softwar effort predict regress rule extract neural network neural network rule extract led displai recognit problem understand consum heterogen busi intellig applic neural network note knowledg discoveri neural network applic credit card screen predict consum prefer fast dfood franchis data mine approach market research applic artifici neural network rule extract support vector machin overview issu applic credit score improv som label methodolog data mine applic recurs neural network rule extract data mix attribut minerva sequenti cover rule extract greedi rule gener discret data neural network rule extract greedi rule gener discret data neural network rule extract knowledg discoveri implement busi intellig approach neural network rule extract decis tabl separ core noncor knowledg applic neural network rule extract cross dnation studi brand imag percept extract salient dimens automat som label hybrid som dsvm method analyz zebra fish gene express appli conjug gradient method text document categor approach gener rule neural network regress problem visual global dorgan map approach custom profil neural network rule extract decis tabl credit risk evalu compar studi centroid dbase effect queri size estim neural network gener concis set linear regress rule artifici neural network input select data ddriven fuzzi model iter backward select method determin relev input variabl data driven fuzzi model present method util paramet takagi sugeno model factor determin signific input variabl result computation intens exist method input variabl select build credit drisk evalu expert system neural network rule extract decis tabl effect method gener multipl linear regress rule artifici neural network gener linear regress rule neural network local squar approxim feedforward neural network construct cross valid problem determin architectur multilay perceptron disadvantag standard backpropag algorithm direct research algorithm determin weight structur network learn data propos construct algorithm multipl oper statist test determin architectur network construct multipl hidden layer multipl hidden unit layer algorithm node remov addit layer addit determin number node layer heurist appli statist test compar architectur result promis optim feedforward neural network construct cross valid problem determin architectur multilay perceptron disadvantag standard backpropag algorithm direct research algorithm determin weight structur network learn data propos construct algorithm multipl oper statist test determin architectur network construct multipl hidden layer multipl hidden unit layer algorithm node remov addit layer addit determin number node layer heurist appli statist test compar architectur result promis optim rapid model acquisit imag small object open neural network black box algorithm extract rule function approxim artifici neural network neural network prune function approxim neural network successfulli appli solv varieti applic problem includ classif function approxim function approxim requir prior knowledg input data distribut shown univers approxim applic desir extract knowledg explain problem solv network exist approach focus extract symbol rule classif method devis extract rule train neural network regress articl present approach extract rule train neural network regress rule extract rule set correspond subregion input space linear function involv relev input attribut data approxim network output data sampl subregion extens experiment result benchmark data set demonstr effect propos approach gener accur regress rule prune neural network regress neural network wide tool regress capabl approxim function requir assumpt distribut data commonli architectur regress feedforward neural network hidden layer paper present network prune algorithm determin number unit input hidden layer network compar perform prune network regress method linear regress lr naiv bay nb nearest neighbor knn decis tree predictor publicli data set test neural network method outperform nb knn predict error comput term root squar error measur metric perform lr hand absolut error measur metric neural network metho fernn algorithm fast extract rule neural network symbol rule extract train neural network network prune obtain concis rule typic prune algorithm requir retrain network incur addit cost paper present fernn fast method extract rule train neural network network retrain fulli connect train feedforward network singl hidden layer fernn identifi relev hidden unit comput inform gain relev hidden unit activ valu divid subinterv inform gain maxim fernn find set relev network connect input unit hidden unit check magnitud weight connect larg weight identifi relev final fernn gener rule distinguish subinterv hidden activ valu term network input experiment result show gener concis accur classif rule breast cancer diagnosi previou work present algorithm extract classif rule train neural network discuss applic breast cancer diagnosi paper describ accuraci network accuraci rule extract improv simpl pre process data data pre process involv select relev input attribut remov sampl miss attribut valu rule gener neural network rule extract algorithm concis accur gener rule gener method report literatur keyword neural network rule extract wisconsin breast cancer diagnosi data pre process attribut select introduct neural network shown valuabl tool pattern classif area medic field input pattern neural network predict norm comparison neural network rule extract techniqu diagnosi hepatobiliari disord neural network wide tool predict medicin expect applic neural network medic diagnosi recent develop neural network rule extract algorithm make decis process train network express classif rule rule comprehens human user classif process network involv complex nonlinear map input data paper report result neural network rule extract techniqu neurolinear neurorul appli diagnosi hepatobiliari disord data set consist measur collect patient japanes hospit measur continu valu neurolinear gener piec wise linear discrimin function data set continu measur previous discret domain expert neurorul appli discret data connectionist approach gener obliqu decis tree map decis tree neural network exist method transform decis tree neural network method typic construct network directli map decis node rule neural unit result network construct larger paper describ prune base method map decis tree neural network compress network remov unimport redund unit connect addit equival decis tree extract prune network simpler induc algorithm id keyword decis tree neural network prune introduct decis tree wide nonparametr pattern classif task involv pattern class larg number featur input pattern tree classifi perform comparison stipul decis node tree branch left subtre base explan virtual input phenomenon obviou activ corner unit synchron likewis side unit result train network produc train output figur train sampl consist bit input theta output pattern pattern pattern gener train network connect prune pattern depend hidden unit activ hidden layer real number binari unit network gener train output pattern connect weight equal nomin valu gamma network gener train output pattern long weight close nomin valu closer weight nomin valu higher frequenc gen featur transform multivari decis tree induct univari decis tree udt inher problem replic repetit fragment multivari decis tree mdt propos overcom problem close examin convent wai build mdt reveal fragment problem persist approach suggest minim fragment problem separ hyperplan search decis tree build achiev featur transform initi featur vector featur vector featur transform obtain mdt build udt replac featur node combin initi featur elabor advantag approach detail expect perform experi conduct order confirm analysi result compar oc cart introducti increment featur select featur select problem find relev featur number featur dataset larg number pattern huge effect method featur select dimension reduct increment probabilist algorithm design implement altern exhaust heurist approach theoret analysi support idea probabilist algorithm find optim optim subset featur experiment result suggest probabilist algorithm effect obtain optim suboptim featur subset increment version expedit featur select number pattern larg scale sacrif qualiti select featur keyword pattern recognit machin learn featur select dimension reduct appli intellig kluwer academ publish introduct featur select analysi hidden represent greedi cluster neurolinear system extract obliqu decis rule neural network present neurolinear system extract obliqu decis rule neural network train classif pattern condit obliqu decis rule correspond partit attribut space hyperplan necessarili axi parallel allow set hyperplan form boundari decis region lead signific reduct number rule gener maintain accuraci rate network describ compon neurolinear detail exampl artifici dataset experiment result real world dataset show system effect extract compact comprehens rule high predict accuraci neural network keyword rule extract obliqu rule prune discret introduct neural network wide appli solv classif problem comparison neural network decis tre featur select discret discret turn numer attribut discret featur select elimin irrelev redund attribut chi simpl gener algorithm statist discret numer attribut repeatedli inconsist found data achiev featur select discret handl mix attribut work multiclass data remov irrelev redund attribut keyword discret featur select pattern classif introduct featur select elimin irrelev redund attribut relev featur classif algorithm gener improv predict accuraci shorten learn period form simpler concept abund featur select algorithm method principl compon compos smaller number featur select subset origin attribut paper consi penalti dfunction approach prune feedforward neural network paper propos penalti function prune feedforward neural network weight elimin penalti function propos consist term term discourag unnecessari connect term prevent weight connect take excess larg valu simpl criteria elimin weight network effect penalti function test problem test problem contigu problem pariti problem monk problem result prune network obtain problem fewer connect previous report literatur introduct concern paper find minim feedforward backpropag neural network solv problem distinguish pattern set dimension space backpropag feedforward neural network gain ac extract rule neural network prune hidden dunit split algorithm extract rule standard layer feedforward neural network propos train network prune remov redund connect network importantli detect relev input algorithm gener rule prune network small number activ valu hidden unit number input connect hidden unit suffici small rule describ activ valu obtain readili gener hidden unit split treat output unit output unit activ hidden layer insert subnetwork form train prune process repeat hidden unit network small number input unit connect exampl propos algorithm work shown real world data aris molecular bio solut pariti problem singl hidden layer feedforward neural network bit pariti problem solvabl standard feedforward neural network singl hidden layer consist hidden unit hidden unit odd network direct connect input layer output layer transfer function hidden unit output unit usual sigmoid function oe exp gammax show solut easili obtain solv system linear equat keyword pariti problem feedforward network sigmoid function introduct pariti problem wide problem test neural network train algorithm problem map problem domain set consist distinct bit binari vector result map number vector problem consid hard output sing neurolinear neural network obliqu decis rule present neurolinear system extract obliqu decis rule neural network train classif pattern condit obliqu decis rule correspond partit attribut space hyperplan necessarili axi parallel allow set hyperplan form boundari decis region lead signific reduct number rule gener maintain accuraci rate network describ compon neurolinear detail exampl artifici dataset experiment result real world dataset show system effect extract compact comprehens rule high predict accuraci neural network keyword rule extract obliqu rule prune discret introduct neural network wide appli solv classif problem comparison neural network decis tre probabilist approach featur select filter solut featur select defin problem find minimum set relev attribut describ dataset origin attribut examin problem exhaust heurist approach featur select paper propos probabilist approach theoret analysi experiment studi show propos approach simpl implement guarante find optim resourc permit fast obtain result effect select featur improv perform learn algorithm site applic involv huge dataset conduct independ prove effect scalabl propos algorithm discuss aspect applic featur select algorithm introduct problem featur select defin find relev attribut origin attrib featur select classif probabilist wrapper approach featur select defin problem find minimum set featur induct algorithm achiev highest predict accuraci data origin featur probabilist wrapper model propos method exhaust search heurist approach aim model avoid local minima exhaust search highest predict accuraci criterion search smallest analysi experi show model effect find relev featur remov irrelev context improv predict accuraci induct algorithm simpl straightforward provid fast solut search optim applic model futur work relat issu discuss introduct problem featur select defin find relev featur origin attribut improv backpropag learn featur select exist redund irrelev noisi data proper data train network speed train simplifi learn structur improv perform phase train algorithm propos phase number input unit network determin inform base method attribut meet criteria inclus consid input network phase number hidden unit network select automat base perform network train data hidden unit ad time experiment result show algorithm achiev faster learn time simpler network improv perform improv backpropag learn featur select exist redund irrelev noisi data proper data train network speed train simplifi learn structur improv perform phase train algorithm propos phase number input unit network determin inform base method attribut meet criteria inclus consid input network phase number hidden unit network select automat base perform network train data hidden unit ad time experiment result show algorithm achiev faster learn time simpler network improv perform extract rule prune network breast cancer diagnosi extract rule prune network breast cancer diagnosi symbol represent neural network backpropag neural network gener predict decis tree pattern classif problem regard black box predict explain decis tree applic explicit knowledg need human expert work deriv symbol represent neural network make epxlicit predict network algorithm propos implement extract symbol rule neural network effect data mine neural network classif data mine problem receiv great attent recent databas commun paper present approach discov symbol classif rule neural network neural network thought suit data mine classif made explicitli state symbol rule suitabl verif interpret human propos approach concis symbol rule high accuraci extract neural network network train achiev requir accuraci rate redund connect network remov network prune algorithm activ valu hidden unit network analyz classif rule gener result analysi effect propos approach demonstr experiment result set standard data mine test problem introduct dimension reduct discret exist numer data larg amount record databas pose challeng task explicit concept extract raw data paper introduc method reduc data vertic horizont discrimin power origin data pave extract concept method base discret vertic reduct featur select horizont reduct experiment result show data effect reduc propos method predict accuraci classifi improv data dimension reduct classif rule learn simpler kei word dimension reduct discret knowledg discoveri introduct wide comput bring prolifer databas aid comput flood raw data exploit human knowledg discoveri system databas understand neural network rule extract backpropag neural network gener predict decis tree pattern classif problem regard black box predict interpret decis tree paper argu proper techniqu enabl algorithm extract rule draw parallel decis tree show predict network explain rule extract network understood experi demonstr rule extract neural network compar decis tree term predict accuraci number rule averag number condit rule preserv high predict accuraci origin network effect classfic rule extract abstract neurorul connectionist approach data mine classif involv find rule partit data set disjoint group class data mine problem approach propos mine classif rule larg databas decis tree base symbol learn method connectionist approach base neural network thought suit data mine major reason cite knowledg gener neural network explicitli repres form rule suitabl verif interpret human paper examin issu newli develop algorithm rule similar concis gener symbol method extract neural network data mine process neural network emphasi rule extract experiment result comparison previous publish work present neural network construct algorithm maxim likelihood function effici neural network train crai dmp neural network construct algorithm applic imag 