类依赖的相关性多类分类器.pdfVIP

ComputerEngineeringandApplications计算机工程与应用 2010，46(2) 7 类依赖的相关性多类分类器 陈晓红 1,2 CHEN Xiao-hong 1．南京航空航天大学 理学院，南京 210016 2。南京航空航天大学 信息科学与技术学院，南京 210016 1．CollegeofScience，NanjingUniversityofAeronauticsandAstronautics，Nanjing210016，China 2．CollegeofInformationScienceandTechnology，NanjingUniversityofAeronauticsandAstronautics，Nanjing210016，China E—mail：lyandcxh@nuaa．edu．cn CHEN Xiao-hong．Class-specfiiccorrelationmutli-classclassfiier．ComputerEngineeringandApplications。2010，46(2)：7—10． Abstract：CanonicalCorrelationAnalysis(CCA)isanimportantdimensionreductionmethodforfeatureextractionbyvirtueofthe correlationbetweensamples．CorrelationDiscriminantAnalysis(CDA)maximizesthecorrelationofthesameclasssamplesandmini- mizesthecorrelation betweendifierentclasssamplesin the feature space．whichcan be treated asclass-specific CCA．n efeature extractionofthetwomethodsisindependentofthefollowingclassifier，i．e．theexrtacted featuresunnecessarilyfavortheclassification． Thisindependence affectstheperformanceofclassifiers unavoidably．In virtueofthecharacteristicsthattheverticesoftheregular simplex areequidistantand affine invariant，usetheverticesofthereuglarsimplextocode the class1abelsSOthattheclassinofr mationisincorporatedintheclassifierdesign．Thealgorithm。termedasClass—specificCorrelationMulti-classClassifier(CCMC)，maxi— mizesthecorrelation betweeneach sampleanditsclasslabe1．minimizesthecorrelation betweeneach sample and the others class labelsatthesametime．Furthemr ore，thispaperalso proposesthecorrespondingkernelized version ofCCMC bycombiningwiththe EmpiricalKernelmapping．called asEK-CCMC．TheexperimentalresultsonbothartificialdatasetandUC1benchmark datasetsshow thattheproposedCCMC and EK—CCMC enhancetheclassifierperfomr ancebyusingclass—specificcorrelationtodesign classifier． Keywords：CanonicalCorrelationAnalysis(CCA)；lineardiseriminantanalysis；featureextraction；class-specific；multi-classclassi- tierdesign 摘 要：典型相关