Académique Documents
Professionnel Documents
Culture Documents
InternationalJointConferenceonArtificialIntelligence
Beijing,August4,2013
DietmarJannach
TUDortmund
GerhardFriedrich
AlpenAdriaUniversittKlagenfurt
-1-
-2-
RecommenderSystems
Applicationareas
-3-
IntheSocialWeb
-4-
Evenmore
Personalizedsearch
"Computationaladvertising"
-5-
Aboutthespeakers
GerhardFriedrich
ProfessoratUniversityKlagenfurt,Austria
DietmarJannach
ProfessoratTUDortmund,Germany
Researchbackgroundandinterests
ApplicationofIntelligentSystemstechnologyinbusiness
Recommendersystemsimplementation&evaluation
Productconfigurationsystems
Webmining
Operationsresearch
-6-
Agenda
Whatarerecommendersystemsfor?
Introduction
Howdotheywork(PartI)?
CollaborativeFiltering
Howtomeasuretheirsuccess?
Evaluationtechniques
Howdotheywork(PartII)?
ContentbasedFiltering
KnowledgeBasedRecommendations
HybridizationStrategies
Advancedtopics
Explanations
Humandecisionmaking
Dietmar Jannach, Markus Zanker and Gerhard Friedrich
-7-
-8-
WhyusingRecommenderSystems?
Valueforthecustomer
Findthingsthatareinteresting
Narrowdownthesetofchoices
Helpmeexplorethespaceofoptions
Discovernewthings
Entertainment
Valuefortheprovider
Additionalandprobablyuniquepersonalizedserviceforthecustomer
Increasetrustandcustomerloyalty
Increasesales,clicktroughrates,conversionetc.
Opportunitiesforpromotion,persuasion
Obtainmoreknowledgeaboutcustomers
-9-
Realworldcheck
Mythsfromindustry
Amazon.comgeneratesXpercentoftheirsalesthroughtherecommendation
lists(30<X<70)
Netflix(DVDrentalandmoviestreaming)generatesXpercentoftheirsales
throughtherecommendationlists(30<X<70)
Theremustbesomevalueinit
Seerecommendationofgroups,jobsorpeopleonLinkedIn
FriendrecommendationandadpersonalizationonFacebook
Songrecommendationatlast.fm
NewsrecommendationatForbes.com(plus37%CTR)
Academia
Afewstudiesexistthatshowtheeffect
increasedsales,changesinsalesbehavior
Dietmar Jannach, Markus Zanker and Gerhard Friedrich
- 10 -
Problemdomain
Recommendationsystems(RS)helptomatchuserswithitems
Easeinformationoverload
Salesassistance(guidance,advisory,persuasion,)
RSaresoftwareagentsthatelicittheinterestsandpreferencesofindividual
consumers[]andmakerecommendationsaccordingly.
Theyhavethepotentialtosupportandimprovethequalityofthe
decisionsconsumersmakewhilesearchingforandselectingproductsonline.
[Xiao&Benbasat,MISQ, 2007]
Differentsystemdesigns/paradigms
Basedonavailabilityofexploitabledata
Implicitandexplicituserfeedback
Domaincharacteristics
Dietmar Jannach, Markus Zanker and Gerhard Friedrich
- 11 -
Recommendersystems
RSseenasafunction[AT05]
Given:
Usermodel(e.g.ratings,preferences,demographics,situationalcontext)
Items(withorwithoutdescriptionofitemcharacteristics)
Find:
Relevancescore.Usedforranking.
Finally:
Recommenditemsthatareassumedtoberelevant
But:
Rememberthatrelevancemightbecontextdependent
Characteristicsofthelistitselfmightbeimportant(diversity)
- 12 -
Paradigmsofrecommendersystems
Recommendersystemsreduce
informationoverloadbyestimating
relevance
- 13 -
Paradigmsofrecommendersystems
Personalizedrecommendations
- 14 -
Paradigmsofrecommendersystems
Collaborative:"Tellmewhat'spopular
amongmypeers"
- 15 -
Paradigmsofrecommendersystems
Contentbased:"Showmemoreofthe
samewhatI'veliked"
- 16 -
Paradigmsofrecommendersystems
Knowledgebased:"Tellmewhatfits
basedonmyneeds"
- 17 -
Paradigmsofrecommendersystems
Hybrid:combinationsofvariousinputs
and/orcompositionofdifferent
mechanism
- 18 -
Recommendersystems:basictechniques
Pros
Cons
Collaborative
Noknowledge
engineeringeffort,
serendipityofresults,
learnsmarketsegments
Requiressomeformofrating
feedback,coldstartfornewusers
andnewitems
Contentbased
Nocommunityrequired,
comparisonbetween
itemspossible
Contentdescriptionsnecessary,
coldstartfornewusers,no
surprises
Knowledgebased
Deterministic
recommendations,
assuredquality,nocold
start,canresemblesales
dialogue
Knowledgeengineeringeffortto
bootstrap,basicallystatic,does
notreacttoshorttermtrends
- 19 -
- 20 -
CollaborativeFiltering(CF)
Themostprominentapproachtogeneraterecommendations
usedbylarge,commercialecommercesites
wellunderstood,variousalgorithmsandvariationsexist
applicableinmanydomains(book,movies,DVDs,..)
Approach
usethe"wisdomofthecrowd"torecommenditems
Basicassumptionandidea
Usersgiveratingstocatalogitems(implicitlyorexplicitly)
Customerswhohadsimilartastesinthepast,willhavesimilartastesinthe
future
- 21 -
1992: Usingcollaborativefilteringtoweaveaninformation
tapestry,D.Goldbergetal.,CommunicationsoftheACM
Basicidea:"Eagerreadersreadalldocsimmediately,casualreaderswait
fortheeagerreaderstoannotate"
ExperimentalmailsystematXeroxParcthatrecordsreactionsofusers
whenreadingamail
Usersareprovidedwithpersonalizedmailinglistfiltersinsteadofbeing
forcedtosubscribe
Contentbasedfilters(topics,from/to/subject)
Collaborativefilters
E.g.Mailsto[all]whichwererepliedby[JohnDoe]andwhichreceived
positiveratingsfrom[X]and[Y].
- 22 -
1994: GroupLens:anopenarchitectureforcollaborativefilteringof
netnews,P.Resnicketal.,ACMCSCW
Tapestrysystemdoesnotaggregateratingsandrequiresknowingeach
other
Basicidea:"Peoplewhoagreedintheirsubjectiveevaluationsinthe
pastarelikelytoagreeagaininthefuture"
Buildsonnewsgroupbrowserswithratingfunctionality
- 23 -
Userbasednearestneighborcollaborativefiltering(1)
Thebasictechnique:
Givenan"activeuser"(Alice)andanitemInotyetseenbyAlice
ThegoalistoestimateAlice'sratingforthisitem,e.g.,by
findasetofusers(peers)wholikedthesameitemsasAliceinthepastand
whohaverateditemI
use,e.g.theaverageoftheirratingstopredict,ifAlicewilllikeitemI
dothisforallitemsAlicehasnotseenandrecommendthebestrated
Item1
Item2
Item3
Item4
Item5
Alice
User1
User2
User3
User4
- 24 -
Userbasednearestneighborcollaborativefiltering(2)
Somefirstquestions
Howdowemeasuresimilarity?
Howmanyneighborsshouldweconsider?
Howdowegenerateapredictionfromtheneighbors'ratings?
Item1
Item2
Item3
Item4
Item5
Alice
User1
User2
User3
User4
- 25 -
Measuringusersimilarity
ApopularsimilaritymeasureinuserbasedCF:Pearsoncorrelation
a,b:users
ra,p :ratingofuseraforitemp
P :setofitems,ratedbothbyaandb
Possiblesimilarityvaluesbetween1and1;
=user'saverageratings
Item1
Item2
Item3
Item4
Item5
Alice
User1
User2
User3
User4
sim =0,85
sim =0,70
sim =0,79
- 26 -
Pearsoncorrelation
Takesdifferencesinratingbehaviorintoaccount
6
Alice
User1
User4
4
Ratings
3
2
1
0
Item1
Item2
Item3
Item4
Workswellinusualdomains,comparedwithalternativemeasures
suchascosinesimilarity
Dietmar Jannach, Markus Zanker and Gerhard Friedrich
- 27 -
Makingpredictions
Acommonpredictionfunction:
Calculate,whethertheneighbors'ratingsfortheunseenitemi arehigher
orlowerthantheiraverage
Combinetheratingdifferences usethesimilarityasaweight
Add/subtracttheneighbors'biasfromtheactiveuser'saverageanduse
thisasaprediction
- 28 -
Makingrecommendations
Makingpredictionsistypicallynottheultimategoal
Usualapproach(inacademia)
Rankitemsbasedontheirpredictedratings
However
Thismightleadtotheinclusionof(only)nicheitems
Inpracticealso: Takeitempopularityintoaccount
Approaches
"Learningtorank"
Optimizeaccordingtoagivenrankevaluationmetric(seelater)
- 29 -
Improvingthemetrics/predictionfunction
Notallneighborratingsmightbeequally"valuable"
Agreementoncommonlylikeditemsisnotsoinformativeasagreementon
controversialitems
Possiblesolution:Givemoreweighttoitemsthathaveahighervariance
Valueofnumberofcorateditems
Use"significanceweighting",bye.g.,linearlyreducingtheweightwhenthe
numberofcorateditemsislow
Caseamplification
Intuition:Givemoreweightto"verysimilar"neighbors,i.e.,wherethe
similarityvalueiscloseto1.
Neighborhoodselection
Usesimilaritythresholdorfixednumberofneighbors
- 30 -
Memorybasedandmodelbasedapproaches
UserbasedCFissaidtobe"memorybased"
theratingmatrixisdirectlyusedtofindneighbors/makepredictions
doesnotscaleformostrealworldscenarios
largeecommercesiteshavetensofmillionsofcustomersandmillionsof
items
Modelbasedapproaches
basedonanofflinepreprocessingor"modellearning"phase
atruntime,onlythelearnedmodelisusedtomakepredictions
modelsareupdated/retrainedperiodically
largevarietyoftechniquesused
modelbuildingandupdatingcanbecomputationallyexpensive
- 31 -
2001: Itembasedcollaborativefilteringrecommendationalgorithms,B.
Sarwaretal.,WWW2001
ScalabilityissuesarisewithU2Uifmanymoreusersthanitems
(m>>n,m=|users|,n=|items|)
e.g.Amazon.com
SpacecomplexityO(m2)whenprecomputed
TimecomplexityforcomputingPearsonO(m2n)
Highsparsityleadstofewcommonratingsbetweentwousers
Basicidea:"ItembasedCFexploitsrelationshipsbetweenitemsfirst,
insteadofrelationshipsbetweenusers"
- 32 -
Itembasedcollaborativefiltering
Basicidea:
Usethesimilaritybetweenitems(andnotusers)tomakepredictions
Example:
LookforitemsthataresimilartoItem5
TakeAlice'sratingsfortheseitemstopredicttheratingforItem5
Item1
Item2
Item3
Item4
Item5
Alice
User1
User2
User3
User4
- 33 -
Thecosinesimilaritymeasure
Producesbetterresultsinitemtoitemfiltering
forsomedatasets,noconsistentpictureinliterature
Ratingsareseenasvectorinndimensionalspace
Similarityiscalculatedbasedontheanglebetweenthevectors
Adjustedcosinesimilarity
takeaverageuserratingsintoaccount,transformtheoriginalratings
U:setofuserswhohaveratedbothitemsaandb
- 34 -
Preprocessingforitembasedfiltering
Itembasedfilteringdoesnotsolvethescalabilityproblemitself
PreprocessingapproachbyAmazon.com(in2003)
Calculateallpairwiseitemsimilaritiesinadvance
Theneighborhoodtobeusedatruntimeistypicallyrathersmall,because
onlyitemsaretakenintoaccountwhichtheuserhasrated
Itemsimilaritiesaresupposedtobemorestablethanusersimilarities
Memoryrequirements
UptoN2 pairwisesimilaritiestobememorized(N=numberofitems)in
theory
Inpractice,thisissignificantlylower(itemswithnocoratings)
Furtherreductionspossible
Minimumthresholdforcoratings(items,whichareratedatleastbyn users)
Limitthesizeoftheneighborhood(mightaffectrecommendationaccuracy)
Dietmar Jannach, Markus Zanker and Gerhard Friedrich
- 35 -
Moreonratings
PureCFbasedsystemsonlyrelyontheratingmatrix
Explicitratings
Mostcommonlyused(1to5,1to7Likert responsescales)
Researchtopics
"Optimal"granularityofscale;indicationthat10pointscaleisbetteracceptedin
moviedomain
Multidimensionalratings(multipleratingspermovie)
Challenge
Usersnotalwayswillingtoratemanyitems;sparseratingmatrices
Howtostimulateuserstoratemoreitems?
Implicitratings
clicks,pageviews,timespentonsomepage,demodownloads
Canbeusedinadditiontoexplicitones;questionofcorrectnessofinterpretation
Dietmar Jannach, Markus Zanker and Gerhard Friedrich
- 36 -
Datasparsityproblems
Coldstartproblem
Howtorecommendnewitems?Whattorecommendtonewusers?
Straightforwardapproaches
Ask/forceuserstorateasetofitems
Useanothermethod(e.g.,contentbased,demographicorsimplynon
personalized)intheinitialphase
Alternatives
Usebetteralgorithms(beyondnearestneighborapproaches)
Example:
Innearestneighborapproaches,thesetofsufficientlysimilarneighborsmight
betosmalltomakegoodpredictions
Assume"transitivity"ofneighborhoods
- 37 -
Examplealgorithmsforsparsedatasets
RecursiveCF
Assumethereisaverycloseneighborn ofuwhohoweverhasnotratedthe
targetitemi yet.
Idea:
ApplyCFmethodrecursivelyandpredictaratingforitemi fortheneighbor
Usethispredictedratinginsteadoftheratingofamoredistantdirect
neighbor
Item1
Item2
Item3
Item4
Item5
Alice
User1
User2
User3
User4
sim =0,85
Predict
ratingfor
User1
- 38 -
Graphbasedmethods
"Spreadingactivation"(sketch)
Idea:Usepathsoflengths>3
torecommenditems
Length3:RecommendItem3toUser1
Length5:Item1alsorecommendable
- 39 -
Moremodelbasedapproaches
Plethoraofdifferenttechniquesproposedinthelastyears,e.g.,
Matrixfactorizationtechniques,statistics
singularvaluedecomposition,principalcomponentanalysis
Associationrulemining
compare:shoppingbasketanalysis
Probabilisticmodels
clusteringmodels,Bayesiannetworks,probabilisticLatentSemanticAnalysis
Variousothermachinelearningapproaches
Costsofpreprocessing
Usuallynotdiscussed
Incrementalupdatespossible?
- 40 -
2000: ApplicationofDimensionalityReductionin
RecommenderSystem,B.Sarwaretal.,WebKDDWorkshop
Basicidea:Trademorecomplexofflinemodelbuildingforfasteronline
predictiongeneration
SingularValueDecompositionfordimensionalityreductionofrating
matrices
Capturesimportantfactors/aspectsandtheirweightsinthedata
factorscanbegenre,actorsbutalsononunderstandableones
Assumptionthatkdimensionscapturethesignalsandfilteroutnoise(K=20to100)
Constanttimetomakerecommendations
ApproachalsopopularinIR(LatentSemanticIndexing),data
compression,
- 41 -
Apicturesays
1
Sue
0,8
0,6
0,4
0,2
Bob
Mary
0
-1
-0,8
-0,6
-0,4
-0,2
0,2
0,4
0,6
0,8
-0,2
-0,4
Alice
-0,6
-0,8
-1
- 42 -
Matrixfactorization
Mk Uk k Vk
SVD:
Uk
Dim1
Dim2
Vk T
Alice
0.47
0.30
Dim1
0.44
0.57
0.06
0.38
0.57
Bob
0.44
0.23
Dim2
0.58
0.66
0.26
0.18
0.36
Mary
0.70
0.06
Sue
0.31
0.93
Prediction: rui
ru U k ( Alice) k V ( EPL)
=3+0.84=3.84
Dietmar Jannach, Markus Zanker and Gerhard Friedrich
T
k
Dim1 Dim2
Dim1
5.63
Dim2
3.23
- 43 -
Associationrulemining
Commonlyusedforshoppingbehavioranalysis
aimsatdetectionofrulessuchas
"Ifacustomerpurchasesbabyfoodthenhealsobuysdiapers
in70%ofthecases"
Associationruleminingalgorithms
candetectrulesoftheformX=>Y(e.g.,babyfood=>diapers)fromasetof
salestransactionsD={t1,t2,tn}
measureofquality:support,confidence
- 44 -
Probabilisticmethods
Basicidea(simplisticversionforillustration):
giventheuser/itemratingmatrix
determinetheprobabilitythatuserAlicewilllikeanitemi
basetherecommendationonsuchtheseprobabilities
CalculationofratingprobabilitiesbasedonBayes Theorem
Howprobableisratingvalue"1"forItem5givenAlice'spreviousratings?
CorrespondstoconditionalprobabilityP(Item5=1|X),where
X=Alice'spreviousratings=(Item1=1,Item2=3,Item3=)
CanbeestimatedbasedonBayes'Theorem
Usuallymoresophisticatedmethodsused
Clustering
pLSA
- 45 -
2008: Factorizationmeetstheneighborhood:amultifacetedcollaborative
filteringmodel,Y.Koren,ACMSIGKDD
StimulatedbyworkonNetflixcompetition
Prizeof$1,000,000foraccuracyimprovementof10%RMSE
comparedtoownCinematch system
Verylargedataset(~100Mratings,~480Kusers,~18K
movies)
Lastratings/userwithheld(setK)
Rootmeansquarederrormetricoptimizedto
0.8567
RMSE
(rui rui ) 2
( u ,i )K
- 46 -
2008: Factorizationmeetstheneighborhood:amultifacetedcollaborative
filteringmodel,Y.Koren,ACMSIGKDD
Mergesneighborhoodmodelswithlatentfactormodels
Latentfactormodels
goodtocaptureweaksignalsintheoveralldata
Neighborhoodmodels
goodatdetectingstrongrelationshipsbetweencloseitems
Combinationinonepredictionsinglefunction
Localsearchmethodsuchasstochasticgradientdescenttodetermine
parameters
Addpenaltyforhighvaluestoavoidoverfitting
rui bu bi puT qi
min
p* , q* ,b*
(rui bu bi puT qi ) 2 ( pu
( u ,i )K
qi bu2 bi2 )
2
- 47 -
Summarizingrecentmethods
Recommendationisconcernedwithlearningfromnoisyobservations
(x,y),where f (x) y
hastobedeterminedsuchthat
isminimal.
(
y
y
)
Avarietyofdifferentlearningstrategieshavebeenappliedtryingto
estimatef(x)
Nonparametricneighborhoodmodels
MFmodels,SVMs,NeuralNetworks,BayesianNetworks,
- 48 -
CollaborativeFilteringIssues
Pros:
wellunderstood,workswellinsomedomains,noknowledgeengineeringrequired
Cons:
requiresusercommunity,sparsityproblems,nointegrationofotherknowledgesources,
noexplanationofresults
WhatisthebestCFmethod?
Inwhichsituationandwhichdomain?Inconsistentfindings;alwaysthesamedomains
anddatasets;differencesbetweenmethodsareoftenverysmall(1/100)
Howtoevaluatethepredictionquality?
MAE/RMSE:WhatdoesanMAEof0.7actuallymean?
Serendipity:Notyetfullyunderstood
Whataboutmultidimensionalratings?
- 49 -
- 50 -
RecommenderSystemsineCommerce
OneRecommenderSystemsresearchquestion
Whatshouldbeinthatlist?
- 51 -
RecommenderSystemsineCommerce
Anotherquestionbothinresearchandpractice
Howdoweknowthatthesearegood
recommendations?
- 52 -
RecommenderSystemsineCommerce
Thismightleadto
Whatisagoodrecommendation?
Whatisagoodrecommendationstrategy?
Whatisagoodrecommendationstrategyformy
business?
We
hope
youbeen
will buy
also for
quite a while now
These
have
in stock
- 53 -
Whatisagoodrecommendation?
Whatarethemeasuresinpractice?
Totalsalesnumbers
Promotionofcertainitems
Clickthroughrates
Interactivityonplatform
Customerreturnrates
Customersatisfactionandloyalty
- 54 -
Purposeandsuccesscriteria(1)
Differentperspectives/aspects
Retrievalperspective
Dependsondomainandpurpose
Noholisticevaluationscenarioexists
Reducesearchcosts
Provide"correct"proposals
Assumption:Usersknowinadvancewhattheywant
Recommendationperspective
Serendipity identifyitemsfromtheLongTail
Usersdidnotknowaboutexistence
- 55 -
WhendoesaRSdoitsjobwell?
Recommenditems
fromthelongtail
"Recommendwidely
unknownitemsthat
usersmightactually
like!"
20%ofitems
accumulate74%ofall
positiveratings
- 56 -
Purposeandsuccesscriteria(2)
Predictionperspective
Interactionperspective
Predicttowhatdegreeuserslikeanitem
Mostpopularevaluationscenarioinresearch
Giveusersa"goodfeeling"
Educateusersabouttheproductdomain
Convince/persuadeusers explain
Finally,conversionperspective
Commercialsituations
Increase"hit","clickthrough","lookerstobookers"rates
Optimizesalesmarginsandprofit
- 57 -
Howdoweasresearchers
know?
Testwithrealusers
A/Btests
Examplemeasures:salesincrease,clickthroughrates
Laboratorystudies
Controlledexperiments
Examplemeasures: satisfactionwiththesystem(questionnaires)
Offlineexperiments
Basedonhistoricaldata
Examplemeasures:predictionaccuracy,coverage
- 58 -
Empiricalresearch
Characterizingdimensions:
Whoisthesubject thatisinthefocusofresearch?
Whatresearchmethodsareapplied?
Inwhichsetting doestheresearchtakeplace?
Subject
Onlinecustomers,students,historical online
sessions,computers,
Researchmethod
Experiments,quasiexperiments, nonexperimental
research
Setting
Lab,realworld scenarios
- 59 -
Researchmethods
Experimentalvs.nonexperimental(observational)researchmethods
Experiment(test,trial):
"Anexperimentisastudyinwhichatleastonevariableismanipulatedand
unitsarerandomlyassignedtodifferentlevelsorcategoriesofmanipulated
variable(s)."
Units:users,historicsessions,
Manipulatedvariable:typeofRS,groupsofrecommendeditems,
explanationstrategies
Categoriesofmanipulatedvariable(s):contentbasedRS,collaborativeRS
- 60 -
Experimentdesigns
- 61 -
Evaluationininformationretrieval(IR)
Recommendationisviewedasinformationretrievaltask:
Retrieve(recommend)allitemswhicharepredictedtobe"good"or
"relevant".
Commonprotocol:
Hidesomeitemswithknowngroundtruth
Rankitemsorpredictratings>Count>Crossvalidate
Groundtruthestablishedbyhumandomainexperts
Prediction
Reality
ActuallyGood
ActuallyBad
Rated
Good
TruePositive(tp)
FalsePositive(fp)
Rated
Bad
FalseNegative (fn)
True Negative(tn)
- 62 -
Metrics:PrecisionandRecall
Precision:ameasureofexactness,determinesthefractionofrelevant
itemsretrievedoutofallitemsretrieved
E.g.theproportionofrecommendedmoviesthatareactuallygood
Recall:ameasureofcompleteness,determinesthefractionofrelevant
itemsretrievedoutofallrelevantitems
E.g.theproportionofallgoodmoviesrecommended
- 63 -
DilemmaofIRmeasuresinRS
IRmeasuresarefrequentlyapplied,however:
Groundtruthformostitemsactuallyunknown
What isarelevantitem?
Differentwaysofmeasuringprecisionpossible
Resultsfromofflineexperimentationmayhavelimitedpredictivepowerfor
onlineuserbehavior.
- 64 -
Metrics:RankScore positionmatters
Forauser:
Recommended
(predicted as good)
Actually good
Item 237
Item 899
hit
Item 345
Item 237
Item 187
RankScoreextendsrecallandprecisiontotakethepositionsofcorrect
itemsinarankedlistintoaccount
Particularlyimportantinrecommendersystemsaslowerrankeditemsmaybe
overlookedbyusers
Learningtorank:Optimizemodelsforsuchmeasures(e.g.,AUC)
- 65 -
Accuracymeasures
Datasetswithitemsratedbyusers
MovieLensdatasets100K10Mratings
Netflix100Mratings
Historicuserratingsconstitutegroundtruth
Metricsmeasureerrorrate
MeanAbsoluteError(MAE)computesthedeviationbetween
predictedratingsandactualratings
RootMeanSquareError(RMSE)issimilartoMAE,butplaces
moreemphasisonlargerdeviation
- 66 -
Offlineexperimentationexample
Netflixcompetition
Webbasedmovierental
Prizeof$1,000,000foraccuracyimprovement(RMSE)of10%comparedtoown
Cinematch system.
Historicaldataset
~480Kusersrated~18Kmoviesonascaleof1to5(~100Mratings)
Last9ratings/userwithheld
Probeset forteamsforevaluation
Quizset evaluatesteamssubmissionsforleaderboard
Testset usedbyNetflixtodeterminewinner
Today
Ratingpredictiononlyseenasanadditionalinputintotherecommendationprocess
- 67 -
Animperfectworld
Offlineevaluationisthecheapestvariant
Still,givesusvaluableinsights
andletsuscompareourresults(intheory)
Dangersandtrends:
Dominationofaccuracymeasures
Focusonsmallsetofdomains(40%onmoviesinCS)
Alternativeandcomplementarymeasures:
Diversity,Coverage,Novelty,Familiarity,Serendipity,Popularity,
Concentrationeffects(Longtail)
- 68 -
Onlineexperimentationexample
Effectivenessofdifferentalgorithmsfor
recommendingcellphonegames
[Jannach,Hegelich09]
Involved150,000usersonacommercialmobile
internetportal
Comparisonofrecommendermethods
- 69 -
Detailsandresults
Recommendervariantsincluded:
Itembasedcollaborativefiltering
SlopeOne (alsocollaborativefiltering)
Contentbasedrecommendation
Hybridrecommendation
Toprateditems
}nonpersonalized
Topsellers
Findings:
Personalizedmethodsincreasedsalesupto3.6%comparedtonon
personalized
Choiceofrecommendationalgorithmdependsonusersituation
(e.g.avoidcontentbasedRSinpostsalessituation)
- 70 -
Nonexperimentalresearch
Quasiexperiments
Lackrandomassignmentsofunitstodifferenttreatments
Nonexperimental/observationalresearch
Surveys/Questionnaires
Longitudinalresearch
Observationsoverlongperiodoftime
E.g.customerlifetimevalue,returningcustomers
Casestudies
Focusgroup
Interviews
Thinkaloudprotocols
- 71 -
Quasiexperimental
SkiMatcher ResortFinderintroducedbySkiEurope.comtoprovideusers
withrecommendationsbasedontheirpreferences
ConversationalRS
questionandanswerdialog
matchingofuserpreferenceswithknowledgebase
DelgadoandDavidsonevaluatedthe
effectivenessoftherecommenderovera
4monthperiodin2001
Classifiedasaquasiexperiment
asusersdecideforthemselvesifthey
wanttousetherecommenderornot
Dietmar Jannach, Markus Zanker and Gerhard Friedrich
- 72 -
SkiMatcherResults
July
August
September
October
10,714
15,560
18,317
24,416
SkiMatcherUsers
1,027
1,673
1,878
2,558
NonSkiMatcher Users
9,687
13,887
16,439
21,858
272
506
445
641
75
143
161
229
197
363
284
412
Conversion
2.54%
3.25%
2.43%
2.63%
SkiMatcherUsers
7.30%
8.55%
8.57%
8.95%
NonSkiMatcher Users
2.03%
2.61%
1.73%
1.88%
IncreaseinConversion
359%
327%
496%
475%
UniqueVisitors
RequestsforProposals
SkiMatcherUsers
NonSkiMatcher Users
- 73 -
InterpretingtheResults
Thenatureofthisresearchdesignmeansthatquestionsofcausality
cannotbeanswered(lackofrandomassignments),suchas
Areusersoftherecommendersystemsmorelikelyconvert?
Doestherecommendersystemitselfcauseuserstoconvert?
SomehiddenexogenousvariablemightinfluencethechoiceofusingRSaswell
asconversion.
However,significantcorrelationbetweenusingtherecommender
systemandmakingarequestforaproposal
Sizeofeffecthasbeenreplicatedinotherdomains
Tourism[Jannachetal.,JITT2009]
Electronicconsumerproducts
Dietmar Jannach, Markus Zanker and Gerhard Friedrich
- 74 -
Observationalresearch
Increaseddemandinniches/longtailproducts
Expostfromwebshopdata[Zankeretal.,ECWeb,2006]
- 75 -
Whatispopular?
From:Jannachetal.,Proceedings
ECWeb2012
Usercentricevaluation/Userstudies
Increasedinterestinrecentyears
Variousnumbersofworkshops
- 76 -
Whatarethenexttopics?
Twoadditionalmajorparadigmsofrecommendersystems
Contentbased
Knowledgebased
Hybridization:takethebestofdifferentparadigms
Advancedtopics:recommendersystemsareabouthumandecisionmaking
- 77 -
- 78 -
Contentbasedrecommendation
CollaborativefilteringdoesNOT requireanyinformationabouttheitems,
However,itmightbereasonabletoexploitsuchinformation
E.g.recommendfantasynovelstopeoplewholikedfantasynovelsinthepast
Whatdoweneed:
Someinformationabouttheavailableitemssuchasthegenre("content")
Somesortofuserprofile describingwhattheuserlikes(thepreferences)
Thetask:
Learnuserpreferences
Locate/recommenditemsthatare"similar"totheuserpreferences
- 79 -
Paradigmsofrecommendersystems
Contentbased:"Showmemoreofthe
samewhatI'veliked"
- 80 -
Whatisthe"content"?
Thegenreisactuallynotpartofthecontentofabook
MostCBrecommendationmethodsoriginatefromInformationRetrieval
(IR)field:
Theitemdescriptionsareusuallyautomaticallyextracted(importantwords)
Goalistofindandrankinterestingtextdocuments(newsarticles,webpages)
Here:
ClassicalIRbasedmethodsbasedonkeywords
Noexpertrecommendationknowledgeinvolved
Userprofile(preferences)areratherlearnedthanexplicitlyelicited
- 81 -
Contentrepresentationanditemsimilarities
Simpleapproach
Computethesimilarityofanunseenitemwiththeuserprofilebasedonthe
keywordoverlap(e.g.usingtheDicecoefficient)
sim(bi,bj)=
|
|
- 82 -
TermFrequency InverseDocumentFrequency(TFIDF)
Simplekeywordrepresentationhasitsproblems
Inparticularwhenautomaticallyextractedbecause
Noteverywordhassimilarimportance
Longerdocumentshaveahigherchancetohaveanoverlapwiththeuserprofile
Standardmeasure:TFIDF
Encodestextdocumentsasweightedtermvector
TF:Measures,howoftenatermappears(densityinadocument)
Assumingthatimportanttermsappearmoreoften
Normalizationhastobedoneinordertotakedocumentlengthintoaccount
IDF:Aimstoreducetheweightoftermsthatappearinalldocuments
- 83 -
TFIDF
Computetheoverallimportanceofkeywords
Givenakeywordiandadocumentj
Termfrequency(TF)
Letfreq i,j numberofoccurrencesofkeywordi indocumentj
LetmaxOthers i,j denotethehighestnumberofoccurrencesofanother
keywordofj
,
,
InverseDocumentFrequency(IDF)
N:numberofallrecommendabledocuments
n(i):numberofdocumentsinwhichkeywordi appears
- 84 -
ExampleTFIDFrepresentation
- 85 -
Moreonthevectorspacemodel
Vectorsareusuallylongandsparse
Improvements
Removestopwords("a","the",..)
Usestemming
Sizecutoffs(onlyusetopnmostrepresentativewords,e.g.around100)
Useadditionalknowledge,usemoreelaboratemethodsforfeatureselection
Detectionofphrasesasterms(suchasUnitedNations)
Limitations
Semanticmeaningremainsunknown
Example:usageofawordinanegativecontext
"thereisnothing onthemenuthatavegetarianwouldlike.."
Usualsimilaritymetrictocomparevectors:Cosinesimilarity(angle)
Dietmar Jannach, Markus Zanker and Gerhard Friedrich
- 86 -
Recommendingitems
Simplemethod:nearestneighbors
GivenasetofdocumentsDalreadyratedbytheuser(like/dislike)
Findthennearestneighborsofanotyetseenitemi inD
Taketheseratingstopredictarating/votefori
(Variations:neighborhoodsize,lower/uppersimilaritythresholds)
Querybasedretrieval:Rocchio's method
TheSMARTSystem:Usersareallowedtorate(relevant/irrelevant)retrieved
documents(feedback)
Thesystemthenlearnsaprototypeofrelevant/irrelevantdocuments
Queriesarethenautomaticallyextendedwithadditionalterms/weightof
relevantdocuments
- 87 -
Rocchio details
DocumentcollectionsD+ andD
,, usedtofinetune
thefeedback
oftenonlypositivefeedback
isused
- 88 -
Probabilisticmethods
Recommendationasclassicaltextclassificationproblem
Longhistoryofusingprobabilisticmethods
Simpleapproach:
2classes:like/dislike
SimpleBooleandocumentrepresentation
Calculateprobabilitythatdocumentisliked/dislikedbasedonBayestheorem
Remember:
P(Label=1|X)=
k*P(X|Label=1) * P(Label=1)
- 89 -
Improvements
Sidenote:Conditionalindependenceofeventsdoesinfactnothold
New/YorkandHong/Kong"
Still,goodaccuracycanbeachieved
Booleanrepresentationsimplistic
Keywordcountslost
Moreelaborateprobabilisticmethods
E.g.estimateprobabilityoftermv occurringinadocumentofclassC by
relativefrequencyofv inalldocumentsoftheclass
Otherlinearclassificationalgorithms(machinelearning)canbeused
SupportVectorMachines,..
- 90 -
Limitationsofcontentbasedrecommendationmethods
Keywordsalonemaynotbesufficienttojudgequality/relevanceofa
documentorwebpage
Uptodateness,usability,aesthetics,writingstyle
Contentmayalsobelimited/tooshort
Contentmaynotbeautomaticallyextractable(multimedia)
Rampupphaserequired
Sometrainingdataisstillrequired
Web2.0:Useothersourcestolearntheuserpreferences
Overspecialization
Algorithmstendtopropose"moreofthesame"
E.g.toosimilarnewsitems
- 91 -
- 92 -
Whydoweneedknowledgebasedrecommendation?
Productswithlownumberofavailableratings
Timespanplaysanimportantrole
Fiveyearoldratingsforcomputers
Userlifestyleorfamilysituationchanges
Customerswanttodefinetheirrequirementsexplicitly
Thecolorofthecarshouldbeblack"
- 93 -
Knowledgebasedrecommendation
Knowledgebased:"Tellmewhatfits
basedonmyneeds"
- 94 -
KnowledgebasedrecommendationI
Explicitdomainknowledge
Salesknowledgeelicitationfromdomainexperts
Systemmimicsthebehaviorofexperiencedsalesassistant
Bestpracticesalesinteractions
Canguaranteecorrectrecommendations(determinism)withrespectto
expertknowledge
Conversationalinteractionstrategy
Opposedtooneshotinteraction
Elicitationofuserrequirements
Transferofproductknowledge(educatingusers)
- 95 -
KnowledgeBasedRecommendationII
Differentviewsonknowledge
Similarityfunctions
Determinematchingdegreebetweenqueryanditem(casebasedRS)
UtilitybasedRS
E.g.MAUT Multiattributeutilitytheory
Logicbasedknowledgedescriptions(fromdomainexpert)
E.g.Hardandsoftconstraints
- 96 -
ConstraintbasedrecommendationI
AknowledgebasedRSformulatedasconstraintsatisfactionproblem
CSP ( X I X U , D , SRS KB I )
Def.
XI,XU:VariablesdescribingitemsandusermodelwithdomainD
(e.g.lowerfocallength,purpose)
KB:Knowledgebase comprisingconstraintsand domainrestrictions
(e.g.IF purpose=ontravelTHEN lowerfocallength<28mm)
SRS:Specificrequirementsofauser(e.g.purpose=ontravel)
I:Productcatalog(e.g.(id=1 lfl =28mm) (id=2 lfl=35mm) )
Solution:AssignmenttupleassigningvaluestoallvariablesX
I
s.t. SRS KB I issatisfiable.
Dietmar Jannach, Markus Zanker and Gerhard Friedrich
- 97 -
Itemranking
MultiAttributeUtilityTheory(MAUT)
Eachitemisevaluatedaccordingtoapredefinedsetofdimensionsthatprovide
anaggregatedviewonthebasicitemproperties
value
quality
economy
price
250
>250
8
>8
9
>9
...
5
10
4
10
6
10
...
10
5
10
6
9
6
...
mpix
optzoom
...
- 98 -
CustomerspecificitemutilitieswithMAUT
Customerinterests:
customer
quality
economy
Cu1
80%
20%
Cu2
40%
60%
Itemutilities:
quality
economy
utility:cu1
utility:cu2
P1 (5,4,6,6,3,7,10) = 41
(10,10,9,10,10,10,6) = 65
45.8 [8]
55.4 [6]
P2 (5,4,6,6,10,10,8) = 49
(10,10,9,10,7,8,10) = 64
52.0 [7]
58.0 [1]
P3 (5,4,10,6,10,10,8) = 53 (10,10,6,10,7,8,10) = 61
54.6 [5]
57.8 [2]
...
...
...
...
*
Dietmar Jannach, Markus Zanker and Gerhard Friedrich
**
**
- 99 -
ConstraintbasedrecommendationII
BUT:Whatifnosolutionexists?
KB I
SRS KB I
KB I
notsatisfiable
debuggingofknowledgebase
notsatisfiable but
correct
debuggingofuserrequirements
Applicationofmodelbaseddiagnosisfordebugginguser
requirements
( SRS \ ) KB
Diagnoses:
Repairs:
( SRS \ ) repair KB
CS SRS : CS KB I
Conflictsets:
issatisfiable
issatisfiable
notsatisfiable
- 100 -
Example:findminimalrelaxations(minimaldiagnoses)
KnowledgeBase:
Productcatalogue:
LHS
RHS
PowershotXY
C1
TRUE
Brand =Brandpref.
Brand
Canon
C2
Motives=Landscape
Low.foc.Length=< 28
Lowerfocallength
35
C3
TRUE
Price =<Max.cost
Upperfocallength
140
Price
420 EUR
Currentuser:
Usermodel(SRS)
CS1
CS2
Lumix
R1
Motives
Landscape
R2
Brandpreference
Canon
R3
Max.cost
350EUR
Diagnoses:
Brand
Panasonic
Lowerfocallength
28
Upperfocallength
112
Price
319 EUR
- 101 -
Askuser
Computationofminimalrevisionsofrequirements
Doyouwanttorelaxyourbrandpreference?
AcceptPanasonic insteadofCanon brand
Orisphotographinglandscapeswithawideanglelensandmaximumcostless
important?
Lowerfocallength>28mmandPrice>350EUR
Optionallyguidedbysomepredefinedweightsorpastcommunitybehavior
Beawareofpossiblerevisions(e.g.age,familystatus,)
- 102 -
ConstraintbasedrecommendationIII
Morevariantsofrecommendationtask
Customersmaybenotknowwhattheyareseeking
Find"diverse"setsofitems
Notionofsimilarity/dissimilarity
Ideathatusersnavigateaproductspace
Ifrecommendationsaremorediversethanuserscannavigateviacritiqueson
recommended"entrypoints"moreefficiently(lessstepsofinteraction)
Bundlingofrecommendations
Finditembundlesthatmatchtogetheraccordingtosomeknowledge
E.g.travelpackages,skincaretreatmentsorfinancialportfolios
RSfordifferentitemcategories,CSPrestrictsconfiguringofbundles
- 103 -
Conversationalstrategies
Processconsistingofmultiple
conversationalmoves
Resemblesnaturalsalesinteractions
Notalluserrequirementsknownbeforehand
Customersarerarelysatisfiedwiththeinitial
recommendations
Differentstylesofpreferenceelicitation:
Freetextqueryinterface
Askingtechnical/genericproperties
Images/inspiration
ProposingandCritiquing
- 104 -
Example:adaptivestrategyselection
Statemodel,differentactionspossible
Proposeitem,askuser,relax/tightenresultset,
[Riccietal.,JITT,2009]
- 105 -
Limitationsofknowledgebasedrecommendationmethods
Costofknowledgeacquisition
Fromdomainexperts
Fromusers
Remedy:exploitwebresources
Accuracyofpreferencemodels
Veryfinegranularpreferencemodelsrequiremanyinteractioncycleswiththe
userorsufficientdetaileddataabouttheuser
Remedy:usecollaborativefiltering,estimatesthepreferenceofauser
However:preferencemodelsmaybeinstable
E.g.asymmetricdominanceeffectsanddecoyitems
- 106 -
- 107 -
Hybridrecommendersystems
Allthreebasetechniquesarenaturallyincorporatedbyagoodsalesassistance
(atdifferentstagesofthesalesact)buthavetheirshortcomings
Ideaofcrossingtwo(ormore)species/implementations
hybrida [lat.]:denotesanobjectmadebycombiningtwodifferentelements
Avoidsomeoftheshortcomings
Reachdesirablepropertiesnotpresentinindividualapproaches
Differenthybridizationdesigns
Monolithicexploitingdifferentfeatures
Paralleluseofseveralsystems
Pipelinedinvocationofdifferentsystems
- 108 -
Monolithichybridizationdesign
Onlyasinglerecommendationcomponent
Hybridizationis"virtual"inthesensethat
Features/knowledgesourcesofdifferentparadigmsarecombined
- 109 -
Monolithichybridizationdesigns:Featurecombination
"Hybrid"userfeatures:
Socialfeatures:Movieslikedbyuser
Contentfeatures:Comedieslikedbyuser,dramaslikedbyuser
Hybridfeatures:userswholikemanymoviesthatarecomedies,
thecommonknowledgeengineeringeffortthatinvolvesinventinggood
featurestoenablesuccessfullearning[BHC98]
- 110 -
Monolithichybridizationdesigns:Featureaugmentation
Contentboostedcollaborativefiltering[MMN02]
Basedoncontentfeaturesadditionalratingsarecreated
E.g.AlicelikesItems1and3(unaryratings)
Item7issimilarto1and3byadegreeof0,75
ThusAlicelikesItem7by0,75
Itemmatricesbecomelesssparse
Recommendationofresearchpapers[TMA+04]
Citationsinterpretedascollaborativerecommendations
Integratedincontentbasedrecommendationmethod
- 111 -
Parallelizedhybridizationdesign
Outputofseveralexistingimplementationscombined
Leastinvasivedesign
Weightingorvotingschemeapplied
Weightscanbelearneddynamically
- 112 -
Parallelizedhybridizationdesign:Switching
Specialcaseofdynamicweights(allweightsexceptoneare0)
Requiresanoraclethatdecideswhichrecommenderisused
Example:
Switchingisbasedonsomequalitycriteria:
E.g.iftoofewratingsinthesystem,useknowledgebased,elsecollaborative
- 113 -
Pipelinedhybridizationdesigns
Onerecommendersystempreprocessessomeinputforthesubsequent
one
Cascade
Metalevel
Refinementofrecommendationlists(cascade)
Learningofmodel(e.g.collaborativeknowledgebasedmetalevel)
- 114 -
Pipelinedhybridizationdesigns:Cascade
Recommender 2
Recommender 1
Item1
Item2
Item3
Item4
Item5
0.5
0
0.3
0.1
0
0.8
0.9
0.4
0
0
Item1
Item2
Item3
Item4
Item5
1
2
3
2
1
3
Item1
Item2
Item3
Item4
Item5
0,80
0,00
0,40
0,00
0,00
1
2
Recommendationlistiscontinuallyreduced
Firstrecommenderexcludesitems
Removeabsolutenogoitems(e.g.knowledgebased)
Secondrecommenderassignsscore
Orderingandrefinement(e.g.collaborative)
Dietmar Jannach, Markus Zanker and Gerhard Friedrich
- 115 -
Pipelinedhybridizationdesigns:Metalevel
Successorexploitsamodel builtbypredecessor
ismodelbuiltbyRSn1 exploitedbyRSn
Examples:
Fabsystem:contentbased,collaborativerecommendation[BS97]
Onlinenewsdomain
Contendbasedrecommenderbuildsusermodelsbasedonweightedtermvectors
Collaborativefilteringidentifiessimilarpeersbasedonweightedtermvectorsbut
makesrecommendationsbasedonratings
Collaborative,constraintbasedmetalevelRS
Collaborativefilteringidentifiessimilarpeers
Aconstraintbaseislearnedbyexploitingthebehaviorofsimilarpeers
Learnedconstraintsareemployedtocomputerecommendations
Dietmar Jannach, Markus Zanker and Gerhard Friedrich
- 116 -
Whatisthebesthybridizationstrategy?
Onlyfewworksthatcomparestrategiesfromthemetaperspective
Forinstance,[Burke02]
Mostdatasetsdonotallowtocomparedifferentrecommendationparadigms
I.e.ratings,requirements,itemfeatures,domainknowledge,critiquesrarely
availableinasingledataset
Someconclusionsaresupportedbyempiricalfindings
Monolithic:preprocessingefforttradedinformoreknowledgeincluded
Parallel:requirescarefuldesignofscoresfromdifferentpredictors
Pipelined:workswellfortwoantitheticapproaches
Netflixcompetition stackingrecommendersystems
Weighteddesignbasedon>100predictors recommendationfunctions
Adaptiveswitchingofweightsbasedonusermodel,parameters
- 117 -
AdvancedtopicsI
- 118 -
Explanationsinrecommendersystems
Motivation
ThedigitalcameraProfishot isamustbuyforyoubecause....
Whyshouldrecommendersystemsdealwithexplanationsatall?
Theanswerisrelatedtothetwopartiesprovidingandreceiving
recommendations:
Asellingagentmaybeinterestedinpromotingparticularproducts
Abuyingagentisconcernedaboutmakingtherightbuyingdecision
- 119 -
Explanationsinrecommendersystems
Additionalinformationtoexplainthesystemsoutputfollowingsome
objectives
- 120 -
Objectivesofexplanations
Transparency
Efficiency
Validity
Satisfaction
Trustworthiness
Relevance
Persuasiveness
Comprehensibility
Effectiveness
Education
- 121 -
Explanationsingeneral
How? andWhy? explanationsinexpertsystems
Formofabductive reasoning
Given:
(itemiisrecommendedbymethodRS)
s.t.
Find
Principleofsuccinctness
Findsmallestsubsetof
i.e.forall
holds
s.t.
Butadditionalfiltering
Somepartsrelevantfor
deduction,mightbeobvious
forhumans
[Friedrich&Zanker,AIMagazine,2011]
Dietmar Jannach, Markus Zanker and Gerhard Friedrich
- 122 -
TaxonomyforgeneratingexplanationsinRS
Majordesigndimensionsofcurrentexplanationcomponents:
Categoryofreasoningmodelforgeneratingexplanations
Whitebox
Blackbox
RSparadigmforgeneratingexplanations
Determinestheexploitablesemanticrelations
Informationcategories
- 123 -
RSparadigmsandtheirontologies
Classesofobjects
Users
Items
Properties
Nary relationsbetweenthem
Collaborativefiltering
NeighborhoodbasedCF(a)
Matrixfactorization(b)
Introducesadditionalfactorsasproxiesfor
determiningsimilarities
- 124 -
RSparadigmsandtheirontologies
Contentbased
Propertiescharacterizingitems
TF*IDFmodel
Knowledgebased
Propertiesofitems
Propertiesofusermodel
Additionalmediatingdomainconcepts
- 125 -
Similaritybetweenitems
Similaritybetweenusers
Tags
Tagrelevance(foritem)
Tagpreference(ofuser)
- 126 -
Thermencheck.com(hotspringresorts)
- 127 -
Resultsfromtestingtheexplanationfeature
+*
Explanation
Perceived
Utility
+**
+**
+**
**sign.<1%,*sign.<5%
Trust
+**
Positive
Usageexp.
Recommend
others
Intentionto
repeated
usage
Knowledgeableexplanationssignificantlyincreasetheusers
perceivedutility
Perceivedutilitystronglycorrelateswithusageintentionetc.
- 128 -
Explanationsinrecommendersystems:Summary
Therearemanytypesofexplanationsandvariousgoalsthatan
explanationcanachieve
Whichtypeofexplanationcanbegenerateddependsgreatlyonthe
recommenderapproachapplied
Explanationsmaybeusedtoshapethewishesanddesiresofcustomers
butareadoubleedgedsword
Ontheonehand,explanationscanhelpthecustomertomakewisebuying
decisions;
Ontheotherhand,explanationscanbeabusedtopushacustomerina
directionwhichisadvantageoussolelyfortheseller
Asaresultadeepunderstandingofexplanationsandtheireffectson
customersisofgreatinterest.
- 129 -
AdvancedtopicsII
Onlineconsumerdecisionmaking
RSareaboutHuman decisionmaking
- 130 -
RealitycheckregardingF1 andaccuracymeasuresforRS
Realvalueforcompaniesliesinincreasingconversions
...andsatisfactionwithboughtitems,lowchurnrate
SomereasonswhyitmightbeafallacytothinkF1 onhistoricaldataisa
goodestimateforrealconversion:
Recommendationcanbeselffulfillingprophecy
Userspreferencesarenotinvariant,butcanbeconstructed[ALP03]
Position/Rankiswhatcounts(e.g.serialpositioneffects)
Actualchoicesareheavilybiasedbytheitemsposition[FFG+07]
Inclusionofweak(dominated)itemsincreasesusersconfidence
Replacingsomerecommendeditemsbydecoy itemsfosterschoicetowards
theremainingoptions[TF09]
- 131 -
Consequently
Theunderstandingofonlineusers' purchasingbehaviorisofhigh
importanceforcompanies
Thispurchasingbehaviorcanbeexplainedbydifferentmodelsofhuman
decisionmaking
- 132 -
Effortofdecisionversusaccuracy
Modelfocusesoncostbenefitaspects
Peopleattempttomakeaccuratechoices
Peoplewanttominimizeeffort
Somemethodsformakingchoicesarehighlyaccurate
Theyinvolveconsideringalotofinformation
Calculatingexpectedutilityisahighaccurateandhigheffortwayofmakinga
choice
Somemethodsaresimpler
Theyinvolveconsideringlessinformation
Alsocalledheuristics
- 133 -
Examplesofdecisionheuristics
Simplificationisanunderlyingconceptofheuristics
Satisficing
Choosethefirstitemthatissatisfactory
EliminationbyAspects
Startwiththemostimportantattribute
Eliminateallitemthatarenotsatisfactory
Proceedwiththenextmostimportantattribute
Comeupwithevolvedset
Reasonbasedchoice
Peoplewanttobeabletojustifytheirchoices
Maymakedecisionsthatareeasiesttojustify
- 134 -
Somedecisionphenomenaduetoheuristics
Phenomenon/Effect
Description
Decoyeffects
Additionalirrelevant(inferior)itemsinanitemsetsignificantly
influencetheselectionbehavior
Primacy/recencyeffects
Itemsatthebeginningandtheendofalistareanalyzed
significantlymoreoften/deeplythanitemsinthemiddleofa
list
Framingeffects
Thewayinwhichdifferentdecisionalternativesarepresented
influencesthefinaldecisiontaken
Priming
Ifspecificdecisionpropertiesaremademoreavailablein
memory,thisinfluencesaconsumer'sitemevaluations
(backgroundpriming)
Defaults
Presetoptionsbiasthedecisionprocess
- 135 -
Decoy:asymmetricdominanceeffect
Product
pricepermonth
30
20
50
downloadlimit
10GB
6GB
9GB
ProductAdominatesDinbothdimensions(priceanddownloadlimit)
ProductBdominatesalternativeDinonlyonedimension(price)
TheadditionalinclusionofDintothechoicesetcouldtriggeranincrease
oftheselectionprobabilityofA
- 136 -
Exampleimpactofdecoyeffect
- 137 -
Personality
Differentpersonalitypropertiesposespecificrequirementsonthedesign
ofrecommenderuserinterfaces
Somepersonalitytraitsaremoresusceptibletoheuristical simplifications
Providevariousinterfaces
- 138 -
Personalitytraits
Theory
Description
Internalvs.externalLocusof Externallyinfluencedusersneedmoreguidance;internally
control(LOC)
controlleduserswanttoactivelyandselectivelysearchfor
additionalinformation
Needforclosure
Describestheindividualpursuitofmakingadecisionassoonas
possible
Maximizervs.satisficer
- 139 -
Summaryofonlineconsumerdecisionmaking
Recommendersystemsarepersuasivesystems
Estimatedutilityisoftennotagoodmodelofhumandecisionmaking
Severalsimplifyingheuristics
Boundedrationality/accuracyefforttradeoffmakesuserssusceptiblefor
decisionbiases
Decoyeffects,positioneffects,framing,priming,defaults,
Differentpersonalitycharacteristicsrequiredifferentrecommender
interactionmethods(Max/sat.,need4closure,trust,locOfcontrol)
- 140 -
Outlook
AdditionaltopicscoveredbythebookRecommenderSystems An
Introduction
CasestudyontheMobileInternet
AttacksonCFRecommenderSystems
RecommenderSystemsinthenextgenerationWeb(SocialWeb,SemanticWeb)
Moreonconsumerdecisionmaking
Recommendinginubiquitousenvironments
CurrentandemergingtopicsinRS
SocialWebrecommendations
Contextawarerecommendation
Learningtorank
- 141 -
Thankyouforyourattention!
http://recsys.acm.org
Questions?
Questions?
Questions?
DietmarJannach
eServicesResearchGroup
DepartmentofComputerScience
TUDortmund,Germany
M:dietmar.jannach@tudortmund.de
P:+492317557272
http://www.recommenderbook.net
MarkusZanker andGerhardFriedrich
IntelligentSystemsandBusinessInformatics
InstituteofAppliedInformatics
UniversityKlagenfurt,Austria
M:firstname.lastname@uniklu.ac.at
P:+4346327003705
- 142 -
References
[Adomavicius &Tuzhilin,IEEETKDE,2005]Adomavicius G.,Tuzhilin,A.Towardthenextgenerationofrecommendersystems:asurveyof
thestateoftheartandpossibleextensions,IEEETKDE,17(6),2005,pp.734749.
[ALP03]Ariely,D.,Loewenstein,G.,Prelec,D.(2003)CoherentArbitrainess:StableDemandCurvesWithoutStablePreferences.The
QuarterlyJournalofEconomics,February2003,73105.
[BKW+10]Bollen,D.,Knijnenburg,B.,Willemsen,M.,Graus,M.(2010)UnderstandingChoiceOverloadinRecommenderSystems.ACM
RecommenderSystems,6370.
[Brynjolfsson etal.,Mgt.Science,2003]Brynjolfsson,E.,Hu,Y.,Smith,M.:ConsumerSurplusintheDigitalEconomy:EstimatingtheValue
ofIncreasedProductVarietyatOnlineBooksellers,ManagementScience,Vol 49(11),2003,pp.15801596.
[BS97]Balabanovic,M.,Shoham,Y.(1997)Fab:contentbased,collaborativerecommendation, CommunicationsoftheACM,Vol.40(3),
pp.6672.
[FFG+07]Felfernig,A.,Friedrich,G.,Gula,B.etal.(2007)Persuasiverecommendation:serialpositioneffectsinknowledgebased
recommendersystems.2ndinternationalconferenceonPersuasivetechnology,Springer,283294.
[Friedrich&Zanker,AIMag,2011]Friedrich,G.,Zanker,M.:ATaxonomyforGeneratingExplanationsinRecommenderSystems.AI
Magazine,Vol.32(3),2011.
[Jannach etal.,CUP,2010]Jannach D.,Zanker M.,Felfernig,A.,Friedrich,G.:RecommenderSystemsanIntroduction,Cambridge
UniversityPress,2010.
[Jannach etal.,JITT,2009]Jannach,D.,Zanker,M.,Fuchs,M.:Constraintbasedrecommendationintourism:Amultiperspectivecase
study,InformationTechnology&Tourism,Vol 11(2),pp.139156.
[Jannach,Hegelich 09]Jannach,D.,Hegelich K.:AcasestudyontheeffectivenessofrecommendationsintheMobileInternet,ACM
ConferenceonRecommenderSystems,NewYork,2009,pp.205208
[Riccietal.,JITT,2009]Mahmood,T.,Ricci,F.,Venturini,A.:ImprovingRecommendationEffectivenessbyAdaptingtheDialogueStrategy
inOnlineTravelPlanning.InformationTechnology&Tourism,Vol 11(4),2009,pp.285302.
[Teppan&Felfernig,CEC,2009]Teppan,E.,Felfernig,A.:AsymmetricDominance andCompromiseEffectsintheFinancialServices
Domain.IEEEInternationalConferenceonECommerceandEnterpriseComputing,2009,pp.5764
[TF09]Teppan,E.,Felfernig,A.(2009)Impactsofdecoyelementsonresultsetevaluationsinknowledgebasedrecommendation.Int.
J.Adv.Intell.Paradigms1,358373.
- 143 -
References
[Xiao&Benbasat,MISQ,2007]Xiao,B.,Benbasat,I.:ECommerceProductRecommendationAgents:Use,Characteristics,andImpact,MIS
Quarterly, Vol 31(1),pp.137209.
[Zanker etal.,ECWeb,2006]Zanker,M.,Bricman,M.,Gordea,S.,Jannach,D.,Jessenitschnig,M.:Persuasiveonlinesellinginquality&
tastedomains,7thInternationalConferenceonElectronicCommerceandWebTechnologies,2006,pp.5160.
[Zanker,RecSys,2008]Zanker M.,ACollaborativeConstraintBasedMetaLevelRecommender.ACMConferenceonRecommender
Systems, 2008,pp.139146.
[Zanker etal.,UMUAI,2009]Zanker,M.,Jessenitschnig,M.,Casestudiesonexploitingexplicitcustomerrequirementsinrecommender
systems,UserModelingandUserAdaptedInteraction,Springer,2009,pp.133166.
[Zanker etal.,JITT,2009]Zanker M.,Jessenitschnig M.,Fuchs,M.:AutomatedSemanticAnnotationsofTourismResourcesBasedon
GeospatialData,InformationTechnology&Tourism,Vol 11(4),2009,pp.341354.
[Zanker etal.,Constraints,2010]Zanker M.,Jessenitschnig M.,SchmidW.:Preferencereasoningwithsoftconstraintsinconstraintbased
recommendersystems.Constraints,Springer,Vol 15(4),2010,pp.574595.
- 144 -