fast condit densiti estim quantit structur dactiv relationship method quantit structur activ relationship qsar deliv point estim quantifi uncertainti inher predict quantifi uncertaini qsar predict predict condit densiti activ structur point estim condit densiti estim easi deriv predict interv activ paper experiment evalu compar method condit densiti estim suitabl qsar model contrast tradit method condit densiti estim base gener machin learn scheme specif class probabl estim experi show kernel estim base class probabl estim random forest classifi highli competit gaussian process regress take fraction time train gener machin learn base method condit densiti estim good fast option quantifi uncertainti qsar model integ linear program model constrain cluster equat discoveri model identif respiratori mechan mechan ventil human lung mine class dcorrel pattern sequenc label adapt transfer distanc measur quantit structur dactiv relationship latent structur pattern mine onlin structur graph cluster frequent subgraph mine pitfal supervis featur select predict biodegrad product pathwai hybrid knowledg machin learn dbase approach studi hierarch flat classif protein abstract automat classif protein machin learn import problem receiv signific attent literatur featur problem expert defin hierarchi protein class exist potenti exploit improv classif perform articl investig empir case hierarchi compar multi class classif techniqu exploit inform class hierarchi logist regress decis tree bag decis tree support vector machin underli base learner compar hierarch flat variant ensembl nest dichotomi shown deliv strong classif perform multi class set present experiment result synthet fold recognit enzym classif remot homolog detect data result show exploit class hierarchi improv perform synthet data case protein classif problem base recommend strong flat multi class method baselin establish benefit exploit class hierarchi area index term protein classif hierarch classif multi class classif interpret pet scan structur patient data data mine case studi dementia research predict mechan lung paramet gaussian process model data deffici inform dtheoret test select subgroup discoveri test select approach applic breast cancer diagnosi find relat associ hiv resist mutat data larg dscale graph mine backbon refin class enhanc navig biomed databas commun vote databas ddriven text classif kernel dbase induct transfer abstract method induct transfer advantag knowledg previou learn task solv newli task context supervis learn task find suitabl bia dataset set dataset paper kernelbas approach induct transfer aim find suitabl kernel data setup kernel linear span set predefin kernel find kernel appli convex optim level base level propos iter procedur gener kernel gener dataset meta level combin kernel minim criterion predict suitabl kernel data criterion base meta kernel captur similar dataset experi small molecul text data kernel base induct transfer show statist signific improv individu kernel case sindbad siql induct databas queri languag relat model induct databas queri languag relat model interpret pet scan structur patient data data mine case studi dementia research data ddriven extract rel reason rule limit combinatori explos biodegrad pathwai predict margin dbase dorder rule learn learn set order rule long tradit machin learn induct logic program tradit system follow separateand conquer approach modern system base statist consider ensembl theori larg margin classif graphic model induct logic program gene regul predict central goal comput system biologi understand mechan gene transcript regul system wide level effort base high throughput genom data model organ cerevisia goal work learn model gene regul optim featur set structur data abstract choos suitabl featur represent structur data trivial task due vast number potenti candid ideal pick small inform set structur featur provid complementari inform instanc frame search suitabl featur set combinatori optim problem purpos defin score function favor featur dissimilar featur score stochast local search sl procedur maxim divers featur set experi small molecul data investig effect forward select approach linear classif scheme leverag chemic background knowledg predict growth inhibit statist approach rule learn present statist approach rule learn address problem inher tradit rule learn comput hard find rule set low train error capac control avoid overfit chosen represent involv weight attach rule optim error rate directli optim rule set larg margin low varianc formul convex optim problem allow effici comput represent optim procedur effect yield weight claus cnflike represent avoid overfit propos model select strategi util concentr inequ empir test show system competit exist rule learn algorithm flexibl learn bia adjust improv predict accuraci consider induct logic program gene regul predict central goal comput system biologi understand mechan gene transcript regul system wide level effort base high throughput genom data model organ cerevisia goal work learn model gene regul margin dbase dorder rule learn learn set order rule long tradit machin learn induct logic program tradit system follow separateand conquer approach modern system base statist consider ensembl theori larg margin classif graphic model optim string mine frequenc constraint abstract propos algorithm framework solv frequenc relat data mine queri databas string optim time time linear input output size addit space linear input size framework mine frequent string emerg string string pass statist test test contrast present result string optim algorithm pattern domain itemset kei approach recent result index structur string suffix lcp arrai preprocess scheme rang minimum queri advantag arrai base data structur compar dynam data structur tree good local behavior extens secondari memori test algorithm real world data comput biologi demonstr approach work practic learn predict model growth inhibit nci dtp human tumor cell line screen data gene express make differ address problem learn predict model growth inhibit nci dtp human tumor cell line screen data extend classic quantit structur activ relationship paradigm investig includ gene express data lead statist signific improv predict qualiti analysi show straightforward approach includ individu gene express featur necessarili improv contrari degrad perform significantli gene express inform aggreg instanc featur repres correl refer cell line perform improv significantli improv expect learn task structur group featur instanc learn predict model growth inhibit nci dtp human tumor cell line screen data gene express make differ address problem learn predict model growth inhibit nci dtp human tumor cell line screen data extend classic quantit structur activ relationship paradigm investig includ gene express data lead statist signific improv predict qualiti analysi show straightforward approach includ individu gene express featur necessarili improv contrari degrad perform significantli gene express inform aggreg instanc featur repres correl refer cell line perform improv significantli improv expect learn task structur group featur instanc represent protein secondari structur predict base frequent pattern introduct special issu multi drelat data mine statist relat learn induct logic program analyz microarrai data quantit associ rule fast frequent string mine suffix arrai extend function point analysi object dorient requir specif induct databas relat model data bridg abstract present comprehens approach induct databas relat model main contribut induct queri languag extend sql goal support knowledg discoveri process pre process data mine post process prototyp system support queri languag develop sindbad structur induct databas develop project set model focus distanc base instanc base method closur easili achiev scenario area gene express data analysi demonstr power simplic concept hope preliminari work bring fundament issu integr pattern domain data mine techniqu attent induct databas commun ensembl balanc nest dichotomi multi dclass problem abstract system nest dichotomi hierarch decomposit multi class problem class classproblem repres tree structur ensembl randomlygener nest dichotomi proven effect approach multi class learn problem sampl tree give tree equal probabl mean depth tree limit number class unbalanc tree neg affect runtim paper investig approach build balanc nest dichotomi class balanc nest dichotomi data balanc nest dichotomi evalu ensembl set decis tree base model show approach reduc runtim effect accuraci problem class investig effect cach model build ensembl nest dichotomi quantit associ rule base half dspace optim approach ensembl nest dichotomi multi dclass problem nest dichotomi standard statist techniqu tackl polytom classif problem logist regress repres binari tree recurs split multi class classif task system dichotomi provid statist sound appli class learn algorithm multi class problem assum algorithm gener class probabl estim candid tree problem standard approach choic tree base domain knowledg practic altern treat system nest dichotomi equal form ensembl classifi base assumpt show approach produc accur classif appli logist regress directli multi class problem result show ensembl nest dichotomi produc accur classifi pairwis classif techniqu compar result logist regress compar error correct output code prefer logist regress compar case addit benefit gener class probabl estim good generalpurpos method appli binari classifi multi class problem tight bound rule learn lot empir evid show tradit rule learn approach work practic imposs deriv analyt result predict accuraci paper investig rule learn theoret perspect show applic mcallest pac bayesian bound rule learn yield practic learn algorithm base ensembl weight rule set experi result learn algorithm show competit state art rule learner error rate bound tightli fact bound turn tighter bound practic learn scheme set cover machin final prove bound improv allow learner abstain uncertain predict frequent free tree discoveri graph data recent year research graph mine explor linear path subgraph pattern languag paper investig middl ground extrem mine free unroot tree graph data motiv upgrad linear path pattern avoid complex issu subgraph pattern start complex consider defin free tree canon form present freetreemin algorithm make effici canon form search experi dataset nation cancer institut development therapeut program dtp anti hiv anti cancer screen data report coupl cohes metric knowledg dbase system frame rule experi predict biodegrad present applic induct logic program ilp area quantit structur activ relationship qsar activ predict biodegrad chemic compound water target variabl half life water aerob aqueou biodegrad structur descript chemic term atom bond deriv chemic smile encod definit substructur background knowledg predict biodegrad essenti regress problem discret version target variabl emploi number relat classif regress method relat represent compar proposit method appli propositionalis problem expert comment induc theori introduct persist chemic environ environmen data mine machin learn techniqu identif mutagen induc substructur structur activ relationship noncongener compound metamodel dcomplianc check requir semiform represent check requir highli desir hard achiev practic word processor project case review mean formal represent varieti autom check formal represent requir rare real world project semiform represent easier obtain offer possibl autom check base object orient hypertext represent present implement approach complianc check metamodel stochast local search dterm dnf learn nativ stochast local search algorithm solv term dnf problem present evalu hard term dnf problem lie phase transit compar perform gsat walksat type algorithm sat encod term dnf problem evalu state art separ conquer algorithm problem final demonstr practic relev algorithm chess endgam databas gener version space tree introduc gener version space tree data structur serv condens represent induct databas graph mine gener version space tree comfort represent version space natur effici process induct queri oper version space focu gener version space tree field mine free unroot tree graph databas experi data set nation cancer institut development therapeut program dtp gener version space tree boost induct queri consider featur moder increas space requir comparison version space represent base boundari set discov structur signatur protein fold base logic hidden markov model kristian kerst tapani raiko stefan kramer luc de raedt machin learn lab comput scienc depart univers freiburg georg koehler all freiburg germani laboratori comput inform scienc helsinki univers technolog box helsinki fin hut finland keyword hidden markov model probabilist model order logic structur signatur scop grow number determin protein structur avail classif scheme increasingli import develop comput method automat extract structur signatur class protein present machin learn techniqu logic hidden markov model lohmm appli task find structur signatur fold classif scheme scop survei predict toxicolog challeng statist evalu predict toxicolog challeng motiv develop silico model predict chemic carcinogenesi molecular structur greatli prevent environment caus cancer predict toxicolog challeng ptc competit wasorgan test state art appli machin learn form predict model phase transit stochast local search dterm dnf learn past decad lot interest phase transit artifici intellig recent machin learn induct logic program investig phase transit learn term dnf boolean formula practic relev class concept show exist phase transit character locat phase transit paramet number posit neg exampl number boolean variabl subsequ investig stochast local search sl term dnf learn compar variant reduc term dnf sat appli sl algorithm gsat walksat experi walksat solv largest fraction hard problem instanc transform dbase regress paper introduc transform base regress tbr rule base symbol regress techniqu base transform base learn tbl transform base learn introduc coupl year ago consid regress type task propos method learn exampl neighborhood relat depend variabl depend neighbor exampl method potenti learn sequenc spatial data paper demonstr capabl limit approach highli complex real world domain musicolog speech synthesi analysi respiratori pressur dvolum curv intens care medicin induct machin learn present case studi machin learn data mine intens care medicin studi compar method measur pressur volum curv artifici ventil patient suffer adult respiratori distress syndrom ard levelwis version space algorithm applic molecular fragment find tight integr mitchel version space algorithm agraw al apriori algorithm present algorithm gener pattern satisfi varieti constraint data constraint impoes featur construct version space biochem applic levelwis version space algorithm applic molecular fragment find tight integr mitchel version space algorithm agraw al apriori algorithm present algorithm gener pattern satisfi varieti constraint data constraint impoes demand ddriven construct structur featur ilp paper tackl problem method proposition featur construct order logic date construct featur unspecif construct featur demand advanc detect represent chang structur featur requir current method construct featur goal direct fashion previou work present molecular featur mine hiv data predict toxicolog challeng motiv develop silico model predict chemic carcinogenesi molecular structur greatli prevent environment caus cancer predict toxicolog challeng ptc competit wasorgan test state art appli machin learn form predict model predict ordin class regress tree paper devot problem learn predict ordin order discret class classif regress tree start cart tree induct algorithm studi wai transform learner ordin classif task algorithm variant compar number benchmark data set verifi rel strength weak strategi studi trade optim categor classif accuraci hit rate minimum distanc base error preliminari result promis avenu algorithm combin aspect classif regress learn oper advic address problem advic take domain build game plai program approach solv strive applic machin learn techniqu avoid knowledg elicit mean build exist work operation advic machin assum advic oper form rel import advic util program paper present approach determin rel import situat reinforc learn implement approach game heart gather empir evid us experi result show program built approach learn make good oper advic learn oper advic address problem advic take domain build game plai program approach solv strive applic machin learn techniqu avoid knowledg elicit mean build exist work operation advic machin assum advic oper form rel import advic util program paper present approach determin rel import situat reinforc learn implement approach game heart gather empir evid us experi result show program built approach learn make good oper advic bottom dup proposition paper present method proposition work bottom data driven manner tailor biochem databas exampl descript chemic compound method gener frequent fragment linearli connect atom user speci length preliminari experi domain carcinogen predict show bottomup proposition promis approach featur construct relat data introduct approach handl relat learn problem transform proposit problem subsequ appli proposit learn algorithm transform requir construct featur captur relat properti learn exampl featur construct model construct decoupl transform relat learn problem proposit call proposition predict ordin class regress tree paper devot problem learn predict ordin order discret class classif regress tree start cart tree induct algorithm studi wai transform learner ordin classif task algorithm variant compar number benchmark data set verifi rel strength weak strategi studi trade optim categor classif accuraci hit rate minimum distanc base error preliminari result promis avenu algorithm combin aspect classif regress thesi relat learn proposition semiautomat gener glossari link practic solut paper appli linguist theori relev studi link work insist lyric qualiti link interpret activ argu pragmat approach understand hypertext reader behavior author toolbuild alik experi predict biodegrad present applic induct logic program ilp area quantit structur activ relationship qsar activ predict biodegrad chemic compound water target variabl half life water aerob aqueou biodegrad structur descript chemic term atom bond deriv chemic smile encod definit substructur background knowledg predict biodegrad essenti regress problem discret version target variabl emploi number relat classif regress method relat represent compar proposit method appli propositionalis problem expert comment induc theori introduct persist chemic environ environmen combin structur search content search world dwide web case studi decompos function requir scenario stochast proposition ddetermin background knowledg proposit relat learn algorithm requir good represent perform practic represent engin manual domain expert deriv automat mean call construct induct induct logic program ilp algorithm put burden data engin effort structur relat represent background knowledg chemic engin domain common represent devic graph structur call determin relat manual engin featur domain typic test count occurr specif substructur specif properti represent determin relat pose effici problem standard ilp algorithm devis stochast algorithm automat deriv featur determin background knowledg algorithm conduc integr forward revers object dorient softwar engin benefit metric kb develop mine cancer machin learn experi level detail paper present methodolog point view result interdisciplinari project scientif data mine analyz data carcinogen chemic deriv carcinogenesi bioassai program long term research studi perform nation institut environment health scienc databas detail descript test perform compound anim speci strain sex chemic structur atom bond level term relev structur properti goal paper investig effect level detail amount inform result hypothes quantit qualit appli relat proposit machin learn algorithm learn problem formul regress classif task addit experi conduct learn structur regress tree real world domain task machin learn algorithm learn theori predict numer valu standard test domain induct logic program ilp concern predict numer valu exampl relat determin background knowledg ilp algorithm predict number cope determin background knowledg except cover algorithm call for paper present structur regress tree srt algorithm appli class problem integr statist method regress tree ilp srt construct tree liter atom formula negat conjunct liter node assign numer leaf srt comprehens result pure statist method appli class problem ilp syst effici search strong partial determin work offer solut problem find function depend distort nois open problem effici find strong highli compress partial determin se briefli introduc restrict form search partial determin base function depend focus attent sole partial determin deriv overfit function depend enabl effici search strong partial determin gener compress base measur evalu partial determin valu attribut applic real world data suggest restrict search retriev subset strong partial determin shorter runtim show feasibl us approach introduct function depend mannila amp raiha fundament form knowledg discov databas real world databas compress dbase evalu partial determin work tackl problem find partial determin databas propos compress base measur evalu partial determin view gener function depend associ rule relat natur except extend measur evalu associ rule support confid partial determin lead problem propos measur base minimum descript length mdl principl remedi problem assum hypothet task transmit databas effici measur estim compress achiev transmit partial determin origin data take account complex correct partial determin avoid overfit presenc nois describ kind search ne 