模拟国际会议演讲稿

Recsplorer:RecommendationMechanismBasedonStudents’WorkingMemoryCapacityinLearningSystems1.IntroductionThankyouverymuch,Dr.Li,foryourkindintroduction.Ladiesandgentlemen,Goodmorning!Iamhonoredtohavebeeninvitedtospeakatthisconference.BeforeIstartmyspeech,letmeaskaquestion.Doyouthinkrecomemdationsfromothersareusefulforyourinternetshopping?Thankyou.Itisobviousthatrecommendationsplayanimportantroleinourdailyconsumptiondecisions.Today,mytopicisaboutRecommendationMechanismBasedonStudents’WorkingMemoryCapacityinLearningSystems.Iwanttoshareourinterestingresearchresultonrecommendationalgorithmswithyou.Thecontentofthispresentationisdividedinto5parts:insession1,Iwillintruducethetradictionalrecommendationandournewstrategy;insession2,IwillgivetheformaldefinitionofPrecedenceMining;insession3,Iwilltalkaboutthenovelrecommendationalgorithms;experimentalresultwillbeshowedinsession4;andfinally,Iwillmakeaconclusion.2.BodySession1:IntroductionThepictureonthisslideisaninstanceofrecommemdationapplicationonamazon.Recommendersystemsprovideadviceonproducts,movies,webpages,andmanyothertopics,andhavebecomepopularinmanysites,suchasAmazon.Manysystemsusecollaborativefilteringmethods.ThemainprocessofCFisorganizedasfollow:first,identifyuserssimilartotargetuser;second,recommenditemsbasedonthesimilarusers.Unfortunately,theorderofconsumeditemsisneglect.Inourpaper,weconsideranewrecommendationstrategybasedonprecedencepatterns.Thesepatternsmayencompassuserpreferences,encodesomelogicalorderofoptionsandcapturehowinterestsevolve.Precedenceminingmodelestimatetheprobabilityofuserfutureconsumptionbasedonpastbehavior.Andtheseprobabilitiesareusedtomakerecommendations.Throughourexperiment,precedenceminingcansignificantlyimproverecommendationperformance.Futhermore,itdoesnotsufferfromthesparsityofratingsproblemandexploitpatternsacrossallusers,notjustsimilarusers.Thisslidedemonstratesthedifferencesbetweencollaborativefilteringandprecedencemining.Supposethatthescenarioisaboutcourseselection.Eachquarter/semesterastudentchoosesacourse,andratesitfrom1to5.Figurea)showsfivetranscripts,atranscriptmeansalistofcourse.Uisourtargetstudentwhoneedrecommendations.Figureb)illustrateshowCFwork.Assumesimilarusersshareatleasttwocommoncoursesandhavesimilarrating,thenu3andu4aresimilartou,andtheircommoncoursehwillbearecommendationtou.Figurec)presentshowprecedenceminingwork.Forthisexample,weconsiderpatternswhereonecoursefollowsanother.Supposepatternsoccouratleasttwotranscripsarerecognizedassignificant,then(a,d),(e,f)and(g,h)arefoundout.Andd,h,andfarerecommendationtouwhohastakena,gande.NowIwillaprobabilisticframeworktosolvetheprecedenceminingproblems.Ourtargetuserhasselectedcoursea,wewanttocomputetheprobabilitycoursexwillfollow,i.e.,Pr[x|a].howerve,whatwereallyneedtocalculateisPr[x|a﹁X]ratherthanPr[x|a].Becauseinourcontext,wearedecidingifxisagoodrecommendationforthetargetuserthathastakena.Thusweknowthatourtargetuser’stranscriptdoesnothavexbeforea.Forinstance,thetranscriptno.5willbeomitted.Inmorecommonsituation,ourtargetuserhastakenalistofcourses,T={a,b,c,…}notjusta.Thus,whatreallyneedisPr[x|T﹁X].Thequestionishowtofigureoutthisprobability.Iwillansweritlater.Session2:PrecedenceMiningWeconsiderasetDofdistinctcourses.Weuselowercaseletters(e.g.,a,b,…)torefertocoursesinD.AtranscriptTisasequenceofcourses,e.g.,a-b-c-d.ThenthedefinitionofTop-kRecommendationProblemisasfollows.GivenasettranscriptsoverDfornusers,theextratranscriptTofatargetuser,andadesirednumberofrecommendationsk,ourgoalisto:1.Assignascorescore(x)(between0and1)toeverycoursex∈Dthatreflectshowlikelyitisthetargetstudentwillbeinterestedintakingx.Ifx∈T,thenscore(x)=0.2.Usingthescorefunction,selectthetopkcoursestorecommendtothetargetuser.Tocomputescores,weproposetousethefollowingstatistics,wherex,y∈D:f(x):thenumberoftranscriptsthatcontainx.g(x;y):thenumberoftranscriptsinwhichxprecedescoursey.Thisslideshowsthecalculationresultoff(x)andg(x,y).Forexample,fromthetable,weknowthatf(a)is10andg(a,c)is3.WeproposeaprecedenceminingmodeltosolvetheTop-kRecommendationProblem.Herearesomenotation:x﹁y,whichwehavememtionedinsession1,referstotranscriptwherexoccurswithoutaprecedingy;x﹁yreferstotranscriptwherexoccurswithoutyfollowingit.Weusequantitiesf(x)andg(x,y)tocompteprobabilitiesthatencodetheprecedenceinformation.Forinstance,fromformular1to7.Iwouldnottellthedetailofallformulars.Wejustpayattentiontoformular5,notethatthisquantityaboveisthesameas:Pr[x﹁y|y﹁x]whichwillbeusedtocomputescore(x).Asweknow,thetargetuserusuallyhastakenalistofcoursesratherthanacourse,soweneedtoextentourprobabilitycalculationformulars.Forexample,supposeT={a,b},Pr[x﹁T]theprobabilityxoccurswithouteitheranaorbprecedingit;Pr[x﹁T]theprobabilityxoccurswithouteitheranaorbfollowingit.Thisprobabilitycanbecalculatedexactly.Sohowtocalculateit?Session3:RecommendationAlgorithmsLet’sreviewsession2.Themaingoaloftherecommendationalgorithmsistocalculatethescore(x),andthenselectthetopkcoursesbasedonthesescores.TraditionalrecommendationalgorithmscomputearecommendationscoreforacoursexinDonlybase