optim high probabl algorithm contextu bandit problem agnost activ learn constraint error dcorrect tournament abstract present famili adapt pairwis tournament provabl robust larg error fraction determin largest element set tournament nk pairwis comparison log depth number player robust paramet reason valu tournament give reduct multiclass binari classif machin learn yield analysi problem import weight activ learn present practic statist consist scheme activ learn binari classifi gener loss function algorithm import weight correct sampl bia control varianc give rigor label complex bound learn process offset tree learn partial label error dcorrect tournament abstract present famili adapt pairwis tournament provabl robust larg error fraction determin largest element set tournament nk pairwis comparison log depth number player robust paramet reason valu tournament give reduct multiclass binari classif machin learn yield analysi problem condit probabl tree estim analysi algorithm main result problem estim condit probabl label time log number label analyz natur reduct problem set binari regress problem organ tree structur prove regret bound scale depth tree motiv analysi propos onlin algorithm provabl construct logarithm depth tree set label solv problem test algorithm empir show work succesfulli dataset roughli label agnost activ learn state analyz activ learn algorithm work presenc arbitrari form nois algorithm agnost activ reli assumpt sampl drawn fix distribut show achiev exponenti improv requir ln sampl find optim classifi usual sampl complex supervis learn set consid realiz case includ learn threshold classifi learn homogen linear separ respect input distribut uniform unit sphere scalabl topolog discoveri link state detect rout event offset tree learn partial label offset tree learn partial label import weight activ learn present practic statist consist scheme activ learn binari classifi gener loss function algorithm import weight correct sampl bia control varianc give rigor label complex bound learn process robust reduct rank classif reduc rank measur area receiv oper characterist curv auc binari classif core theorem show binari classif regret induc binari problem impli auc regret larg improv approach order regress score regret transform nr number element evalu optim method network bottleneck diagnosi agnost activ learn state analyz activ learn algorithm work presenc arbitrari form nois algorithm agnost activ reli assumpt sampl drawn fix distribut show achiev exponenti improv requir ln sampl find optim classifi usual sampl complex supervis learn set consid realiz case includ learn threshold classifi learn homogen linear separ respect input distribut uniform unit sphere agnost activ learn state analyz activ learn algorithm work presenc arbitrari form nois algorithm agnost activ reli assumpt sampl drawn fix distribut show achiev exponenti improv requir ln sampl find optim classifi usual sampl complex supervis learn set consid realiz case includ learn threshold classifi learn homogen linear separ respect input distribut uniform unit sphere cover tree nearest neighbor present tree data structur fast nearest neighbor oper gener npoint metric space data set consist point data structur requir space metric structur maintain perform properti navig net krauthgam amp lee point set bound expans constant measur intrins dimension defin karger amp ruhl cover tree data structur construct log time nearest neighbor queri requir time logarithm log time experiment result show speedup brute forc search vari order magnitud natur machin learn dataset cover tree nearest neighbor present tree data structur fast nearest neighbor oper gener npoint metric space data set consist point data structur requir space metric structur maintain perform properti navig net krauthgam amp lee point set bound expans constant measur intrins dimension defin karger amp ruhl cover tree data structur construct log time nearest neighbor queri requir time logarithm log time experiment result show speedup brute forc search vari order magnitud natur machin learn dataset 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 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 sensit error correct output code 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 test dbase diagnosi tree matrix represent common problem encount applic scenario repres prior knowledg system order determin true state effici inform typic form test question system test potenti reduc uncertainti system state problem repres inform captur depend test outcom state effici navig aid diagnosi common represent flowchart leaf node state leaf node test state problem flowchart notori difficult maintain addit knowledg manual integr system make imposs track decis path optim flow maxim perform propos effici method optim exist flowchart base convers auxiliari matrix represent main goal paper show synergi represent hope practition choos strategi applic show convers suggest wai improv represent wai envis represent final show represent information equival sens transform black box state system statict model unequ space time seri effici test select activ diagnosi entropi approxim problem diagnos fault system repres bayesian network diagnosi correspond recov state unobserv node outcom test observ node find optim subset test set intract gener show difficult comput inform test greedi test select involv entropi term exact comput intract propos approxim approach util loopi belief propag infrastructur simultan comput approxim margin condit entropi multipl subset node appli method fault diagnosi comput network show algorithm effect realist internet topolog provid theoret justif greedi test select approach perform guarante enumer collaps nc investig altern notion approxim problem insid deter minist polynomi time show slightli nontrivi inform problem hard obtain solut prove elimin singl possibl arithmet circuit input impli class fast polygarithm time parallel solut word constitut proof inher sequenti problem result robust respect elimin procedur allow err exclud correct small probabl show fundament linear algebra problem hard sens turn hard substanti reduc number valu determin rank comput final show precis sens random nontrivi substitut nondetermin space believ random give constant factor advantag space determin power nondetermin show true restrict version probabilist logspac error potenti larger achiev amplif improv network robust import properti network system robust remov network node random node failur target attack design method propos creat ab initio network optim robust measur face exist network feasibl substanti modifi redesign robust improv lesser degre modif present empir result show robust measur size largest connect compon shortest path length pair node affect strategi alter network rewir fraction edg ad edg find modest alter initi scale free network usefulli improv robust attack fraction attack node small identifi modif scheme effect purpos enumer collaps nc investig altern notion approxim problem insid deter minist polynomi time show slightli nontrivi inform problem hard obtain solut prove elimin singl possibl arithmet circuit input impli class fast polygarithm time parallel solut word constitut proof inher sequenti problem result robust respect elimin procedur allow err exclud correct small probabl show fundament linear algebra problem hard sens turn hard substanti reduc number valu determin rank comput final show precis sens random nontrivi substitut nondetermin space believ random give constant factor advantag space determin power nondetermin show true restrict version probabilist logspac error potenti larger achiev amplif 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 enumer determin rank investig complex enum approxim elementari problem linear algebra comput rank determin matrix show exist enumer matrix output list constantli number guarante rank matrix determin ac oracl access enumer number rank enumer fl function problem comput rank fl result hold matric commut ring size grow polynomi size matrix exist enumer impli slightli stronger collaps exact count logspac hierarchi approxim probabl distribut question distribut approxim probabilist graphic model introduc paramet effect treewidth captur degre approxim tradeoff accuraci complex approxim enumer determin rank investig altern notion approxim problem insid deter minist polynomi time show slightli nontrivi inform problem hard obtain solut prove elimin singl possibl arithmet circuit input impli class fast polygarithm time parallel solut word constitut proof inher sequenti problem result robust respect elimin procedur allow err exclud correct small probabl show fundament linear algebra problem hard sens turn hard substanti reduc number valu determin rank comput final show precis sens random nontrivi substitut nondetermin space believ random give constant factor advantag space determin power nondetermin show true restrict version probabilist logspac error potenti larger achiev amplif enumer determin rank investig complex enum approxim elementari problem linear algebra comput rank determin matrix show exist enumer matrix output list constantli number guarante rank matrix determin ac oracl access enumer number rank enumer fl function problem comput rank fl result hold matric commut ring size grow polynomi size matrix exist enumer impli slightli stronger collaps exact count logspac hierarchi halv line arrang enumer determin rank investig complex enum approxim elementari problem linear algebra comput rank determin matrix show exist enumer matrix output list constantli number guarante rank matrix determin ac oracl access enumer number rank enumer fl function problem comput rank fl result hold matric commut ring size grow polynomi size matrix exist enumer impli slightli stronger collaps exact count logspac hierarchi fast order larg categor dataset visual fast order larg categor dataset visual dwai function worst dcase cryptographi algebra secur properti survei recent develop studi worst case function strong algebra secur properti rs line research initi rivest sherman design parti secretkei agreement protocol strongli noninvert total associ function kei build block commut ad ingredi protocol parti note rabi sherman rs develop digit signatur protocol base enhanc function recent open question function algebra secur properti protocol requir creat function recent hemaspaandra roth hr resolv open issu arm show function exist strong total commut associ function exist dwai function worst dcase cryptographi algebra secur properti hous 