简介:Alotof3Dshapedescriptorsfor3Dshaperetrievalhavebeenpresentedsofar.Thispaperproposesanewmechanism,whichemploysseveralexistingglobalandlocal3Dshapedescriptorsasinput.Withthesparsetheory,somedescriptorswhichplaythemostimportantroleinmeasuringsimilaritybetweenquerymodelandthemodelinthedatasetareselectedautomaticallyandanaffinitymatrixisconstructed.Spectralclusteringmethodcanbeimplementedtothisaffinitymatrix.Spectralembeddingofthisaffinitymatrixcanbeappliedtoretrieval,whichintegratingalmostalltheadvantagesofselecteddescriptors.Inordertoverifytheperformanceofourapproach,weperformexperimentalcomparisonsonPrincetonShapeBenchmarkdatabase.Testresultsshowthatourmethodisapose-oblivious,efficientandrobustnessmethodforeithercompleteorincompletemodels.
简介:Privacypreservingdataminingalgorithmsarecrucialforthepersonaldataanalysis,suchasmedicalandfinancialrecords.Thispaperfocusesonfeatureselectionandproposesanewprivacypreservingdistributedalgorithm,whichcaneffectivelyselectfeaturesbasedondifferentialprivacyandGiniindexundertheMapReduceframework.Atthesametime,thetheoreticanalysisforprivacyguaranteeisalsopresented.Someexperimentsareconductedonbench-markdatasets,thesimulationresultsindicatethatduringtheselectionofimportantfeatures,theproposedalgorithmcanpreserveprivacyinformationtoacertainextentwithlesstimecostthanoncentralizedcounterpart.