基于約束的中垂面相似度準(zhǔn)則及其應(yīng)用.pdf

1、南京航空航天大學(xué)碩士學(xué)位論文基于約束的中垂面相似度準(zhǔn)則及其應(yīng)用姓名：高山申請(qǐng)學(xué)位級(jí)別：碩士專業(yè)：計(jì)算機(jī)軟件與理論指導(dǎo)教師：張道強(qiáng)201101u??R??qOK9A^ABSTRACTMeasuringthesimilaritybetweendataobjectsisoneoftheprimarytasksfdistancebasedtechniquesindataminingmachinelearninge.g.distancebased

2、clusteringdistancebasedclassification.Facertainproblemusingpropersimilaritymeasurementwillmakeitbettersolved.Memeresearchesshowthatgettingasimilaritymeasurementfittingtoacertainproblemwithpairwiseconstraintscansignifican

3、tlyimprovethealgithmicperfmances.Nowadaysresearchesonsimilaritymeasurementwithpairwiseconstraintsaremainlydistancemetriclearningbasedonpairwiseconstraintswhichusespairwiseconstraintstolearnadistancematrixfclassificationc

4、lustering.Weproposeanewsimilaritymeasurementwithpairwiseconstraintsespeciallywithcannotlinkconstraints.Themaincontributionsofthisthesisaresummarizedasfollows:Firstlybasedontheanalysisof1NNSVMalgithmstheconceptionofperpen

5、dicularhyperplaneisextracted.FromtheperpendicularhyperplaneofcannotlinkconstraintWeproposethePerpendicularHyperplaneSimilaritycriterionalsoshowthathowtocomputethesimilaritymeasurementwithatoyproblem.Secondlythecriterioni

6、sappliedatclusteringweproposetheclusteringalgithmMPHSshtfPerpendicularHyperplaneSimilarity.WedivideMPHSflineardatasetsnonlineardatasetsproposeseveralsubalgithms.OnseveralUCIdatasetsseveralimagerydatasetsMPHSoutperfmsothe

7、rsemisupervisedclusteringalgithms.Thirdlyensemblelearningisintroducedtosemisupervisedclusteringexperimentalresultsshowthatourensemblealgithmisbetterthanothers.Finallythecriterionisappliedatclassification.Whenobtainingthe

8、similaritymatrixbyMPHSPCPprocessweuse1nnsvmtoclassifyrespectivelyproposethealgithmsmphs1nnmphssvm.Wealsointroduceensemblelearningtoclassificationproposethealgithmsmphs1nnbaggingmphssvmbagging.Theexperimentalresultsonseve