支持向量機增量學習算法研究.pdf

1、西安電子科技大學碩士學位論文支持向量機增量學習算法研究姓名：徐新功申請學位級別：碩士專業(yè)：應用數學指導教師：周水生201201AbstractSupportVectorMachine(SVM)proposedbyVapnikinthe1990sisanewtooltosolvemachinelearningproblemswiththehelpofoptimizationtheoryIthasmanyadvantagesinsolvin

2、gthesmallsamples，nonlinearandhigh—dimensionleamingproblemsItCanhandleregressionandpanemrecognitionsuccessfullyandhasbeenwidelyusedinhandwrittendigitrecognition，facerecognition，textclassification，andreviewofpredictionandS

3、OonThelearningdataarealwaysincreasingstepbystepinsomepracticalproblemsTheincrementallearningalgorithmCanmakefillluseofhistoricalinformation，reducethetrainingscalegreatlyandsavethetrainingtime，SOitisaneffectivemethodtosol

4、vetheproblemswithaccumulatedtrainingsamplesByanalyzingtheexistingincrementalleamingalgorithmsforSVM，weproposeanewincrementallearningalgorithmbasedonthecenterdistanceratioandthecenterdensityAttheprepossessingstage，thecent

5、erdistanceratioofeverysampleiSdefinedastheratioofitsselfcenterdistancetoitsmutualcenterdistance，thensomesampleswiththehighercenterdistanceratioareselectedasthetrainingsetteplacethewholetrainingdataandtrainWhilesomenewinc

6、rementalsamplesarecoming，thenewtrainingsetincludestheprevioussupportvectors，thenewincrementalsimplesviolatingtheKKTconditionofthepreviouspattemandapartofwellchosennon—supportvectorsTowellchoosethenon—supportvectorsintotr

7、ainingset，wefirstfindthenumberofsomespecialsampleswhosedistancetotheircenterarelessthanacertainthreshold，anddefinethecenterdensityastheratioofthenumberofthesesamplestothewhole，thenwechoosesomenonsupportvectorsintothetrai

8、ningsetaccordingtotheircenterdensitiesExperimentsshowthatthenewalgorithmnotonlyreducesthecomputationtime，butalsoimprovestheclassificationprecisionKeywords：SupportVectorMachineIncrementallearningCenterdistanceratioCenterd