Feature Selection Algorithms

《Feature Selection Algorithms》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、FeatureSelectionAlgorithms:ASurveyandExperimentalEvaluationLuisCarlosMolina,LluísBelanche,ÀngelaNebotUniversitatPolitècnicadeCatalunyaDepartamentdeLlenguatgesiSistemesInformáticsJordiGirona1-3,CampusNordC6,08034,Barcelona,Spain.{lcmolina,belanche,angela}@lsi.upc.esAbstractTheFSAscanbeclassified

2、accordingtothekindofout-puttheyyield:(1)thosealgorithmsgivinga(weighed)lin-Inviewofthesubstantialnumberofexistingfeaturese-earorderoffeaturesand(2)thosealgorithmsgivingasubsetlectionalgorithms,theneedarisestocountoncriteriathatoftheoriginalfeatures.Bothtypescanbeseeninanunifiedenablestoadequate

3、lydecidewhichalgorithmtouseincer-waybynotingthatin(2)theweightingisbinary.tainsituations.Thisworkreviewsseveralfundamentalal-TheworkpresentedinthispaperiscenteredinFSAsgorithmsfoundintheliteratureandassessestheirperfor-tacklingthefeatureselectionproblemoftype(2),studiedmanceinacontrolledscenar

4、io.Ascoringmeasureranksformanyyearsbythestatistical[18]aswellasthemachinethealgorithmsbytakingintoaccounttheamountofrel-learning[38]communities.Researchdevelopedwithintheevance,irrelevanceandredundanceonsampledatasets.machinelearningareaisusuallyfocusedontheproposalofThismeasurecomputesthedegr

5、eeofmatchingbetweenthenewalgorithms,theoreticallearningresultsofexistingal-outputgivenbythealgorithmandtheknownoptimalsolu-gorithmsorempiricalstudies(evaluationsorapplications).tion.Samplesizeeffectsarealsostudied.Inthisresearch,severalfundamentalalgorithmsfoundintheliteraturearestudiedtoasses

6、stheirperformancein1.Introductionacontrolledscenario.Tothisend,ameasuretoevaluateFSAsisproposedthattakesintoaccounttheparticulari-tiesofrelevance,irrelevanceandredundanceonthesam-Thefeatureselectionproblemintermsofsupervisedin-pledataset.Thismeasurecomputesthedegreeofmatchingductivelearningis:

7、givenasetofcandidatefeaturesselectbetweentheoutputgivenbyaFSAandtheknownoptimalasubsetdefinedbyoneofthreeapproaches:a)thesubsetsolution.Samplesizeeffectsarealsostudied.Theresultswithaspecifiedsizethatoptimizesanevaluationmeasure,illustrat