基于稀疏表示的近鄰保持嵌入數(shù)據(jù)降維研究.pdf

1、天津大學(xué)碩士學(xué)位論文基于稀疏表示的近鄰保持嵌入數(shù)據(jù)降維研究姓名：吳鏑申請(qǐng)學(xué)位級(jí)別：碩士專業(yè)：計(jì)算機(jī)科學(xué)與技術(shù)指導(dǎo)教師：趙政201112ABSTRACTWiththerapiddevelopmentoftheinternet，thestorageandanalysisoflarge—scaledataisgettingmoreandmoreimportantSotheseyears，lotsoftechnologiesforstoring

2、andcomputingoflargescaledataemergeHoweverwecanalsousesomemethodsofdimensionalityreductiontoreducethehighdimensionaldataeffectivelythisreducesthecomplexityofcomputingfromanotherperspectiveInrealworld，highdimensionaldataar

3、eeverywhere，butthenaturestructurebehindthemisalwaysfeaturedbyonlyafewparametersWiththefastdevelopmentofcomputervisionandpatternrecognition，inlotsofapplications，moreandmoredatadimensionalityreductionproblemsareinvolvedFor

4、example，infacerecognition，afaceimageusuallyconsistsoftensofthousandsoffeaturesbutitincludesredundantinformationsandimportantrelationsamongfeatureshidebehinditThiswillincreasethedifficultyoftrainingclassifierandcomputingI

5、t’SSOcalled”curseofdimensionality”AllofthechallengesabovelargelypremotethedevelopmentofdimensionalityreductionLinearmethodsuchasLPP(LocalPreservingProjection)，NPE(NeighborhoodPreservingEmbedding)andSPP(SparsityPreserving

6、Projections)comeupwithrecentlynolinearmethodsuchasLLE(LocalLinearEmbedding)andimprovementversionkernelNPEOneparticularlysimplebuteffectiveassumptioninfacerecognitionisthatthesamplesfromthesameclasslieonalinearsubspace，SO

7、lotsofnolinearmethodsonlyperformwellonsomeartificialdatasetsThispaperemphasizesonNPEandtherepresentationofsparsitycomeupwithrecentlNandcombinesthesetwomethods，theexperimentsshowtheeffectofnewmethodoutperformsomeclassicun