r statistical application development cart and beyond

《r statistical application development cart and beyond》由会员上传分享，免费在线阅读，更多相关内容在工程资料-天天文库。

1、10CARTandBeyondInthepreviouschapter,westudiedCARTasapowerfulrecursivepartitioningmethod,usefulforbuilding(non-linear)models.Despitetheoverallgenerality,CARTdoeshavecertainlimitationsthatnecessitatesomeenhancements.Itistheseextensionsthatformthecruxofthefinal

2、chapterofthisbook.Forsometechnicalreasons,wewillfocussolelyontheclassificationtreesinthischapter.WewillalsobrieflylookatsomelimitationsoftheCARTtool.ThefirstimprovementthatcanbemadetotheCARTisprovidedbythebaggingtechnique.Inthistechnique,webuildmultipletrees

3、onthebootstrapsamplesdrawnfromtheactualdataset.Anobservationisputthrougheachofthetreesandapredictionismadeforitsclass,andbasedonthemajoritypredictionofitsclass,itispredictedtobelongtothemajoritycountclass.AdifferentapproachisprovidedbyRandomForests,whereyouc

4、onsiderarandompoolofcovariatesagainsttheobservations.WefinallyconsideranotherimportantenhancementofaCARTbyusingtheboostingalgorithms.Thechapterwilldiscussthefollowing:Cross-validationerrorsforCARTThebootstrapaggregation(bagging)techniqueforCARTExtendingtheCA

5、RTwithrandomforestsAconsolidationoftheapplicationsdevelopedfromChapter6toChapter10,CARTandBeyondImprovingCARTIntheAnotherlookatmodelassessmentsectionofChapter8,wesawthatthetechniqueoftrain+validate+testmaybefurtherenhancedbyusingthecross-validationtechnique.

6、Inthecaseof,linearregressionmodel,wehadusedtheCVlmfunctionfromtheDAAGpackageforthepurposeofcross-validationoflinearmodels.Thecross-validationtechniqueforthelogisticregressionmodelsmaybecarriedoutbyusingtheCVbinaryfunctionfromthesamepackage.Profs.TherneauandA

7、tkinsoncreatedthepackagerpart,andadetaileddocumentationoftheentirerpartpackageisavailableontheWebathttp://www.mayo.edu/hsr/techrpt/61.pdf.RecalltheslightimprovementprovidedinthePruningandotherfineraspectsofatreesectionofthepreviouschapter.Thetwoaspectsconsid

8、eredthererelatedtothecomplexityparametercpandtheminimumsplitcriteriaminsplit.Now,theproblemofoverfittingwiththeCARTmaybereducedtoanextentbyusingthecross-validationtechnique.Intheridgeregressionm