R

packagesfor

SVMSoftwarereviewMostoftheexistingSVMsoftwareiswritteninCorC++e.glibsvmSVMlightSVMTorchmySVMM-SVMPuttingSVMspecificsoftwareaside,thereislittlesoftwareavailableimplementingawiderrangeofkernelmethodswithsomeexceptionsliketheSpider

softwarewhichprovidesaMATLABinterfacetovariousC/C++SVMlibrariesandMATLABimplementationsofvariouskernel-basedalgorithms,Torchwhichalsoincludesmoretraditionalmachinelearningalgorithms,andtheoccasionalMATLABorCprogramfoundonapersonalwebpagewhereanauthorincludescodefromapublishedpaper.Rpackagese1071offersaninterfacetolibsvmprovidesarobustandfastSVMimplementationandproducesstateoftheartresultsonmostclassificationand

regression.klaRprovidesaninterfacetoSVMlightapopularSVMimplementationalongwithotherclassificationtoolslikeRegularizedDiscriminantAnalysis.LiblineaRLinear,fasteritcanbeabletohandlelarge-scaleddataset.However,mostofthelibsvmandklaRSVMcodeisinC++.Therefore,ifonewouldliketoextendorenhancethecodewithe.g.newkernelsordifferentoptimizers,onewouldhavetomodifythecoreC++code.Kernlabkernlabaimstoallowtheusertoswitchbetweenkernelsonanexistingalgorithmandevencreateanduseownkernelfunctionsforthekernelmethodsprovidedinthepackage.supportvectormachinestherelevancevectormachineGaussianprocessesarankingalgorithmkernelPCAkernelCCAkernelfeatureanalysisonlinekernelmethodsaspectralclusteringalgorithm.

e1071C-andν-classificationone-class-classification

(noveltydetection)

ε-andν-regressionlinear,polynomial,radialbasisfunction,andsigmoidalkernelsK-foldCrossValidationMulti-classclassification:one-against-onee1071Svm

letter.Rpredict.svmtunee1071ICABaggedclusteringkmeansclusteringc-shellclusteringNaiveBayesClassifierkernlabKernelfunctionifaprojectionΦ:X→Hisused,thedotproduct?Φ(x),Φ(y)?canberepresentedbyakernelfunctionkk(x,y)=?Φ(x),Φ(y)?k(x,x′)=?x,x′?

largesparsedata.egtextcategorizationk(x,x′)=exp(?σ∥x?x′∥2)general.nopriorinformationk(x,x′)=(scale·?x,x′?+offset)degree

classificationofimagesk(x,x′)=tanh(scale·?x,x′?+offset)

neuralworksMoretypesandkernelfunctionsC-svcCclassificationnu-svcnuclassificationC-bsvcbound-constraintsvmclassificationspoc-svcCrammer,Singernativemulti-classkbb-svcWeston,Watkinsnativemulti-classone-svcnoveltydetectioneps-svrepsilonregressionnu-svrnuregressioneps-bsvrbound-constraintsvmregressionletter.Rregression.RUseyourownkernelfunctionownkernel.RLSSVM

llsvm.RRVMsparsermodelsthentheSVM.Italsoperformsbetterinmanycases(speciallyinregression)rvm.RkernlabTheKernelPrincipalComponentsAnalysisclasCanonicalcorrelationanalysisLiblineaRcanproduce10typesof(generali