Bagging Boosting and C4.5

《Bagging Boosting and C4.5》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、Bagging,Boosting,andC4.5J.R.QuinlanUniversityofSydneySydney,Australia2006quinlan@cs.su.oz.auAbstractHeath,Kasif,andSalzberg1993),andcountingop-erations(MurphyandPazzani1991Zheng1995).Breiman'sbaggingandFreundandSchapire'sUseoferror-correctingcodeswhentherearemoreboostingare

2、recentmethodsforimprovingthethantwoclasses(DietterichandBakiri1995).predictivepowerofclassierlearningsystems.BothformasetofclassiersthatarecombinedDecisiontreesthatincorporateclassiersofotherbyvoting,baggingbygeneratingreplicatedboot-kinds(Brodley1993Ting1994).strapsampl

3、esofthedata,andboostingbyad-justingtheweightsoftraininginstances.ThisAutomaticmethodsforsettinglearningsystempa-paperreportsresultsofapplyingbothtechniquesrameters(KohaviandJohn1995).toasystemthatlearnsdecisiontreesandtestingOntypicaldatasets,allhavebeenshowntoleadtoonarepre

4、sentativecollectionofdatasets.Whilemoreaccurateclassiersatthecostofadditionalcom-bothapproachessubstantiallyimprovepredictiveputationthatrangesfrommodesttosubstantial.accuracy,boostingshowsthegreaterbenet.Ontheotherhand,boostingalsoproducessevereTherehasrecentlybeenrenewedi

5、nterestinincreas-degradationonsomedatasets.Asmallchangeingaccuracybygeneratingandaggregatingmultipletothewaythatboostingcombinesthevotesofclassiers.Althoughtheideaofgrowingmultipletreeslearnedclassiersreducesthisdownsideandalsoisnotnew(see,forinstance,(Quinlan1987Buntinele

6、adstoslightlybetterresultsonmostofthe1991)),thejusticationforsuchmethodsisoftenem-datasetsconsidered.pirical.Incontrast,twonewapproachesforproducingandusingseveralclassiersareapplicabletoawideva-rietyoflearningsystemsandarebasedontheoreticalIntroductionanalysesofthebehavior

7、ofthecompositeclassier.DesignersofempiricalmachinelearningsystemsareThedataforclassierlearningsystemsconsistsofconcernedwithsuchissuesasthecomputationalcostattribute-valuevectorsorinstances.Bothbootstrapofthelearningmethodandtheaccuracyandintel-aggregatingorbagging(Breiman1

8、996)andboostingligibilityofthetheoriesthatitconstructs.Muchof(Freund