人工智能機(jī)器學(xué)習(xí)Alexander Felfernig et al. - Group Recommender Systems. An Introduction (2018,Springer) - libgen.lc

AlexanderFelfernig?LudovicoBoratto

MartinStettinger?MarkoTkalcˇicˇ

GroupRecommender

Systems

AnIntroduction

123

AlexanderFelfernig

InstituteforSoftwareTechnology

GrazUniversityofTechnology

Graz,Austria

LudovicoBoratto

EURECAT

CentreTecnológicodeCatalunya

Barcelona,Spain

MartinStettinger

InstituteforSoftwareTechnology

GrazUniversityofTechnology

Graz,Austria

MarkoTkalcˇicˇ

FacultyofComputerScience

FreeUniversityofBozen-Bolzano

Bolzano,Italy

ISSN2191-8112ISSN2191-8120(electronic)

SpringerBriefsinElectricalandComputerEngineering

ISBN978-3-319-75066-8ISBN978-3-319-75067-5(eBook)

/10.1007/978-3-319-75067-5

LibraryofCongressControlNumber:2018930531

?TheAuthor(s)2018

Preface

Recommendersystemshavebecomeafundamentalmeansforprovidingper-

sonalizedguidancetousersintheirsearchesforinterestingorusefulobjects

suchasmovies,songs,restaurants,softwarerequirements,anddigitalcameras.

Althoughmostexistingrecommendersystemssupportsingleusers,therearemany

scenarioswhereitemsareusedbygroups.Theincreasedinterestinrecommendation

technologiesforgroupsmotivatedustowritethisbookonGroupRecommender

Systems.Theoverallpurposeofthisbookistoprovideaneasytounderstand

introductiontothe?eldofgrouprecommendersystems.Itisintendedforpersons

newtothe?eld,andalsoasreferencematerialforresearchersandpractitionersthat

providesanoverviewoftheexistingstateoftheartandissuesforfuturework.

Includedarecontributionsrelatedtoalgorithms,userinterfaces,psychological

issues,andresearchchallenges.Thebookentailsanintroductorypresentation

ofdifferentgrouprecommendationalgorithms.Beyondalgorithms,itdealswith

additionalrelevantaspectssuchasgrouprecommenderuserinterfaces,evaluation

techniques,approachestothehandlingofpreferences,differentwaystoinclude

explanationsintoagrouprecommendationprocess,andpsychologicalfactorsthat

havetobetakenintoaccountwhenbuildinggrouprecommendersystems.Thebook

alsoprovidesanoverviewofgrouprecommendationscenariosthatgobeyondthe

basicrankingofalternatives.Relatedexamplesincludegroup-basedcon?guration,

recommendationofsequencestogroups,resourcebalancingforgroups,andrelease

planningforgroups.

Graz,AustriaAlexanderFelfernig

Barcelona,SpainLudovicoBoratto

Graz,AustriaMartinStettinger

Bolzano,ItalyMarkoTkalcˇicˇ

Jan2018

Contents

PartIGroupRecommendationTechniques

1DecisionTasksandBasicAlgorithms3

1.1Introduction3

1.2CharacteristicsofDecisionTasks4

1.3RecommendationAlgorithmsforIndividualUsers7

1.4RelationshipBetweenAlgorithmsandChoicePatterns19

1.5BookOverview22

References23

2AlgorithmsforGroupRecommendation27

2.1Introduction27

2.2PreferenceAggregationStrategies29

2.3SocialChoiceBasedPreferenceAggregationFunctions31

2.4CollaborativeFilteringforGroups34

2.5Content-BasedFilteringforGroups37

2.6Constraint-BasedRecommendationforGroups40

2.7HandlingInconsistencies44

2.8Critiquing-BasedRecommendationforGroups47

2.9HybridRecommendationforGroups51

2.10MatrixFactorizationforGroups52

2.11ConclusionsandResearchIssues55

References56

3EvaluatingGroupRecommenderSystems59

3.1Introduction59

3.2Classi?cationMetrics60

3.3ErrorMetrics63

3.4RankingMetrics64

3.5CoverageandSerendipity66

3.6ConsensusandFairness67

3.7ConclusionsandResearchIssues69

References70

PartIIGroupRecommenderUserInterfaces

4GroupRecommenderApplications75

4.1Introduction75

4.2MusicRecommendation75

4.3RecommendationofMoviesandTVPrograms78

4.4RecommendationofTravelDestinationsandEvents79

4.5RecommendationofNewsandWebPages81

4.6GroupRecommendersforHealthyLiving82

4.7GroupRecommendersinSoftwareEngineering82

4.8Domain-IndependentGroupRecommenders84

4.9ConclusionsandResearchIssues87

References88

5HandlingPreferences91

5.1Introduction91

5.2CollectingPreferences93

5.3PreferenceHandlingPractices95

5.4ConsistencyManagement97

5.5ConclusionsandResearchIssues98

References99

6ExplanationsforGroups105

6.1Introduction105

6.2CollaborativeFiltering109

6.3Content-BasedFiltering112

6.4Constraint-BasedRecommendation115

6.5Critiquing-BasedRecommendation121

6.6ConclusionsandResearchIssues122

References123

PartIIIGroupDecisionProcesses

7FurtherChoiceScenarios129

7.1Introduction129

7.2Ranking133

7.3Packaging134

7.4Parametrization134

7.5Con?guration135

7.6ReleasePlanning136

7.7Triage137

7.8ResourceBalancing138

7.9Sequencing138

7.10PollsandQuestionnaires140

7.11Voting141

7.12FurtherAspectsofChoiceScenarios141

7.13ConclusionsandResearchIssues143

References143

8BiasesinGroupDecisions145

8.1Introduction145

8.2DecoyEffects146

8.3SerialPositionEffects147

8.4Framing149

8.5Anchoring149

8.6GroupThink150

8.7EmotionalContagion151

8.8Polarization152

8.9ConclusionsandResearchIssues152

References153

9Personality,Emotions,andGroupDynamics157

9.1PersonalityandEmotions157

9.2GroupDynamics159

9.3Example:TakingintoAccountPersonalityandConformity161

9.4ConclusionsandResearchIssues164

References165

10Conclusions169

Index171

PartIGroupRecommendationTechniques

PartIofthisbookfocusesondifferentrecommendationtechniquesforgroups.

InChap.1,recommendersystemsareintroducedasaspeci?ctypeofdecision

supportsystem.Inthiscontext,thebasicapproachesofcollaborative?ltering,

content-based?ltering,constraint-based,critiquing-based,andhybridrecommen-

dationareintroducedwithaworkingexamplefromthedomainoftravel.InChap.2,

weintroducedifferentconceptsofgrouprecommendersystemsandshowhowsuch

systemscanbebuiltonthebasisoftherecommendationapproachesintroducedin

Chap.1.Finally,inChap.3wediscusstechniquesthatcanbeusedtoevaluategroup

recommendersystems.

Chapter1

DecisionTasksandBasicAlgorithms

AlexanderFelfernig,MüslümAtas,andMartinStettinger

AbstractRecommendersystemsaredecisionsupportsystemshelpingusersto

identifyoneormoreitems(solutions)that?ttheirwishesandneeds.Themost

frequentapplicationofrecommendersystemsnowadaysistoproposeitemsto

individualusers.However,therearemanyscenarioswhereagroupofusersshould

receivearecommendation.Forexample,thinkofagroupdecisionregardingthe

nextholidaydestinationoragroupdecisionregardingarestauranttovisitforajoint

dinner.Thegoalofthisbookistoprovideanintroductiontogrouprecommender

systems,i.e.,recommendersystemsthatdeterminerecommendationsforgroups.

Inthischapter,weprovideanintroductiontobasictypesofrecommendation

algorithmsforindividualusersandcharacterizerelateddecisiontasks.Thisintro-

ductionservesasabasisfortheintroductionofgrouprecommendationalgorithms

inChap.2.

1.1Introduction

Arecommendersystemisaspeci?ctypeofadvice-givingordecisionsupport

systemthatguidesusersinapersonalizedwaytointerestingorusefulobjects

inalargespaceofpossibleoptionsorthatproducessuchobjectsasoutput

[14,17,23,29,43,66,76,81,82].Adecisionproblem/taskemergesifaperson

oragroupofpersonshaveanideaaboutadesiredstate[33].Iftherearedifferent

optionstoachievethedesiredstate,thedecisiontasktobesolvedistoidentifyitems

oractionsthathelptoapproachthetargetstateinasuitablefashion.Recommender

systemscanprovidehelpinsuchacontextbytryingto?ndthesuitableitems

oractionsthathelptobestreachtheenvisionedtarget.Arrivingatachoicecan

beseenastheresultofacollaborationbetweentheuserandtherecommender

system.Recommendersystemssupport“good”choiceswithinreasonabletime

spansincludingcorrespondingjusti?cationsprovidedintermsofexplanations

[42,80].

Therearedifferentwaysinwhicharecommendercansupportusersindecision

makingprocesses.First,itcanactasasupporterto?gureoutcandidateitems,i.e.,

alargenumberofalternativesisreducedtoaso-calledconsiderationset—selecting

thefavoriteoptionislefttotheuser.Second,therecommendercanhelptheuser

selectfromtheitemsintheconsiderationset,forexample,byrepresentingthemin

convenientwaysandprovidingexplanationsofwhytheyhavebeenrecommended.

Onlyin“extreme”casesisthedecisionmakingauthoritytakenoverbythe

recommenderitself.Examplesincludemusicrecommendationina?tnessstudio,

therecommendationofinformationunitsonapublicdisplay,andtheautomated

adaptationofparametersettingssuchaslightintensityinasmarthome[50].

Anexampleofasingle-userrecommendationscenarioisthefollowing:when

navigatinganonlinesalesplatformwiththegoalto?ndabookrelatedtothe

topicof,forexample,deeplearning,arecommendersystemwillidentifyrelated

booksandproposethesetotheuseroftheonlineplatform.Inthisscenario,the

envisionedtargetstateisto?ndasuitablebookonthementionedtopicandthe

optionstoachievethisstatearethedifferentexistingbooksonthetopicofdeep

learning.Findingabookinanonlinesalesplatformistypicallyasingleuser

decisionscenario.However,therearealsoscenarioswhereagroupofusershas

tomakeadecision.Inthiscontext,arecommendersystemmusttakeintoaccount

thepotentiallycon?ictingpreferencesofdifferentgroupmembers.Suchasituation

makestherecommendationtaskdifferentandoftenmorechallenging.

Themainfocusofthisbookisrecommendationtechniquesthatprovidehelpin

scenarioswhereagroupofusersisengaged.Inmanyscenarios,thepresentation

ofrecommendationstogroupsisamorenaturalapproachthantryingtoaddress

individualusers[40,56,58].Forexample,musicrecommendationsin?tnessstudios

havetotakeintoaccountthepreferencesofallindividualscurrentlypresentinthe

studio[59].Stakeholdersinasoftwareprojecthavetoestablishagreementregarding

therequirements/featuresthathavetobedevelopedwithinthescopeofthenext

release[64].Personneldecisionsareoftentakeningroups,i.e.,agrouphastodecide

whichjobapplicantwillbehired[78].Groupsoffriendshavetodecideaboutthe

hotelforthenextsummerholidaysoraskiingresortforthenextwinterholidays

[39,61].Apublicdisplayshouldbepersonalizedinordertobeabletodisplay

informationtopersonscurrentlyinthesurrounding[40].Finally,travelgroups

shouldreceiveapersonalizedmuseumguidanceinsuchawaythatthepersonal

preferencesofgroupmembersareful?lled[28,47].

1.2CharacteristicsofDecisionTasks

Decisiontasksdifferwithregardtovariousaspects[69].Inthefollowing,we

introducebasicdimensionsofdecisiontasks(see[33])whichwillhelptobetter

understandwhichdecisionsaresupportedbygrouprecommenders(seeTable1.1).1

1Asmentionedin[33],thischaracterizationofdecisiontasksisnotcompletebutagoodbasisfor

discussingpropertiesrelevantforgrouprecommenders.

Table1.1Characteristicsof

decisiontasks[33]

DimensionCharacteristic

Complexitylow..high

Structurednesslow..high

Decisiontypechoice..design

Sentimentopportunity..threat

Dependenceyes..no

Leveloriginal..meta

Actorperson..group

ComplexityWeinterpretcomplexityofdecisiontasksintermsofthenumberof

decisiondimensionsandthedegreeofiteminvolvement[33,38].Dependingon

theunderlyingitemdomain,humanswillinvestmoreorlesstimetocometo

adecision,i.e.,toachieveanacceptabletrade-offbetweendecisioneffortand

accuracy[67].Itemswithhigherrelateddecisioneffortsareoftendenotedashigh-

involvementitemswhereasitemswithlessrelateddecisioneffortsaredenotedas

low-involvementitems[70].Suboptimaldecisionshaveamuchhighernegative

impactinthecontextofhigh-involvementitems.Forexample,whenpurchasingan

apartment,asuboptimaldecisionmanifestsinsearcheffortsforanewapartment,

unnecessarypayments,relocationcosts,andadditionaltimeefforts[27].Incontrast,

risksrelatedtothepurchasingofalow-qualitybookarenegligible—intheworst

case,theuserwillprovidenegativefeedbackonthebookandtryto?ndother

alternativesthatbetter?this/herwishesandneeds.Whenpurchasingabook,the

numberofdecisiondimensionsislow—examplesofdimensionsarepriceand

quality.Thenumberofdecisiondimensionsofapartmentsismuchhigher(e.g.,

price,qualityofpublictransport,neighborhood,schoolsintheneighborhood,etc.).

Aspectsthatfurtherincreasedecisioncomplexityespeciallyingroupdecision

scenariosare,forexample,contradictingpreferencesofgroupmembers,personal

relationships,personalityfactors,andemotion-relatedaspects(seeChap.9).

StructurednessWeinterpretstructurednessofdecisiontasksasthedegreetowhich

underlyingprocessesanddecisionpoliciesarede?ned.Decisiontasksareoften

characterizedbyunde?nedprocessesandrelateddecisionpoliciesarenotpre-

de?nedbutdevelopedandadaptedinthecourseofthedecisionprocess.Ifagroup

ofusershastodecideonaholidaydestinationforthenextsummer,arecommender

systemcanproposealternativedestinationsbutitisunclearwhichofthealternatives

willbechosenbythegroup.The?naldecisionissomethingthathastobemade

bygroupmembers(oranindividualuser)andisinmanycasesnothandledby

theunderlyingdecisionsupportenvironment.Thereareexceptionstotherule,for

example,musicrecommendationsin?tnesscentersandinformationunitsshownon

publicdisplays.Speci?cdecisiontypesfollowaformalizedprocess.Forexample,

electoralsystemsarede?nedbypreciserulesthatdeterminehowelectionsand

referendumshavetobeconducted(theprocess)andhowtheresultsaredetermined

(decisionmaking).

DecisionTypeDecisiontasksareoftende?nedonthebasisofknownalternatives

orparametersoutofwhichoneormorealternatives(values)shouldbeselected.

Ifalternatives(parameters)areprede?ned,theunderlyingdecisiontaskcanbe

regardedasabasicchoicetask[42].Choiceproblemsareconsideredascentral

applicationareaforrecommendersystems[42].Theotherextremeareso-called

designtasks,wherealternativesarenotprede?nedbutcreatedthroughoutadecision

process.Designtasksareoftenrelatedtocreativeactswherepersonsdevelop

ideasandsolutions.In“pure”designscenarios,theapplicationofrecommendation

technologiesisnotwidespread[75].However,therearemanyscenarioslocated

in-betweenbasicchoicetasksand“pure”designtasks.Forexample,knowledge-

basedcon?guration[25]canbeconsideredasasimplertypeofdesigntaskwhere

asolutionisidenti?ed(con?gured)outofasetofpre-de?nedcomponenttypes.In

thiscontext,thealternatives(parametervalues)areknownbeforehand;duetothe

largeoptionspace,notallpotentialalternativescanbeenumeratedforperformance

reasons—billionsofalternativeswouldhavetobemanagedandthecorresponding

recommendationalgorithmswouldbecomeinef?cient[13](seeChap.7).

SentimentDecisionsupportwithincludedrecommendationsupportisveryoften

opportunity-related,i.e.,thegoalisrelatedtoanopportunityandthebestsolutionto

achieveagoalshouldbeidenti?ed.Examplesthereofarepurchasingabooktobetter

understandacertaintopicorchoosingaholidaydestinationtospendunforgettable

dayssomewhereabroad.Asimilarargumentholdsforitemdomainssuchassongs,

digitalequipment,food,and?nancialservices.Certainly,decisionproblemsalso

existincontextswherealternativeoutcomescanbeconsideredasnegativeones.

Forexample,choosingbetweenalternativeoptionstoliquidateacompany—inthis

scenario,everyoutcomecanbeconsideredasanegativeone(thecompanygets

liquidated).However,recommendersystemscanhelptominimizedamage,for

example,onthebasisofastructuredutilityanalysis[79].

DependenceWeregarddecisiontasksasdependentiftheoutcomeofadecision

hasanimpactonanotherdecision.Forexample,thepurchaseofamovietypically

doesnotrequireafollow-updecisionregardingthepurchaseofthenextmovie

ordifferentitem.Dependentdecisiontasksoccurwhenonedecisionatanearlier

pointoftimeleadstofollow-updecisiontasksatalaterpointoftime.Anexample

ofsuchadecisiontaskisrequirementsreleaseplanningwhereforeachsoftware

requirementithastobedecidedwhentherequirementshouldbeimplemented[64].

Consequently,decisionstakeninearlyphasesofasoftwareprojectcanhavean

impactonortriggerdecisionslaterintheproject.Forexample,adecisionthata

requirementshouldbeimplementedcouldtriggerdecisionsregardingadditional

resourcesinordertobeabletoprovidethepromisedsoftwarefunctionalitiesin

time.

LevelWecandifferentiatebetweenoriginaldecisionsoperatingontheobject

levelanddecisionsonthemeta-level[33].The?rsttypeisomni-presentinmany

recommendation-supportedscenarios—theunderlyingtaskistoidentifyandchoose

itemsofrelevance.Incontrast,meta-decisionsaredecisionsaboutthequalitiesof

adecisionprocessandthewaydecisionsaretaken.Forexample,agroupcould

decidetousemajorityvotingwhenitcomestotheelectionofthenextchairman.

Ameta-decisioninthiscontextistodecideabouttheelectionformalism—related

alternativescanbe,forexample,relativemajorityandasingle-shotelectionor

absolutemajorityinapotentiallymulti-levelelectionprocess[51].Inmanydecision

scenarios—especiallyinthecontextofgroupdecisionmaking—recommendations

haveanadvisoryfunctionbutarenotconsideredimperative.

ActorManyrecommendationapproachessupportindividualdecisionmaking

whererecommendationalgorithmsarefocusingondeterminingrecommendations

forindividualusers.Thefocusofthisbookarerecommendersystemsthatsupport

decisionmakingforgroupsofusers.Thefollowingtypesofgroupsareintroduced

in[7]:(1)establishedgroupswithsharedandlong-termcommoninterests(e.g.,

conferencecommitteestakingdecisionsaboutconferencevenuesorfamiliestaking

adecisionaboutpurchasingahouse),(2)occasionalgroupswithacommonaim

inaparticularmoment(e.g.,agroupofpersonsjointlyparticipatinginamuseum

tour),(3)randomgroups(e.g.,personsina?tnesscenterorpersonsaroundapublic

display),and(4)automaticallyidenti?edgroupswhereindividualswithsimilar

preferenceshavetobegrouped(e.g.,distributionofseminarpaperstostudentsand

distributionofconferencepaperstoreviewers).

1.3RecommendationAlgorithmsforIndividualUsers

Recommendersystems[43,58]proposeitemsofpotentialinteresttoanindividual

useroragroupofusers.2Theyareappliedinitemdomainssuchasbooks[52],

websites[68],?nancialservices[16],andsoftwareartifacts[20,24].Inthe

following,weintroducecollaborative?ltering[31,48],content-based?ltering[68],

constraint-based[8,14],critiquing-based[10,11],andhybridrecommendation

[9]whicharebasicrecommendationapproaches.TheitemsinTable1.2(travel

destinations)serveasexamplestodemonstratehowbasicrecommendationalgo-

rithmsoperate.InChap.2,weshowhowtheseapproachescanbeintegratedinto

correspondinggrouprecommendationscenarios.

CollaborativeFiltering

Collaborative?lteringisbasedontheideaofword-of-mouthpromotionwhereopin-

ionsofrelativesandfriendsplayamajorrolewhentakingadecision[12,48,52].In

onlinescenarios,familymembersandfriendsarereplacedbynearestneighborswho

2Partsofthissectionarebasedonadiscussionofrecommendationtechnologiesgivenin[24].

Table1.2Examplesetoftraveldestinations

Traveldestination(item)NameBeachCitytoursNatureEntertainment

t1Viennaxx

t2Yellowstonex

t3NewYorkxxx

t4BlueMountainsx

t5Londonxx

t6Beijingxx

t7CapeTownxxxx

t8Yosemityx

t9Parisxx

t10Pittsburghxx

Theseitemswillbeusedacrossthefollowingsectionsfordemonstrationpurposes.Eachtravel

destinationisdescribedbyasetofmeta-characteristics(categories),forexample,traveldestination

Yellowstoneisfamousforitsexperienceofnature

areuserswithpreferencessimilartotheonesofthecurrentuser.Incollaborative?ltering,auseritemratingspeci?estowhichextentauserlikesanitem.Rating

predictionsaredeterminedbyarecommenderalgorithmtoestimatetheextentauserwilllikeanitemhe/shedidnotconsume/evaluateuptonow.Acollaborative?lteringrecommender?rstdeterminesknearestneighbors(kNN).

3Thepreferencesofnearestneighborsareusedtoextrapolatefutureitemratingsofthecurrentuser.Auseritemratingmatrixthatwillbeusedinthefollowingforexplaining

collaborative?lteringisshowninTable1.3.Inthisexample,allusersuivisitedtraveldestinationsandprovidedacorrespondingrating.Incollaborative?ltering,useritemratingsserveasinputfortherecommender.

Collaborative?lteringidenti?esthek-nearestneighborsofthecurrentuserua(seeFormula1.1)4and—basedonthenearestneighbors—calculatesaprediction

forthecurrentuser’srating.WhenapplyingFormula1.1,useru2isidenti?edas

thenearestneighborofuserua(seealsoTable1.3).Thesimilaritybetweenua

andanotheruseruxcanbedetermined,forexample,usingthePearsoncorrelation

coef?cient[43](seeFormula1.1)whereTDcisthesetofitemsthathavebeenratedbybothusers(uaandux),rx;tiistheratingofuserxforitemti,andrxistheaverage

ratingofuserx.SimilarityvaluesresultingfromFormula1.1cantakevaluesonascaleof[1::C1].Sometimes,“neighbor”userswithlowornegativecorrelationswiththecurrentuserare?lteredout[1].

3Wefocusonuser-basedcollaborative?lteringwhichisamemory-basedapproachthat—incontrasttomodel-basedones—operatesonanuncompressedversionofauser/itemmatrix[6,12].

4Weassumek=2inourexample.

Table1.3Exampleofcollaborative?lteringratingmatrix:traveldestinations(items)tiandratings

(weassumearatingscaleof1..5)

ItemNameu1u2u3u4ua

t1Vienna5.04.0

t2Yellowstone4.0

t3NewYork3.04.03.0

t4BlueMountains5.05.04.0

t5London3.0

t6Beijing4.54.04.0

t7CapeTown4.0

t8Yosemity2.0

t9Paris3.0

t10Pittsburgh5.03.0

Averagerating(rx)4.333.6254.03.753.67

Table1.4Similarity

betweenuseruaandtheusers

uj¤u

adeterminedbasedon

Formula1.1

u1u2u3u4

ua–0.970.70–

Ifthenumberofcommonly

rateditemsisbelow2,no

similaritybetweenthetwo

usersiscalculated

similarity.ua;ux/DP

c.ra;tir

qPa/2qP

(1.1)

a/.rx;tir

x/

ti2TD

c.ra;tirc.rx;tir

x/2

ti2TDti2TD

ThesimilarityvaluesforuacalculatedbasedonFormula1.1areshowninTable1.4.Forthepurposeofourexample,weassumetheexistenceofatleasttwoitemsperuserpair(ui,uj)(i¤j)inordertobeabletodetermineasimilarity.This

criterionholdsforusersu2andu3.

Achallengewhendeterminingthesimilaritybetweenusersisthesparsityofthe

ratingmatrix.Userstypicallyprovideratingsforonlyaverysmallsubsetofthe

offereditems.Forexample,givenalargemoviedatasetthatcontainsthousandsof

entries,auserwilltypicallybeabletorateonlyafewdozen.Oneapproachtothis

problemistotakeintoaccountthenumberofcommonlyrateditemsasacorrelation

signi?cance[37],i.e.,thehigherthenumberofcommonlyrateditems,thehigheris

thesigni?canceofthecorrespondingcorrelation.Forfurtherinformationregarding

thehandlingofsparsity,wereferto[37,43].

Theinformationaboutthesetofuserswitharatingbehaviorsimilartothatof

thecurrentuser(nearestneighborsNN)isthebasisforpredictingtheratingofuser

uaforanitemtthathassofarnotbeenratedbyua(seeFormula1.2).

prediction.ua;t/DrO.ua;t/Dr

aCP

P

similarity.ua;uj/.rj;tr

j/

uj2NN

similarity.ua;uj/

uj2NN

(1.2)

Table1.5Collaborative?lteringbasedrecommendations(predictions)foritemsthathavenot

beenratedbyuseruauptonow

t1t2t3t4t5t6t7t8t9t10u2––3.05.0–4.5–2.0––

u3––4.05.03.04.0––––

ua–––4.0–4.0––3.0

Predictionforua––3.30p–2.66––2.04––

Basedontheratingsofthenearestneighborsofua,weareabletodeterminea

predictionforua(seeTable1.5).Thenearestneighborsofuaareassumedtobeu2

andu3(seeTable1.4).Thetraveldestinationsratedbythenearestneighborsbut

notratedbyuaaret3,t5,andt8.Duetothedeterminedpredictions(Formula1.2),

itemt3wouldberankedhigherthantheitemst5andt8inarecommendationlist.

Foradiscussionofadvancedcollaborativerecommendationapproaches,werefer

thereaderto[49,74].

Content-BasedFiltering

Thisapproachisbasedontheassumptionofmonotonicpersonalinterests[68].For

example,usersinterestedinpoliticalnewsaretypicallynotchangingtheirinterest

pro?lefromonedaytoanother.Onthecontrary,theywillalsobeinterestedin

thetopicinthe(near)future.Inonlinescenarios,content-basedrecommendersare

applied,forexample,whenitcomestotherecommendationofwebsites[68].

Content-based?lteringisbasedon(a)asetofusersand(b)asetofcategories

(orkeywords)thathavebeenassignedto(orextractedfrom)thesetofitems.It

comparesthecontentofalreadyconsumeditemswithnewitems,i.e.,it?ndsitemsthataresimilartothosealreadyconsumed(positivelyrated)bytheuser(ua).The

basisfordeterminingsuchasimilarityarekeywordsextractedfromtheitemcontent

descriptions(e.g.,keywordsextractedfromnewsarticles)orcategoriesifitems

havebeenannotatedwiththerelevantmeta-information.Readersinterestedinthe

principlesofkeywordextractionarereferredto[43].Inthisbook,wefocuson

content-basedrecommendationwhichexploitsitemcategories(seeTable1.2).

Content-based?lteringwillnowbeexplainedusingTables1.2,1.6,and1.7.

Table1.2providesanoverviewoftherelevantitemsandtheassignmentsofitems

tocategories.Table1.6providesinformation