Comprehensible credit scoring models using rule extraction from support vector machines

《Comprehensible credit scoring models using rule extraction from support vector machines》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、ARTICLEINPRESSEuropeanJournalofOperationalResearchxxx(2007)xxx–xxxwww.elsevier.com/locate/ejorComprehensiblecreditscoringmodelsusingruleextractionfromsupportvectormachinesa,*b,ac,daDavidMartens,BartBaesens,TonyVanGestel,JanVanthienenaDepartmentofDeci

2、sionSciencesandInformationManagement,K.U.LeuvenNaamsestraat69,B-3000Leuven,BelgiumbSchoolofManagement,UniversityofSouthampton,HighfieldSouthampton,SO171BJ,UnitedKingdomcBaselIIModelling,RiskManagement,DexiaGroupPlaceRogier11,1210Brussels,BelgiumdDepar

3、tmentofElectricalEngineering,ESAT-SCD-SISTA,K.U.LeuvenKasteelparkArenberg10,B-3001Leuven(Heverlee),BelgiumReceived1October2005;accepted1April2006AbstractInrecentyears,supportvectormachines(SVMs)weresuccessfullyappliedtoawiderangeofapplications.Howeve

4、r,sincetheclassifierisdescribedasacomplexmathematicalfunction,itisratherincomprehensibleforhumans.Thisopacitypropertypreventsthemfrombeingusedinmanyreal-lifeapplicationswherebothaccuracyandcomprehensibilityarerequired,suchasmedicaldiagnosisandcreditri

5、skevaluation.Toovercomethislimitation,rulescanbeextractedfromthetrainedSVMthatareinterpretablebyhumansandkeepasmuchoftheaccuracyoftheSVMaspossible.Inthispaper,wewillprovideanoverviewoftherecentlyproposedruleextractiontechniquesforSVMsandintroducetwoo

6、therstakenfromtheartificialneuralnetworksdomain,beingTrepanandG-REX.Thedescribedtechniquesarecomparedusingpub-liclyavailabledatasets,suchasRipley’ssyntheticdatasetandthemulti-classirisdataset.Wewillalsolookatmedicaldiag-nosisandcreditscoringwherecompr

7、ehensibilityisakeyrequirementandevenaregulatoryrecommendation.OurexperimentsshowthattheSVMruleextractiontechniquesloseonlyasmallpercentageinperformancecomparedtoSVMsandthereforerankatthetopofcomprehensibleclassificationtechniques.Ó2006ElsevierB.V.Allr

8、ightsreserved.Keywords:Creditscoring;Classification;Supportvectormachine;Ruleextraction1.Introductionscoring[2],financialtimeseriesprediction[14],spamcategorization[9]andbraintumorclassifica-Supportvectormachinesareastate-of-thearttion[19].Thestrengthof