impression discount model in rs

《impression discount model in rs》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、2014年8月12日impressiondiscountmodelinrsModelingImpressionDiscountinginLarge-scaleRecommenderSystems摘要Userbehavioralanalysisanduserfeedback(bothexplicitandimplicit)modelingarecrucialfortheimprovementofanyonlinerecommendersystem.Themaincontributionsofthispaperinclude:(1)large-scaleanalysisofim

2、pressiondatafromLinkedInandKDDCup;数据分析(2)novelanti-noiseregressiontechniques,anditsapplicationtolearnfourdifferentimpressiondiscountingfunctionsincludinglineardecay,inversedecay,exponentialdecay,andquadraticdecay;(3)applyingtheseimpressiondiscountingfunctionstoLinkedIn’s“PeopleYouMayKnow”a

3、nd“SuggestedSkillsEndorsements”recommendersystems.1.介绍Intheimpressiondiscountingproblemweaimtomaximizeconversionofrecommendeditemsgeneratedbyarecommendersystembyapplyingadiscountingfactor,derivedfrompastimpressions,ontopofscoresgeneratedbytherecommendersystem.两个挑战(1)如何结合用户展示和反馈数据，构建有效的响应模型

4、(2)howcanthemodelbeappliedtoimprovetheperformanceofexistingrecommendersystems?thenumberoftimesanitemisimpressedorrecommendedtoauser;whentheitemwasimpressed,andfrequencyofuservisitsonthesiteoruserseeinganyoftherecommendeditems.Themaincontributionsofthispaperareasfollows:file:///C:/Users/afa

5、n/appdata/local/temp/63.html1/42014年8月12日impressiondiscountmodelinrs(1)Performlargescalecorrelationstudiesbetweenimpressionsignalsandconversionrateofimpresseditems;(2)Designeffectiveimpressiondiscountingmodelsbasedonlinear/multiplicativeaggregation,andproposenovelanti-noiseregressionmodelt

6、odealwiththedatasparsityproblem;(3)Evaluatetheseregressionmodelsonreal-worldrecommendationsystemssuchas“PeopleYouMayKnow”and“SuggestedSkillsEndorsements”todemonstratetheireffectivenessbothinofflineanalysisandinonlinesystemsbyA/Btesting.2.IMPRESSIONDATAINLARGE-SCALERECOMMENDERSYSTEMS(1)Form

7、alizingImpressions展示数据规范化T=(user;item;conversion;[behavior1;behavior2;…];t;R)用户id，推荐item，conversion布尔型描述用户是否产生行为，观察交互行为类型，t展示时间，R推荐分值Behaviors.followa“see-think-do”procedure.1.LastSeen,用户最近一次展示和当前展示之间的时间差；2.ImpCount，当前展示之前的历史展示次数之和；3.Position，item展现的位置；4.UserF