machine-learning

MachineLearning北京邮电大学TheLearningProblemWhatisMachineLearningLearningandMachineLearninglearning:acquiringskillwithexperienceaccumulatedfromobservationsobservationslearningskillmachinelearning:acquiringskillwithexperienceaccumulated/computedfromdatadataskillMLWhatisskill?Hsuan-TienLin(NTUCSIE)MachineLearningFoundations6/27TheLearningProblemWhatisMachineLearningdataMLAMoreConcreteDefinitionskill⇔improvesomeperformancemeasure(e.g.predictionaccuracy)machinelearning:improvingsomeperformancemeasurewithexperiencecomputedfromdataimprovedperformancemeasureAnApplicationinComputationalFinancestockdataMLmoreinvestmentgainWhyusemachinelearning?Hsuan-TienLin(NTUCSIE)MachineLearningFoundations7/27TheLearningProblemWhatisMachineLearningYetAnotherApplication:TreeRecognition•‘define’treesandhand-program:difficult•learnfromdata(observations)andrecognize:a3-year-oldcandoso•‘ML-basedtreerecognitionsystem’canbeeasiertobuildthanhand-programmedsystemML:analternativeroutetobuildcomplicatedsystemsHsuan-TienLin(NTUCSIE)MachineLearningFoundations8/27PicturemachineIsadogornot?If(condition1&condition2&…&conditionN){print(“Thisisadog!”);}else{print(“Thisisnotadog!”);}If(condition1&condition2&…&conditionN){print(“Thisisadog!”);}else{print(“Thisisnotadog!”);}TheLearningProblemWhatisMachineLearningTheMachineLearningRouteML:analternativeroutetobuildcomplicatedsystemsSomeUseScenarios•whenhumancannotprogramthesystemmanually—navigatingonMars•whenhumancannot‘definethesolution’easily—speech/visualrecognition•whenneedingrapiddecisionsthathumanscannotdo—high-frequencytrading•whenneedingtobeuser-orientedinamassivescale—consumer-targetedmarketingGiveacomputerafish,youfeeditforaday;teachithowtofish,youfeeditforalifetime.:-)Hsuan-TienLin(NTUCSIE)MachineLearningFoundations9/27TheLearningProblemWhatisMachineLearningdataMLKeyEssenceofMachineLearningmachinelearning:improvingsomeperformancemeasurewithexperiencecomputedfromdataimprovedperformancemeasure123existssome‘underlyingpattern’tobelearned—so‘performancemeasure’canbeimprovedbutnoprogrammable(easy)definition—so‘ML’isneededsomehowthereisdataaboutthepattern—soMLhassome‘inputs’tolearnfromkeyessence:helpdecidewhethertouseMLHsuan-TienLin(NTUCSIE)MachineLearningFoundations10/27TheLearningProblemWhatisMachineLearningFunTimeWhichofthefollowingisbestsuitedformachinelearning?1234predictingwhetherthenextcryofthebabygirlhappensataneven-numberedminuteornotdeterminingwhetheragivengraphcontainsacycledecidingwhethertoapprovecreditcardtosomecustomerguessingwhethertheearthwillbedestroyedbythemisuseofnuclearpowerinthenexttenyearsHsuan-TienLin(NTUCSIE)MachineLearningFoundations11/27TheLearningProblemWhatisMachineLearningFunTimeWhichofthefollowingisbestsuitedformachinelearning?1234predictingwhetherthenextcryofthebabygirlhappensataneven-numberedminuteornotdeterminingwhetheragivengraphcontainsacycledecidingwhethertoapprovecreditcardtosomecustomerguessingwhethertheearthwillbedestroyedbythemisuseofnuclearpowerinthenexttenyearsReferenceAnswer:31234nopatternprogrammabledefinitionpattern:customerbehavior;definition:noteasilyprogrammable;data:historyofbankoperationarguablyno(ornotenough)datayetHsuan-TienLin(NTUCSIE)MachineLearningFoundations11/27TheLearningProblemApplicationsofMachineLearningDailyNeeds:Food,Clothing,Housing,TransportationMLskill1dataFood(Sadileketal.,2013)•data:Twitterdata(words+location)•skill:tellfoodpoisoninglikelinessofrestaurantproperly234Clothing(Abu-Mostafa,2012)•data:salesfigures+clientsurveys•skill:givegoodfashionrecommendationstoclientsHousing(TsanasandXifara,2012)•data:characteristicsofbuildingsandtheirenergyload•skill:predictenergyloadofotherbuildingscloselyTransportation(Stallkampetal.,2012)•data:sometrafficsignimagesandmeanings•skill:recognizetrafficsignsaccuratelyMLiseverywhere!Hsuan-TienLin(NTUCSIE)MachineLearningFoundations12/27TheLearningProblemApplicationsofMachineLearningEducationdataMLskill•data:students’recordsonquizzesonaMathtutoringsystem•skill:predictwhetherastudentcangiveacorrectanswertoanotherquizquestionAPossibleMLSolutionanswercorrectly≈recentstrengthofstudentdifficultyofquestion•giveML9millionrecordsfrom3000students•MLdetermines(reverse-engineers)strengthanddifficultyautomaticallykeypartoftheworld-championsystemfromNationalTaiwanUniv.inKDDCup2010Hsuan-TienLin(NTUCSIE)MachineLearningFoundations13/27TheLearningProblemApplicationsofMachineLearningEntertainment:RecommenderSystem(1/2)dataMLskill•data:howmanyusershaveratedsomemovies•skill:predicthowauserwouldrateanunratedmovieAHotProblem•competitionheldbyNetflixin2006•100,480,507ratingsthat480,189usersgaveto17,770movies•10%improvement=1milliondollarprize•similarcompetition(movies→songs)heldbyYahoo!inKDDCup2011•252,800,275ratingsthat1,000,990usersgaveto624,961songsHowcanmachineslearnourpreferences?Hsuan-TienLin(NTUCSIE)MachineLearningFoundations14/27mTocor?stettbuenckntntncobloeontoidcacymeuCriseinti?edionlockomactcrsliksTTheLearningProblemApplicationsofMachineLearningEntertainment:RecommenderSystem(2/2)MatchmovieandviewerfactorspredictedratingeomCruise?bst