feature-based classification of time-series data.pdf

《feature-based classification of time-series data.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、Feature-basedClassicationofTime-seriesDataALEXNANOPOULOSROBALCOCKYANNISMANOLOPOULOSDataEngineeringLab,DepartmentofInformaticsAristotleUniversityThessaloniki54006GREECEfalex,rob,manolopog@delab.csd.auth.grAbstract:Inthispaperweproposetheuseofstatisticalfeaturesfortime-seriesclassication.Theclas-si

2、cationisperformedwithamulti-layerperceptron(MLP)neuralnetwork.TheproposedmethodisexaminedinthecontextofControlChartPatterndata,whicharetimeseriesusedinStatisticalProcessControl.Experimentalresultsverifytheeciencyofthefeature-basedclassicationmethod,comparedtopreviousmethodswhichclassifytimeseries

3、basedonthevaluesofeachtimepoint.Moreover,theresultsshowtherobustnessoftheproposedmethodagainstnoiseandtime-serieslength.Key-words:Datamining,timeseries,classication,statisticalfeatures1IntroductionDataminingistheprocessofpatternidenticationinlargedatabases[1].Themainobjectivesofdataminingarepredic

4、tionanddescription.Dataminingmethodsbelongtoseveralcategories.Regressionmapsdatatopredictionvalues.Generalizationproducesasimpledescriptionfromcomplexdataandassociationndsdependenciesamongdata.Clusteringidentiesasetoftypeswithwhichdatacanbecategorized,whereasclassicationmapsdatatoasetofapredened

5、types.Datamininghasbeenmainlyappliedtorelationaldata.Non-relationaldatapresentimportantchallengestodataminingduetotheirsizeanddimensionality.Time-seriesdataaresupportedbymanydatabasesystems.Atimeseriesisasequenceofrealnumbersrepresentingthevaluesofavariableovertime.Theyhavefoundapplicationsintempora

6、l[2]andscienticdatabases,aswellasindatawarehousescontainingavarietyofdatatypes,fromstockmarketpricestoelectro-cardiograms.Miningtime-seriesdatacanrevealimportantpatterns,suchassimilarities[3],trends[4]orperiodicity[5].Sincetime-seriesdatatendtogrowrapidlyovertime,theypresentseveralperformanceissues

7、todataminingalgorithms.Time-seriesclassicationinvolvesthelearningofafunctionthatmapsaseriesintoaclassfromasetofpredenedclasses.Therefore,itrequirestheexistenceofasetofwell-denedcategories.Theclassd