1
FinalReportbyJarvisThefollowingreportwascompletedbytheprojectteammembersintheendofOctober2017asafinalreportforthefirstincubationphase.Thisdoesnotmeanthattheincubationnortheprojectareatafinalstage.WhatwasyourVision?Chatbotsareanewwaytointeractwithcompanies,productsandservices.Theproductisthesame,buttheinterfacebetweentheproductandtheuserisdifferent.Accesstoinformationhaschangedintherecentyearsthankstothehugeadvancesinpredictivealgorithms.Therelevantinformationisdisplayedtotheuserevenbeforeherequestsit.Everythingisbasedonthetastesandpreferencesoftheuser.Ourprojectisthecreateapersonalizedrecommendersystemforcookingrecipesavailablethroughafriendlyconversationalinterface.Whatgoalsdidyousetfortheincubationtimeframe?
– Buildarecommendersystembasedonadatasetofrecipesanduserinteractions– Buildachatbotinterfacethatwillguidetheuserthroughtherecommendation
processWhatopendatadidyouuseorgenerate?
– weusetheSwissFoodCompositionDatabase:http://www.naehrwertdaten.ch– we’llgeneratedataaboutpeople’sfoodconsumption
Whatdidyouaccomplish?
– Scrapalltherecipesfromtheallrecipeswebsite(45krecipes,~4.5Minteractionswith~1Musers)
– BuildachatbotprototypeofJarviswithDialogflow(ourfirstideas,butthefinalgoalchangedsincethen)
– GatherinformationaboutchatbotdesignandrecommendersystemsHowdidOpenDatasupportyou?
– Withfinancing– AmeetingwithHannesGassert(29.07.2017):brainstorming,strategies,ideation– AmeetingwithThomasRippel(10.08.2017):brainstorming,knowledgesharing
FeedbackforOpenData
– ThestrategicmeetingwithHanneswasreallyinterestingandhelpfulinthewaythatheknowswhathesaysandhasalotofexperienceinthedomain.
2
– Thewayyouadapttoeachteamisreallygood(eg.wedonotneedofficessothemoneyisusedforsomethingmoreusefulforthesuccessoftheproject,andsoon)
Whatarethenextsteps?
– Design,developmentandvalidationofthereciperecommendersystemmodelusingcollecteddata.~15hours*person/weekforthenext10weeks,withtheassistanceofLSIRlabatEPFL.Thecollecteddatacontainsbothimplicit(madeit/reviews)andexplicitfeedback(ratings)fromusertorecipes.
TherecommendermodelwillbebasedonBayesianPersonalizedRanking,usingimplicitfeedback,whichisexpectedtooutperformstandardlearningtechniques[1].ThecurrentdevelopmentseemstoleadtoaMatrix-factorization[2]basedmodel.
Amodelusingtheexplicitrankingsdatamightbedevelopedinparallelinordertoevaluateandcomparetheirrespectiveperformancesandvalidatethechoiceofusingimplicitdata.Otherwaysofevaluationthemodelwillbeexplored[3].
Rendle[4][5]introducesthefactorizationmachines,anapproachcombiningthepoweroffactorizationmodelswiththegeneralityofstandardfeaturesengineering.Sincethecollecteddatacontainslotsoffeaturesforrecipesitwouldbeinterestingtoimplementitinourmodelandcomparetheresults.
LatestresearchesonNeuralNetworkssuggestthattheuseofanunderlying(deep)neuralnetworksoutperformsstandardfactorizationmethodsforrecommendersystem[6].Onthelong-termitmightbeinterestingtotrysomethingandseeifitisactuallyofanyhelptoenhancetheperformanceofourmodel,butatsuchanearlystagebeginningwithafactorizationbasedmodelisbest.
– Integrateconstraintintherecommendersystem,allowingforexampleto
proposeasetofrecipesthatcomplywithspecificnutritiongoals.(Vegan,Low-meat,WHO’snutritionalguidelines)
– Adapttherecommendersystemtolegallyavailabledata(TBD).– Creationofthepersonalityofthechatbot– Designoftheconversationalflow– Integrationoftherecommenderinthechatbot.Sincewewillhaveno
informationabouttheusersandrecipesatthebeginning,wewillneedtoimplementasmartcold-startrecommendationsstrategy.[7]proposessuchastrategythatbasicallyconsistinmappingusersoritemsfeaturestothelatentfactorsofourmodel.Weneedtodesignasmartstrategytoaskusersfortheirpreference,forexamplebyproposingthemrecipes/ingredientsandletthemswipeleft/rightaccordingtotheirpreferencesanduseittobuildaprofile.Withasufficientamountofuserwecouldthenusetheprofiletomapthemtothelatentfactorsofourmodelandbeabletorecommendrecipestoanewuser.
– FindanewnameandbrandimagefortheJarvisproject– Regardingopendata,we’llbereleasingaggregatedandanonymizeddata
generatedbyourplatformcontainingusersinteractions,behavioursandevolutions.Thisdatawillbefreelyavailableonline,andwouldprovideaninterestingdatasetforscientisttryingtofindpatternsinpeople’sfoodconsumption.
3
– Expectedtimeframe:January2018:– endoftherecommendersystemprototype– endofthebasicchatbotFebruary2018:– adaptingtherecommendersystemtonewrecipes(legallyavailable)– integratingtherecommendersystemtothechatbotMarch2018:– firsttestofthebotwithrealusers– creationofbranding&launchstrategy)
TeamMembers&Contact
– JackyCasas,[email protected],@jackycasas_– NathanQuinteiro,[email protected],@nathan_quint
OpenDataContactNikkiBöhler,ProjectLeaderOpenData.ch,[email protected]:[1]:Rendle,Freudenthaler,GantnerandSchmidt-Thieme-BPR:BayesianPersonalizedRankingfromImplicitFeedbackhttps://arxiv.org/pdf/1205.2618.pdf[2]:SrebroandJaakkola-WeightedLow-RankApproximationshttps://www.aaai.org/Papers/ICML/2003/ICML03-094.pdf[3]:ShaniandGunawardana-EvaluatingRecommendationSystemshttp://www.bgu.ac.il/~shanigu/Publications/EvaluationMetrics.17.pdf[4]:Rendle-FactorizationMachineshttps://pdfs.semanticscholar.org/2ef7/d506b25731d0f3ec0c8f90b718b6e5bbd069.pdf[5]:Rendle-FactorizationMachineswithlibFMhttp://www.csie.ntu.edu.tw/~b97053/paper/Factorization%20Machines%20with%20libFM.pdf[6]:Xiangnan,Lizi,Hanwang,Liqiang,XiaandTat-Seng-NeuralCollaborativeFilteringhttp://papers.www2017.com.au.s3-website-ap-southeast-2.amazonaws.com/proceedings/p173.pdf[7]:Gantner,Drumond,Freudenthaler,RendleandSchmidt-Thiem-LearningAttribute-to-FeatureMappingsforCold-StartRecommendationshttps://www.semanticscholar.org/paper/Learning-Attribute-to-Feature-Mappings-for-Cold-St-Gantner-Drumond/1535d5db7078c85f0e2d565860a0fb4053a6090c