高維復(fù)雜數(shù)據(jù)的有監(jiān)督特征提取方法.pdf

1、中國(guó)農(nóng)業(yè)大學(xué)碩士學(xué)位論文高維復(fù)雜數(shù)據(jù)的有監(jiān)督特征提取方法姓名：胡曉琴申請(qǐng)學(xué)位級(jí)別：碩士專(zhuān)業(yè)：應(yīng)用數(shù)學(xué)指導(dǎo)教師：經(jīng)玲20070601AbstractNowdaysthefastincreasingdimensfonMityofdatasetshasbroughttremendouscha上lengeindataprooe日ingeoitsurgenttodiscovertheintrinsicetructttt№ofthecomplexd

2、ata日et8withhighdimeusianalityDimensionalityreductionnimportantpreprocessing咖tocopewithhighdimensionsldatasets，hasattractedconsiderableattentionsinrecentyearsItsachallengingworktoextractinformationfromcomplexdatasets，whic

3、hhasimportanttheoreticalsignificanceandwideapplicationsinpatternrecogftition，Bioinfornntion，digitalimageprocessingand∞OnInthisdissertationthetheoriesandmethodsofdimensionalreductioninhighdimensional∞m—plexdatasetaarestud

4、iedfromclarificationandvisualizatioa也wpointsCorrespondingtodifferentca8民耽havedevelopedM”ceraln唧featureextractionmethodsbelowBasedOftclassificationprincipalweproposeAftewsupervisedfeatureextractioftalgorithmcalleddiscriim

5、natemultidimensionalmappingfDMM)，whichisespechlIyeffectiveforsmallsampledatasetswithhighdimensionalityThealgorithmtakesadvantagesofclassicalMDSandLDAandmeanwhileavoidstheirdisadvantagesDMMhasmanyappealing觸tI刪：翱na】4凹compU

6、tatienalcomplexitynoparoJ口aetet8tobeselected，andhavingformalsolutionWegivetheoreticalanalysisofDMM，anddemonstratewithnumericalexperimentsthatDMMiseffectiveinpracticalriseIfthenumberofdataisverylarge，itisexpensiveforDMMin

7、practiceTosolvethisproblemwedescribeavariationOnDMM(LDMM)，whichworkstractablywithverylargedatasets，andpreservestileattractivepropertiesofDMMThe∞lectionoflandmarkparameterisflexibleandmucht／loreefficientcomparingtoMDDThee

8、xperimentsremtltsshowthatLDMMcan舀wresultsapproximatetotheOiltputofDMMWeextendDMMfornonfinsormanifoldsanddevelopan哪nonlhasardimensionalityreductionmethodcalledgeodesicmetricdiscriminatemapping(GDM)GDMIM糟thegeodeBicdistanc

9、etoestimatetheintrinsicgeometryoftheIInderIyiD【grearold，ioUglynomfifteatdatasets∞D(zhuǎn)besuee自sfullyembeddedinthiswayIsomapalgorithmwasproposedforle_aroiftganonlineaxⅡlanifoIdandhadbeendemonstratedits8II代懶fmapplicationstomany

10、d如eetaHowever，wheumanifoldb錢(qián)姍凹disconnected，thebetweenregiongeodesicdiericeswillvarydrasticallyandIsomapcannotachievedesirableembeddingWedevelopamethodtocomputethebetweenregiongeodesicdistances，andthengeneralizeIsonmpforu

11、nconnectedmanifoldsExperimentalresultsOn目／ntheticandrealdataillustratetheefficiencyoftheproposedaigorithntGDMcan幽hegeneralizedforunconnectedmanifoldsasthe8aqleway∞Isomapkeywords：featureextractionmultidimensionalscaling1i