithmasfollowings:following:Step1 Learning theinitialconcept:trainingSVDCAssumewehasknownRklupdatingthecurrentusinginitialdatasetoTS,thenparameterR0ismodel~~~~~~usnSVknlnXkadaaeTSoI/hnpaaeerRmodelusingSJK,l1andnewdataset{(XiY7)}>=1obtained;WeupdatingthecurrentmodelusingthefollowingStep2Updatingthecurrentconcept:whenthenewdataareavailable,usingthemtosolveQPproblemquadraticprogramming(QP)problem:formula(11),andobtainnewconcept;ming(Rk)IRk-R112Step3Repeatingstep2untilthe incremental learning isk(Rk2_(Xk-a)'(XV-a))>XkexiDkoverwhereRk-listheradiusoflastoptimizationproblem(11),4.ExperimentsandResultswhenk=1,RoistheradiusofstandardSVDC.ItisInordertoevaluatethe learning performanceofferedbyobvious,whenRklI=0,the incremental SVDChastheour incremental algorithm, weconductedexperiment on sixdifferentdatasetstakenfromUCIMachineRepository:sameformasthestandardSVDC.WewillfoundtheBanana,Diabetes,Flare-Solar,Heart,Breast-Cancer,German.updatedmodelbythe incremental SVDCalsoownstheNotesomeofthenarenotbinary-classclassificationproblems,butwehavetransformthemtobinary-classproblembyspecialpropertyofsolutionsparsitywhichisownedbythetechnique.ExperimentparametersandDatasetareshowninstandardSVDC...