reciprocity

ReciprocalRecommenderSystemforOnlineDating

LuizPizzato,TomekRej,ThomasChung,IrenaKoprinska,KalinaYacefandJudyKay

SchoolofInformationTechnologies

UniversityofSydney,NSW2006,Australia

{e}@

ABSTRACT

Reciprocalrecommenderisaclassofrecommendersystemsthat

isimportantfortaskswherepeopleareboththesubjectandthe

objectoftherecommendation;

haveimplementedRECON,areciprocalrecommenderforonline

dating,andwehaveevaluateditonamajordatingwebsite.

Resultsshowanimprovedsuccessrateforrecommendationsthat

considerreciprocityincomparisontorecommendationsthatonly

considerthepreferencesoftheursreceivingthe

recommendations.

CategoriesandSubjectDescriptors

H.3.3[InformationStorageandRetrieval]:InformationSearch

andRetrieval–Informationfiltering

GeneralTerms

Algorithms,Experimentation

Keywords

Recommendersystems,Onlinedating,Reciprocity

UCTION

hasbeenevaluatedonalargedatatfromamajorAustralian

ocalrecommenders[1]recommendpeople

topeopleandasuccessfulrecommendationonlyoccurswhen

tiontoonlinedating,other

importantapplicationsofreciprocalrecommendersinclude

matchingemployerswithjobapplicants,matchingmenteeswith

mentorsandidentificationofbusinesspartners.

Areciprocalrecommenderdoesnotimplysocialmatching,butit

doesrelatetocomputer-humaninteractionissuesraidby

TerveenandMcDonald[2]suchasprivacy,trust,relationand

interpersonalattraction.

Urscreateonlineprofileswhichtypicallyconsistofa

predefinedlistofattributesaboutthemlvesandtheirideal

xpressinterestinanotherurbyndinga

predefinedmessage,suchas“Ilikeyou,doyouwanttotalk?”.

Thereceivercanreplywithapositiveornegativemessage,e.g.

“Lovedthemessage,canyoundmeanemail.”or“Thanks,but

Idon’tthinkwearerightforeachother”.

RECONusthepredefinedmessagestolearntheur’s

filtersoutcandidateswhodonotsatisfythe

ur’spreferencesandrankstheremainingcandidatesusinga

rankingcriterion.

RECONpromotescandidateswhoarelikelytoreciprocatethe

ur’smessagebyrespondingpositivelytohim/her,

willappearatthetopofurU'srecommendationlistifUappears

inV'onallyRECONsupports

variousrankingcriteriabadonpopularity,responrateandlast

sopossibletoorderthecandidatesbyacompatibility

score,whichisbadonhowwellthecandidatesmatchtheur's

ecandidatesareranked,RECONdisplaysthe

top-NcandidateswhereNisaurdefinedparameter.

TION

WehaveevaluatedRECONusingurinteractiondatafromasix

weeksperiod,wherethefirstfourweekswereudastraining

iningdata

consistedof1.4millionmessagesntbyover90,000urs.

BytakingintoaccountreciprocityRECONimprovedthesuccess

rateofthetop-10recommendationsfrom23%to42%.

Reciprocityalsohelpedwiththecoldstartproblemprovidingan

improvementofmorethan60%insuccessratefornewurs.

Wealsoobrvedlargeimprovementsinrecall,suchasan

improvementof83%forthetop-100recommendations(from

5.90%to10.80%).

MANCE

ThetimerequiredtorunRECONdependsondifferentstagesof

alculatesthepreferencesformorethan

90,000ursandalltheirmessagesinabout10minutesusingtwo

hepreferences,it

createsalistofrecommendationsforallursinabout2hours.

MMARY

Ourdemoprovidesaninterfacetoviewrecommendations

interfaceconsistsoftwoparts:Apartdisplayingtheprofile

informationoftheurforwhomtherecommendationsare

generated,andapartdisplayingtherecommendationsgenerated.

Aurisspecifiedbyauniquenumber,ur_erfaceisa

webpageandrequiresnothingmorethanabrowrtou.

Copyrightisheldbytheauthor/owner(s).

RecSys’10,September26–30,2010,Barcelona,Spain.

ACM978-1-60558-906-0/10/09.

partprents

informationaboutanexampleurUforwhomrecommendations

aremade;thisincludesasummaryofU’spreferencesbadon

themessagesntandtheidealpartnerprofile,andalsoa

summaryofU’tompartshowsthe

generatedrecommendationsforUandiftheyarelikelytobe

mbolinthelowerleftofarecommendedur

ck

symbolindicatesthatVhasshowninterestinUbyreplying

positivelytoU'smessagewhileagreencrossandblackquestion

ownlistallowsthe

lectionofdifferentrankingcriteria.

Toviewtherecommendationsforaparticularur(example

ur),atextfieldisprovidedtoinputtheur_

processing,theexampleur'sinformationisdisplayedfollowed

mpleur'sinformationincludes

statisticsabouthispreviousinteractionswithotherursandhis

statedpreferencesforthetrainingperiodudbythe

ommendedurshaveabasicprofileand

previousinteractionswiththeexampleurlisted,withalinkto

showapopupboxdisplayingdetailedprofileinformationforthe

particularur.

Thereareveraloptionsforrankingoftherecommendationsand

berof

recommendationstobegeneratedcanalsobelectedsimilarly.

Alinkisalsoprovidedwitheachrecommendedurtoviewthe

anotherwayto

specifyaurtogeneraterecommendationsforwithoutusingthe

ur_id.

TERINFORMATION

LuizPizzato(/People/LuizPizzato)isa

postdoctoralrearcherattheComputerHumanAdapted

Interaction(CHAI)rearchgroupattheUniversityofSydney.

HisworkispartoftheSmartServicesCRCandinvolves

personalization,

activelyinvolvedintherearch,developmentand

implementationoftherecommendersystemforonlinedating

earchinterestsareinInformationRetrievaland

LanguageTechnology.

shotofRECON

LEDGMENTS

TherearchwasfundedbytheSmartServicesCRC.

NCES

[1]o,,,,skaandJ.

eedingsofthe

UMAP-20108thWorkshoponIntelligentTechniquesfor

WebPersonalization,Hawaii,USA

[2]Matching:A

ransactionson

Computer-HumanInteraction,,12(3):401-434