欢迎来到天天文库
浏览记录
ID:56029182
大小:412.25 KB
页数:7页
时间:2020-06-19
《流形学习算法中邻域大小参数的递增式选取.pdf》由会员上传分享,免费在线阅读,更多相关内容在行业资料-天天文库。
1、第40卷第8期计算机工程2014年8月Vo1.40NO.8ComputerEngineeringAugust2014·人工智能及识别技术·文章编号:1000-3428(2014)08-0194-07文献标识码:A中图分类号:TPI8流形学习算法中邻域大小参数的递增式选取邵超,万春红,赵静玉(河南财经政法大学计算机与信息工程学院,郑州450002)摘要:流形学习算法能否成功应用依赖于邻域大小参数的选取是否合适,但该参数在实际中通常难以高效选取。为此,提出一种邻域大小参数的递增式选取方法。按照流形的局部欧氏性,邻域图上的所有邻域都呈线性或近似线性,邻域大小参数若合适,此时所有邻域的线性度量可
2、聚成一类;而邻域大小参数若不合适,邻域图上就会有部分邻域不再线性,其线性度量也不能聚成一类。对邻域图上的每一个邻域执行加权主成分分析,用重建误差对其线性程度进行度量,并计算相应的贝叶斯信息准则,以探测其聚类个数,从而实现对邻域大小参数的递增式选取。实验结果表明,该方法无需任何额外参数,具有较高的运行效率。关键词:流形学习;邻域大小;局部欧氏性;加权主成分分析;重建误差;叶斯信息准则IncrementalSelectionofNeighb0rh00dSizeParameterforManifoldLearningAlgorithmsSHAOChao,WANChun—hong,ZHAOJin
3、g—yu(SchoolofComputerandInformationEngineering,HenanUniversityofEconomicsandLaw,Zhengzhou450002,China)【Abstract】Thesuccessofmanifoldlearningalgorithmsdependsgreatlyuponselectingasuitableneighborhoodsizeparameter,however,itisallopenproblemhowtOdothisefficiently.Tosolvethisproblem,thispaperproposes
4、anefficientmethodtOincrementallyselectasuitableneighborhoodsize.AccordingtothelocalEuclideanpropertyofthemanifold,thatallthenei曲b0rhoodsintheneighborhoodgrapharelinearoralmostlinearisthebasistothinkthecorrespondingneighborhoodsizesuitable,whentheirlinearitymeasurescanremainsmallandfallintooneclus
5、ter.However,oncetheneighborhoodsizebecomesunsuitable,someneighborho0dsarenonlinear,andtheirlinearitymeasurescannotfallintooneclusteranymore.So,thismethodrunstheweightedPrincipalComponentAnalysis(PCA)oneachneighborhoodintheneighborhoodgraph,toobtainitsreconstructionerorasitslinearitymeasure,andcom
6、putesthecorrespondingBayesianInformationCriterion(BIC)todetectthenumberofclustersofallthereconstructionerrorsintheneighborhoodgraph,bywhichtheneighborhoodsizecanbeselectedincrementally.Experimentalresultsthatthismethoddoesnotrequireanyextraparameter,andhashighrunefficiency.【Keywords】manifoldlearn
7、ing;neighborhoodsize;localEuclideanproperty;weightedPrincipalComponentAnalysis(PCA);reconstructionerror;BayesianInformationCriterion(BIC)DOI:10.3969/i.issn.1000—3428.2014.08.037小参数的选取是否合适,然而,目前该参数在实际中1概述通常还难以高效选取,另外,数据
此文档下载收益归作者所有