资源描述：
《多类支持向量机算法的研究和改进》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。
1、上海交通大学硕士学位论文多类支持向量机算法的研究和改进姓名：严慧敏申请学位级别：硕士专业：计算机应用技术指导教师：谢康林20060101上海交通大学硕士论文1)首先转换这两种方案对应的最优化问题到它们的对偶形式从而使用更为方便的数值算法来进行求解并且我们在实现这两种方案的时候使用了适合大规模数据的算法因此可以支持大规模的数据运行2)接着通过实验对这两种方案的有效性进行了验证实验结果显示我们提出的改进方案很有效关键词支持向量机多分类问题多类支持向量机大规模优化算法II上海交通大学硕士论文MULTI-CLASSSUPPORTVE
2、CTORMACHINES’STUDYANDIMPROVEMENTABSTRACTAsthemostsuccessfulmachinelearningmethod,SupportVectorMachinehasmadealotofgoodapplications,includingtextclassification,hand-writtencharactersrecognition,facerecognitionetc.Thebiggestdiscriminationfromothermachineslearningmeth
3、odsisthatSVMiscorrespondingtoseveralprinciplesinstatisticallearningtheory,suchasstructureriskminimization.AnditcouldtheoreticallyprovedthattheexpectedriskofSVMhasanupperbound.SVMisreallysuccessfulexceptforonepoint:itisbinaryinnature.Butinreallife,multi-classproblem
4、isprevailing.AndSVM'sapplicationinmulti-classproblemstillhasalongwaytogo..TheexistingmethodsforSVMtosolvemulti-classproblemgoesintwodirections:firstistoconvertmulti-classtoseveralbinary;secondistherealmulti-classSVM,thatis,consideringallthedataatonce.Inthisthesis,w
5、ehavestudiedtheprinciplesandimplementationmethods,andalsoimprovedthealgorithm.Wehaveproposedtwoimprovementmethods,respectivelyfordifferenttargetsinmulti-classSVM:costfactorandsubproblem.Theformerone'smainideaistoconsidertherelationshipbetweenclasses,forexample,thed
6、istance,andcorporatetherelationshipintooriginalalgorithm.Thelatter'smainideatobalanceeverysubprobleminmulti-classSVM,insteadofjustaddingallthesubproblemstofindasolution.Twomethodscomefromtwopointsofview,buttheyallshowoneimportantidea:togetaglobaloptimalsolutions..A
7、ftertheproposal,wehavedonethefollowingwork:1)Firstweconvertthesetwoproblemstotheirdualform,soastoutilizeeasiernumericalmethodtosolvetheproblem,andwealsousealgorithmwhichissuitableforlarge-scaledata;thereforetheimplementationcouldrunlarge-scaledataset,whichisveryimp
8、ortantnowadayswithsomuchdata.III上海交通大学硕士论文2)Secondwevalidatethesetwomethodsbyexperiments.Andtheexperimentresultsshowourmethodsareveryeffective.Ke
