资源描述：
《a time series forest for classification and feature extraction.pdf》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、ATimeSeriesForestforClassificationandFeatureExtraction✩HoutaoDeng∗Intuit,MountainView,CA,USAGeorgeRungerArizonaStateUniversity,Tempe,AZ,USAEugeneTuv,MartyanovVladimirIntel,Chandler,AZ,USAAbstractAtree-ensemblemethod,referredtoastimeseriesforest(TSF),isproposedfortimeseriesclassification.TSFe
2、mploysacombinationofentropygainandadistancemeasure,referredtoastheEntrance(entropyanddistance)gain,forevaluatingthesplits.ExperimentalstudiesshowthattheEntrancegainimprovestheaccuracyofTSF.TSFrandomlysamplesfeaturesateachtreenodeandhascomputationalcomplexitylinearinthelengthoftimeseries,an
3、dcanbebuiltusingparallelcomputingtechniques.Thetemporalimportancecurveisproposedtocapturethetemporalcharacteristicsusefulforclassification.ExperimentalstudiesshowthatTSFusingsimplefeaturessuchasmean,standarddeviationandslopeiscomputationallyefficientandoutperformsstrongcompetitorssuchasone-ne
4、arest-neighborclassifierswitharXiv:1302.2277v2[cs.LG]18Feb2013dynamictimewarping.Keywords:decisiontree;ensemble;Entrancegain;interpretability;largemargin;timeseriesclassification;∗Correspondingauthor:hdeng3@asu.eduEmailaddresses:hdeng3@asu.edu(HoutaoDeng),george.runger@asu.edu(GeorgeRunger),
5、eugene.tuv@intel.com(EugeneTuv),vladimir.martyanov@intel.com(MartyanovVladimir)PreprintsubmittedtoElsevierFebruary19,20131.IntroductionTimeseriesclassificationhasbeenplayinganimportantroleinmanydisciplinessuchasfinance[25]andmedicine[2].Althoughonecantreatthevalueofeachtimepointasafeatureand
6、usearegularclassifiersuchasone-nearest-neighbor(NN)withEuclideandistancefortimeseriesclassification,theclassifiermaybesensitivetothedistortionofthetimeaxisandcanleadtounsatisfactoryaccuracyperformance.One-nearest-neighborwithdynamictimewarping(NNDTW)isrobusttothedistortionofthetimeaxisandhasp
7、rovenexceptionallydifficulttobeat[20].However,NNDTWprovideslimitedinsightsintothetemporalcharacteristicsusefulfordistinguishingtimeseriesfromdifferentclasses.Thetemporalfeaturescalculatedovertimeseriesintervals[15],referredtoasintervalfeatures,cancapturethetempor
