tutori summari reduct machin learn exploit context deal uncertainti classif guest editori special issu util dbase data mine high dquantil model custom wallet estim applic paper discuss import practic problem custom wallet estim estim potenti spend custom expect spend purpos util quantil model goal estim quantil discrimin condit distribut respons implicit goal standard regress approach argu notion wallet captur high quantil model estim percentil describ wallet estim implement ibm market align program map discuss wide rang domain high quantil model practic import estim opportun sale market domain defin surpris pattern outlier fraud detect survei exist approach quantil model propos adapt nearest neighbor regress tree approach quantil model demonstr model perform high quantil estim domain includ motiv problem estim realist wallet ibm custom rank dbase evalu regress model secur coprocessor privaci preserv collabor data mine analysi secur coprocessor privaci preserv outlier detect activ learn predict condit quantil reduct classif show reduc process predict condit quantil median solv classif accompani theoret statement show regret classifi bound regret quantil regress quantil loss test reduct empir exist quantil regress method larg real world dataset discov state theart perform ubdm util dbase data mine workshop report weight dagainst dall tti chicago abstract reduct convert solver task solver task introduc notion reduct analyz reduct classif task introduct natur supervis learn task classif varieti form binari multiclass cost sensit import weight approach analyz task independ devis algorithm approach reduct convert unexplor task task analyz automat transfer exist theori algorithm understood domain domain investig approach classif problem reduct power tool number desir properti show similar task task reduc busi perform manag system crm sale execut rank dbase evalu regress model improv categor classifi sensit sampl select bia recent paper categor classifi learn algorithm sensit common type sampl select bia chanc select train sampl depend featur vector directli class label classifi learner categor local insensit type sampl select bia consid global paper true model distinguish model algorithm output discuss bayesian classifi logist regress hard margin support vector machin true model model gener true class label implicitli assum contain model space learner true class probabl model estim class probabl assum asymptot converg train data set size increas discuss naiv bay decis tree soft margin support vector machin model space assum true model classif algorithm argu global learner argu classifi learn algorithm includ discuss affect sampl select bia depend dataset heurist induct bia impli classifi learn algorithm appropri dataset make earlier experiment result produc addit result illustr claim relat reinforc learn perform classif perform error limit reduct classif task introduc reduct base model analyz supervis learn task model devis reduct multi class cost sensit classif binari classif guarante learn binari classifi error rate cost sensit classifi cost time expect sum cost labl cost sensit classif emb bound loss finit choic supervis learn task result show task solv binari classif oracl final present experiment result show reduct outperform exist algorithm multi class cost sensit learn report ubdm workshop util dbase data mine learn evalu classifi sampl select bia classifi learn method commonli assum train data consist randomli drawn exampl distribut test exampl learn model expect make predict iter method multi dclass cost dsensit learn cost sensit learn address issu classif presenc vari cost type misclassif paper present method solv multi class cost sensit learn problem binari classif algorithm algorithm deriv kei idea iter weight expand data space gradient boost stochast ensembl establish theoret guarante perform method show variant possess boost properti form weak learn assumpt compon binari classifi empir evalu perform propos method benchmark data set verifi method gener achiev result repres method cost sensit learn term predict perform cost minim case comput effici empir comparison reinforc learn strategi sequenti target market empir evalu perform reinforc learn method applic sequenti target market propos evalu progress reinforc learn method rang direct batch method indirect simul base method call semidirect method fall conduct number control experi evalu perform compet method result indirect method perform situat perfect model realist situat system model paramet restrict attent indirect method perform tend degrad show semi direct method effect reduc amount comput attain level perform result profit polici sequenti cost dsensit decis make reinforc learn recent increas interest issu cost sensit learn decis make varieti applic data mine number approach develop effect optim cost sensit decis decis consid isol issu sequenti decis make goal maxim total benefit accru period time benefit rare address present paper propos approach sequenti decis make base reinforc learn framework approach attempt learn decis rule optim sequenc cost sensit decis maxim total benefit accru time domain target market testb empir evalu propos method conduct experi approxim year monthli promot data deriv kdd cup donat data set experiment result show propos method optim total accru benefit outperform usual target market methodolog optim promot isol analyz behavior target rule obtain discuss appropri applic domain transform classifi score accur multiclass probabl estim class membership probabl estim import applic data mine classif output combin sourc inform decis make depend misclassif cost output classifi domain knowledg previou calibr method appli class problem show obtain accur probabl estim multiclass problem combin calibr binari probabl estim propos method obtain calibr class probabl estim appli classifi produc rank exampl naiv bay support vector machin classifi give experiment result varieti class multiclass domain includ direct market text categor digit recognit obtain calibr probabl estim decis tree naiv bayesian classifi accur calibr estim class membership probabl need supervis learn applic cost sensit decis made exampl depend cost paper present simpl success method obtain calibr probabl estim decis tree naiv bayesian classifi larg challeng kdd contest dataset testb report result detail experiment comparison ten method evalu measur conclud bin succe significantli improv naiv bayesian probabl estim improv decis tree probabl estim recommend smooth estim variant prune call curtail 