【基于用户轨迹的兴趣研究的国内外文献综述4000字】

基于用户轨迹的兴趣推荐研究的国内外文献综述本文主要根据用户的行为轨迹来研究用户的下一个兴趣点推荐方法，具体可分为用户行为预测分析和下一个兴趣点推荐方法这两个方面。下面将从这两个方面来阐述己有的相关工作。1.1兴趣点推荐早期对POI推荐的研究主要集中在使用协同过滤(CF)估计用户偏好，特别是基于矩阵分解(MF)的技术[2-5]。这些方法只能对用户的静态首选项建模。例如，当居住在纽约的用户到夏威夷度假时，这些类型的推荐器可能仍然推荐位于纽约的POIs，因为他们无法捕获用户偏好的动态。最近，基于深度学习的方法，如嵌入学习[6-8]，神经协同过滤[9-10]、潜在因子模型和度量学习模型[11-12]在许多推荐系统中都取得了很好的性能。Cheng等人[13]的开创性的提出了一种嵌入个性化马尔科夫链和局部区域的矩阵分解方法。受RNN在顺序数据建模中的成功启发，基于RNN的方法在next-POI推荐领域变得非常普遍。例如，ST-RNN模型扩展了RNN来建模局部时间和空间上下文。CARA通过利用GRU的gate机制捕捉用户的动态偏好。TMCA和STGN分别采用基于LSTM和门控LSTM框架学习时空上下文。DeepMove设计了一种多模态RNN来捕捉顺序转移。以上这些方法的提出，极大的扩展了人们解决兴趣点推荐问题的思路，并为未来相关问题的解决打下基础。1.2下一个兴趣点推荐兴趣点推荐已经吸引了很多产业界和学术界的研究。下一个兴趣点推荐是一般的兴趣点的延伸，根据用户的历史签到信息给用户推荐将要访问的下一个兴趣点列表，下一个兴趣点推荐经常会被看成序列推荐问题。目前已经有很多方法被应用到下一个兴趣点推荐中，例如基于潜在因子模型的方法、基于马尔可夫链模型的方法、嵌入表示模型的方法、以及神经网络模型的方法等。本文主要研究下一个兴趣点推荐问题。下一个兴趣点与传统的一般兴趣点推荐不同，在传统的兴趣点推荐中，类似于一般的商品（图书，电影，音乐）推荐等，通常使用协同过滤的思想给用户推荐用户在未来将要访问的商品，基于矩阵分解的算法是协同过滤算法最先进的算法。该算法是基于相似的用户通常有相似的品味的思想进行兴趣点推荐，它首先用随机初始的向量表示用户和项目的潜在因子，通过对用户历史访问兴趣得到用户和兴趣点的潜在因子表示，在推荐过程中，根据用户和兴趣点的潜在因子的内积得到用户未来访问兴趣点的概率，最终得到用户TOP-K个推荐列表。在一般的兴趣点推荐中，它只是根据用户的访问偏好推荐用户未来访问兴趣点的可能性（比如未来的一天内或者一年），严重忽略了用户历史签到兴趣点之间的序列关系。比如用户按序列顺序访问A，B，C三个景点，在一般的兴趣点推荐中，推荐的三个景点是由用户的偏好得到，它们之间没有任何顺序关系。然而在下一个兴趣点推荐中，用户的序列访问行为通常对于下一个兴趣点也具有很重要的影响，因此在下一个兴趣点中用户的推荐列表是随着签到信息时刻变化，且用户的每次移动都会导致推荐列表剧烈的变化。许多的研究通常会把下一个兴趣点推荐问题看成是序列推荐问题ADDINCSL_CITATION{"citationItems":[{"id":"ITEM-1","itemData":{"DOI":"10.1016/j.ins.2019.12.006","author":[{"dropping-particle":"","family":"Zhang","given":"Lu","non-dropping-particle":"","parse-names":false,"suffix":""},{"dropping-particle":"","family":"Sun","given":"Zhu","non-dropping-particle":"","parse-names":false,"suffix":""},{"dropping-particle":"","family":"Zhang","given":"Jie","non-dropping-particle":"","parse-names":false,"suffix":""},{"dropping-particle":"","family":"Kloeden","given":"Horst","non-dropping-particle":"","parse-names":false,"suffix":""},{"dropping-particle":"","family":"Klanner","given":"Felix","non-dropping-particle":"","parse-names":false,"suffix":""}],"id":"ITEM-1","issued":{"date-parts":[["2020"]]},"page":"169-190","publisher":"ElsevierInc.","title":"ModelinghierarchicalcategorytransitionfornextPOIrecommendationwithuncertaincheck-ins","type":"article-journal","volume":"515"},"uris":["/documents/?uuid=abbd42db-8316-4608-aebb-353efa9d40ef"]}],"mendeley":{"formattedCitation":"^[27]","plainTextFormattedCitation":"[27]","previouslyFormattedCitation":"^[27]"},"properties":{"noteIndex":0},"schema":"/citation-style-language/schema/raw/master/csl-citation.json"}[27]。文献ADDINCSL_CITATION{"citationItems":[{"id":"ITEM-1","itemData":{"abstract":"Venuerecommendationaimstoassistusersbymakingpersonalisedsuggestionsofvenuestovisit,buildingupondataavailablefromlocation-basedsocialnetworks(LBSNs)suchasFoursquare.Aparticularchallengeforthistaskiscontext-awarevenuerecommendation(CAVR),whichadditionallytakesthesurroundingcontextoftheuser(e.g.theuser'slocationandthetimeofday)intoaccountinordertoprovidemorerelevantvenuesuggestions.ToaddressthechallengesofCAVR,wedescribetwoapproachesthatexploitwordembeddingtechniquestoinferthevector-spacerepresentationsofvenues,users'existingpreferences,andusers'contextualpreferences.OurevaluationuponthetestcollectionoftheTREC2015ContextualSuggestiontrackdemonstratesthatwecansignificantlyenhancetheeffectivenessofastate-of-the-artvenuerecommendationapproach,aswellasproducecontext-awarerecommendationsthatareatleastaseffectiveasthetopTREC2015systems.","author":[{"dropping-particle":"","family":"Manotumruksa","given":"Jarana","non-dropping-particle":"","parse-names":false,"suffix":""},{"dropping-particle":"","family":"Macdonald","given":"Craig","non-dropping-particle":"","parse-names":false,"suffix":""},{"dropping-particle":"","family":"Ounis","given":"Iadh","non-dropping-particle":"","parse-names":false,"suffix":""}],"id":"ITEM-1","issued":{"date-parts":[["2016"]]},"page":"2-5","title":"ModellingUserPreferencesusingWordEmbeddingsforContext-AwareVenueRecommendation","type":"article-journal"},"uris":["/documents/?uuid=14873e01-13ec-48c2-aaf1-5253b6844c0b"]}],"mendeley":{"formattedCitation":"^[32]","plainTextFormattedCitation":"[32]","previouslyFormattedCitation":"^[32]"},"properties":{"noteIndex":0},"schema":"/citation-style-language/schema/raw/master/csl-citation.json"}[14]提出了一种用于POI推荐的基于排名的地理因子分解方法，称为RankGeoFM。该模型考虑到签到频率表征了用户的访问偏好，并通过对兴趣点的正确排序来学习因子分解。该模型可以很容易地将不同类型的上下文信息，如地理影响和时间影响纳入其中。由于用户-兴趣点矩阵的稀疏性带来了严峻的挑战。为了应对这一挑战，文献ADDINCSL_CITATION{"citationItems":[{"id":"ITEM-1","itemData":{"DOI":"10.1145/2623330.2623638","ISBN":"9781450329569","abstract":"Point-of-Interest(POI)recommendationhasbecomeanimportantmeanstohelppeoplediscoverattractivelocations.However,extremesparsityofuser-POImatricescreatesaseverechallenge.Tocopewiththischallenge,viewingmobilityrecordsonlocation-basedsocialnetworks(LBSNs)asimplicitfeedbackforPOIrecommendation,wefirstproposetoexploitweightedmatrixfactorizationforthistasksinceitusuallyservescollaborativefilteringwithimplicitfeedbackbetter.Besides,researchershaverecentlydiscoveredaspatialclusteringphenomenoninhumanmobilitybehaviorontheLBSNs,i.e.,individualvisitinglocationstendtoclustertogether,andalsodemonstrateditseffectivenessinPOIrecommendation,thusweincorporateitintothefactorizationmodel.Particularly,weaugmentusers'andPOIs'latentfactorsinthefactorizationmodelwithactivityareavectorsofusersandinfluenceareavectorsofPOIs,respectively.Basedonsuchanaugmentedmodel,wenotonlycapturethespatialclusteringphenomenonintermsoftwo-dimensionalkerneldensityestimation,butwealsoexplainwhytheintroductionofsuchaphenomenonintomatrixfactorizationhelpstodealwiththechallengefrommatrixsparsity.Wethenevaluatetheproposedalgorithmonalarge-scaleLBSNdataset.Theresultsindicatethatweightedmatrixfactorizationissuperiortootherformsoffactorizationmodelsandthatincorporatingthespatialclusteringphenomenonintomatrixfactorizationimprovesrecommendationperformance.©2014ACM.","author":[{"dropping-particle":"","family":"Lian","given":"Defu","non-dropping-particle":"","parse-names":false,"suffix":""},{"dropping-particle":"","family":"Zhao","given":"Cong","non-dropping-particle":"","parse-names":false,"suffix":""},{"dropping-particle":"","family":"Xie","given":"Xing","non-dropping-particle":"","parse-names":false,"suffix":""},{"dropping-particle":"","family":"Sun","given":"Guangzhong","non-dropping-particle":"","parse-names":false,"suffix":""},{"dropping-particle":"","family":"Chen","given":"Enhong","non-dropping-particle":"","parse-names":false,"suffix":""},{"dropping-particle":"","family":"Rui","given":"Yong","non-dropping-particle":"","parse-names":false,"suffix":""}],"container-title":"ProceedingsoftheACMSIGKDDInternationalConferenceonKnowledgeDiscoveryandDataMining","id":"ITEM-1","issued":{"date-parts":[["2014"]]},"page":"831-840","title":"GeoMF:Jointgeographicalmodelingandmatrixfactorizationforpoint-of-interestrecommendation","type":"paper-conference"},"uris":["/documents/?uuid=3e562ae9-ac7e-40a4-91d1-b8b99cae7122"]}],"mendeley":{"formattedCitation":"^[30]","plainTextFormattedCitation":"[30]","previouslyFormattedCitation":"^[30]"},"properties":{"noteIndex":0},"schema":"/citation-style-language/schema/raw/master/csl-citation.json"}[15]首次提出利用加权矩阵分解来解决这一问题，因为它通常能更好地服务于隐式反馈的协同过滤。文献ADDINCSL_CITATION{"citationItems":[{"id":"ITEM-1","itemData":{"DOI":"10.1145/1772690.1772773","ISBN":"9781605587998","abstract":"Recommendersystemsareanimportantcomponentofmanywebsites.Twoofthemostpopularapproachesarebasedonmatrixfactorization(MF)andMarkovchains(MC).MFmethodslearnthegeneraltasteofauserbyfactorizingthematrixoverobserveduser-itempreferences.Ontheotherhand,MCmethodsmodelsequentialbehaviorbylearningatransitiongraphoveritemsthatisusedtopredictthenextactionbasedontherecentactionsofauser.Inthispaper,wepresentamethodbringingbothapproachestogether.OurmethodisbasedonpersonalizedtransitiongraphsoverunderlyingMarkovchains.Thatmeansforeachuseranowntransitionmatrixislearned-thusintotalthemethodusesatransitioncube.Astheobservationsforestimatingthetransitionsareusuallyverylimited,ourmethodfactorizesthetransitioncubewithapairwiseinteractionmodelwhichisaspecialcaseoftheTuckerDecomposition.WeshowthatourfactorizedpersonalizedMC(FPMC)modelsubsumesbothacommonMarkovchainandthenormalmatrixfactorizationmodel.Forlearningthemodelparameters,weintroduceanadaptionoftheBayesianPersonalizedRanking(BPR)frameworkforsequentialbasketdata.Empirically,weshowthatourFPMCmodeloutperformsboththecommonmatrixfactorizationandtheunpersonalizedMCmodelbothlearnedwithandwithoutfactorization.©2010InternationalWorldWideWebConferenceCommittee(IW3C2).","author":[{"dropping-particle":"","family":"Rendle","given":"Steffen","non-dropping-particle":"","parse-names":false,"suffix":""},{"dropping-particle":"","family":"Freudenthaler","given":"Christoph","non-dropping-particle":"","parse-names":false,"suffix":""},{"dropping-particle":"","family":"Schmidt-Thieme","given":"Lars","non-dropping-particle":"","parse-names":false,"suffix":""}],"container-title":"Proceedingsofthe19thInternationalConferenceonWorldWideWeb,WWW'10","id":"ITEM-1","issued":{"date-parts":[["2010"]]},"note":"FromDuplicate1(FactorizingpersonalizedMarkovchainsfornext-basketrecommendation-Rendle,Steffen;Freudenthaler,Christoph;Schmidt-Thieme,Lars)\n\n2012年的FPMC不加地理位置\n\nFromDuplicate2(FactorizingpersonalizedMarkovchainsfornext-basketrecommendation-Rendle,Steffen;Freudenthaler,Christoph;Schmidt-Thieme,Lars)\n\nFPMC模型没有地理位置限制\n\nFPMC是基于马尔科夫链的模型。通过markov链对用户的签到信息建模。使用个性markov链和矩阵分解算法整合序列信息和用户的一般偏好进行next-basket推荐。","page":"811-820","title":"FactorizingpersonalizedMarkovchainsfornext-basketrecommendation","type":"article-journal"},"uris":["/documents/?uuid=12f1a349-c88e-43b9-b6da-9adb731d1013"]}],"mendeley":{"formattedCitation":"^[39]","plainTextFormattedCitation":"[39]","previouslyFormattedCitation":"^[39]"},"properties":{"noteIndex":0},"schema":"/citation-style-language/schema/raw/master/csl-citation.json"}[16]考虑到矩阵分解的方法难于对兴趣点序列信息建模，而基于马尔科夫链的方法能够捕获用户的签到序列信息，因此文献ADDINCSL_CITATION{"citationItems":[{"id":"ITEM-1","itemData":{"DOI":"10.1145/1772690.1772773","ISBN":"9781605587998","abstract":"Recommendersystemsareanimportantcomponentofmanywebsites.Twoofthemostpopularapproachesarebasedonmatrixfactorization(MF)andMarkovchains(MC).MFmethodslearnthegeneraltasteofauserbyfactorizingthematrixoverobserveduser-itempreferences.Ontheotherhand,MCmethodsmodelsequentialbehaviorbylearningatransitiongraphoveritemsthatisusedtopredictthenextactionbasedontherecentactionsofauser.Inthispaper,wepresentamethodbringingbothapproachestogether.OurmethodisbasedonpersonalizedtransitiongraphsoverunderlyingMarkovchains.Thatmeansforeachuseranowntransitionmatrixislearned-thusintotalthemethodusesatransitioncube.Astheobservationsforestimatingthetransitionsareusuallyverylimited,ourmethodfactorizesthetransitioncubewithapairwiseinteractionmodelwhichisaspecialcaseoftheTuckerDecomposition.WeshowthatourfactorizedpersonalizedMC(FPMC)modelsubsumesbothacommonMarkovchainandthenormalmatrixfactorizationmodel.Forlearningthemodelparameters,weintroduceanadaptionoftheBayesianPersonalizedRanking(BPR)frameworkforsequentialbasketdata.Empirically,weshowthatourFPMCmodeloutperformsboththecommonmatrixfactorizationandtheunpersonalizedMCmodelbothlearnedwithandwithoutfactorization.©2010InternationalWorldWideWebConferenceCommittee(IW3C2).","author":[{"dropping-particle":"","family":"Rendle","given":"Steffen","non-dropping-particle":"","parse-names":false,"suffix":""},{"dropping-particle":"","family":"Freudenthaler","given":"Christoph","non-dropping-particle":"","parse-names":false,"suffix":""},{"dropping-particle":"","family":"Schmidt-Thieme","given":"Lars","non-dropping-particle":"","parse-names":false,"suffix":""}],"container-title":"Proceedingsofthe19thInternationalConferenceonWorldWideWeb,WWW'10","id":"ITEM-1","issued":{"date-parts":[["2010"]]},"note":"FromDuplicate1(FactorizingpersonalizedMarkovchainsfornext-basketrecommendation-Rendle,Steffen;Freudenthaler,Christoph;Schmidt-Thieme,Lars)\n\n2012年的FPMC不加地理位置\n\nFromDuplicate2(FactorizingpersonalizedMarkovchainsfornext-basketrecommendation-Rendle,Steffen;Freudenthaler,Christoph;Schmidt-Thieme,Lars)\n\nFPMC模型没有地理位置限制\n\nFPMC是基于马尔科夫链的模型。通过markov链对用户的签到信息建模。使用个性markov链和矩阵分解算法整合序列信息和用户的一般偏好进行next-basket推荐。","page":"811-820","title":"FactorizingpersonalizedMarkovchainsfornext-basketrecommendation","type":"article-journal"},"uris":["/documents/?uuid=12f1a349-c88e-43b9-b6da-9adb731d1013"]}],"mendeley":{"formattedCitation":"^[39]","plainTextFormattedCitation":"[39]","previouslyFormattedCitation":"^[39]"},"properties":{"noteIndex":0},"schema":"/citation-style-language/schema/raw/master/csl-citation.json"}[16]引进了个性化的马尔科夫链，通过结合矩阵分解和一阶马尔科夫两者的优势分别对用户的偏好和访问的商品的序列信息进行建模，提出个性化的马尔科夫链分解的算法（FPMC）对用户购买商品进行建模，用于给用户推荐用户下一个购买的商品，同时通过在数据集上的实验结果也显示出FPMC的方法要优于一般的矩阵分解算法和单纯的基于马尔科夫模型的算法。与下一个商品推荐不同，由于兴趣点推荐中不仅受到序列信息和用户偏好的影响，用户的地理位置信息对于下一个兴趣点推荐也有一定程度的影响，文献ADDINCSL_CITATION{"citationItems":[{"id":"ITEM-1","itemData":{"ISBN":"9781577357384","ISSN":"10450823","abstract":"TherapidlygrowingofLocation-basedSocialNetworks(LBSNs)providesavastamountofcheck-indata,whichenablesmanyservices,e.g.,point-of-interest(POI)recommendation.Inthispaper,westudythenextnewPOIrecommendationprobleminwhichnewPOIswithrespecttousers'currentlocationaretoberecommended.Thechallengeliesinthedifficultyinpreciselylearningusers'sequentialinformationandpersonalizingtherecommendationmodel.Tothisend,weresorttotheMetricEmbeddingmethodfortherecommendation,whichavoidsdrawbacksoftheMatrixFactorizationtechnique.Weproposeapersonalizedrankingmetricembeddingmethod(PRME)tomodelpersonalizedcheck-insequences.WefurtherdevelopaPRME-Gmodel,whichintegratessequentialinformation,individualpreference,andgeographicalinfluence,toimprovetherecommendationperformance.Experimentsontworeal-worldLBSNdatasetsdemonstratethatournewalgorithmoutperformsthestate-of-the-artnextPOIrecommendationmethods.","author":[{"dropping-particle":"","family":"Feng","given":"Shanshan","non-dropping-particle":"","parse-names":false,"suffix":""},{"dropping-particle":"","family":"Li","given":"Xutao","non-dropping-particle":"","parse-names":false,"suffix":""},{"dropping-particle":"","family":"Zeng","given":"Yifeng","non-dropping-particle":"","parse-names":false,"suffix":""},{"dropping-particle":"","family":"Cong","given":"Gao","non-dropping-particle":"","parse-names":false,"suffix":""},{"dropping-particle":"","family":"Chee","given":"YeowMeng","non-dropping-particle":"","parse-names":false,"suffix":""},{"dropping-particle":"","family":"Yuan","given":"Quan","non-dropping-particle":"","parse-names":false,"suffix":""}],"container-title":"IJCAIInternationalJointConferenceonArtificialIntelligence","id":"ITEM-1","issued":{"date-parts":[["2015"]]},"note":"PRME-G\n本文唯一的创新，通过嵌入的方法对兴趣点序列偏好和用户偏好建模。\n本文使用了2个嵌入向量空间。序列变换嵌入向量空间，用户和兴趣点向量嵌入空间。分别对用户偏好和序列偏好建模。\n针对FPMC的创新有一点。比如A-B，B-C。由于MC方法的独立性，A到C的潜在转移可能性被忽略。\n但是ME使用嵌入方法，A，B，C用嵌入空间表示，A，C的转移可能性也可以被捕获。\n因此ME的方法由于MF的方法。","page":"2069-2075","title":"PersonalizedrankingmetricembeddingfornextnewPOIrecommendation","type":"paper-conference","volume":"2015-Janua"},"uris":["/documents/?uuid=a3d32887-ce52-489b-b7ff-eb79ffda9c4b"]}],"mendeley":{"formattedCitation":"^[42]","plainTextFormattedCitation":"[42]","previouslyFormattedCitation":"^[42]"},"properties":{"noteIndex":0},"schema":"/citation-style-language/schema/raw/master/csl-citation.json"}[17]提出了PRME模型，该模型同时考虑了用户偏好和序列信息的影响，并利用度量嵌入的方法对用户的签到数据信息进行建模。它使用两个嵌入空间，用户偏好嵌入空间和序列信息嵌入空间。在用户偏好嵌入空间中用户与兴趣点在嵌入空间的距离表示用户对兴趣点的偏好。在序列嵌入空间中，兴趣点之间的的嵌入距离表示兴趣点之间转化的可能。然后通过权重的形式整合用户偏好和序列信息的影响。为了整合地理位置影响，文献ADDINCSL_CITATION{"citationItems":[{"id":"ITEM-1","itemData":{"ISBN":"9781577357384","ISSN":"10450823","abstract":"TherapidlygrowingofLocation-basedSocialNetworks(LBSNs)providesavastamountofcheck-indata,whichenablesmanyservices,e.g.,point-of-interest(POI)recommendation.Inthispaper,westudythenextnewPOIrecommendationprobleminwhichnewPOIswithrespecttousers'currentlocationaretoberecommended.Thechallengeliesinthedifficultyinpreciselylearningusers'sequentialinformationandpersonalizingtherecommendationmodel.Tothisend,weresorttotheMetricEmbeddingmethodfortherecommendation,whichavoidsdrawbacksoftheMatrixFactorizationtechnique.Weproposeapersonalizedrankingmetricembeddingmethod(PRME)tomodelpersonalizedcheck-insequences.WefurtherdevelopaPRME-Gmodel,whichintegratessequentialinformation,individualpreference,andgeographicalinfluence,toimprovetherecommendationperformance.Experimentsontworeal-worldLBSNdatasetsdemonstratethatournewalgorithmoutperformsthestate-of-the-artnextPOIrecommendationmethods.","author":[{"dropping-particle":"","family":"Feng","given":"Shanshan","non-dropping-particle":"","parse-names":false,"suffix":""},{"dropping-particle":"","family":"Li","given":"Xutao","non-dropping-particle":"","parse-names":false,"suffix":""},{"dropping-particle":"","family":"Zeng","given":"Yifeng","non-dropping-particle":"","parse-names":false,"suffix":""},{"dropping-particle":"","family":"Cong","given":"Gao","non-dropping-particle":"","parse-names":false,"suffix":""},{"dropping-particle":"","family":"Chee","given":"YeowMeng","non-dropping-particle":"","parse-names":false,"suffix":""},{"dropping-particle":"","family":"Yuan","given":"Quan","non-dropping-particle":"","parse-names":false,"suffix":""}],"container-title":"IJCAIInternationalJointConferenceonArtificialIntelligence","id":"ITEM-1","issued":{"date-parts":[["2015"]]},"note":"PRME-G\n本文唯一的创新，通过嵌入的方法对兴趣点序列偏好和用户偏好建模。\n本文使用了2个嵌入向量空间。序列变换嵌入向量空间，用户和兴趣点向量嵌入空间。分别对用户偏好和序列偏好建模。\n针对FPMC的创新有一点。比如A-B，B-C。由于MC方法的独立性，A到C的潜在转移可能性被忽略。\n但是ME使用嵌入方法，A，B，C用嵌入空间表示，A，C的转移可能性也可以被捕获。\n因此ME的方法由于MF的方法。","page":"2069-2075","title":"PersonalizedrankingmetricembeddingfornextnewPOIrecommendation","type":"paper-conference","volume":"2015-Janua"},"uris":["/documents/?uuid=a3d32887-ce52-489b-b7ff-eb79ffda9c4b"]}],"mendeley":{"formattedCitation":"^[42]","plainTextFormattedCitation":"[42]","previouslyFormattedCitation":"^[42]"},"properties":{"noteIndex":0},"schema":"/citation-style-language/schema/raw/master/csl-citation.json"}[17]进一步通过地理位置作为权重信息通过指数衰减的形式控制用户访问下一个兴趣点的不同偏好。文献ADDINCSL_CITATION{"citationItems":[{"id":"ITEM-1","itemData":{"DOI":"10.1145/2983323.2983711","ISBN":"9781450340731","abstract":"Withtherapidprevalenceofsmartmobiledevicesandthedramaticproliferationoflocation-basedsocialnetworks(LB-SNs),location-basedrecommendationhasbecomeanimportantmeanstohelppeoplediscoverattractiveandinterestingpointsofinterest(POIs).However,extremesparsityofuser-POImatrixandcold-startissuecreateseverechallenges,causingCF-basedmethodstodegradesignificantlyintheirrecommendationperformance.Moreover,location-basedrecommendationrequiresspatiotemporalcontextawarenessanddynamictrackingoftheuser'slatestpreferencesinareal-timemanner.Toaddressthesechallenges,westandonrecentadvancesinembeddinglearningtechniquesandproposeagenericgraph-basedembeddingmodel,calledGE,inthispaper.GEjointlycapturesthesequentialeffect,geographicalinfluence,temporalcycliceffectandsemanticeffectinaunifiedwaybyembeddingthefourcorrespondingrelationalgraphs(POI-POI,POI-Region,POI-TimeandPOI-Word)intoasharedlowdimensionalspace.Then,tosupportreal-timerecommendation,wedevelopanoveltime-decaymethodtodynamicallycomputetheuser'slatestpreferencesbasedontheembeddingofhis/herchecked-inPOIslearntinthelatents-pace.Weconductextensiveexperimentstoevaluatetheperformanceofourmodelontworeallarge-scaledatasets,andtheexperimentalresultsshowitssuperiorityoverothercompetitors,especiallyinrecommendingcold-startPOIs.Besides,westudythecontributionofeachfactortoimprovelocation-basedrecommendation,andfindthatbothsequentialeffectandtemporalcycliceffectplaymoreimportantrolesthangeographicalinfluenceandsemanticeffect.","author":[{"dropping-particle":"","family":"Xie","given":"Min","non-dropping-particle":"","parse-names":false,"suffix":""},{"dropping-particle":"","family":"Yin","given":"Hongzhi","non-dropping-particle":"","parse-names":false,"suffix":""},{"dropping-particle":"","family":"Wang","given":"Hao","non-dropping-particle":"","parse-names":false,"suffix":""},{"dropping-particle":"","family":"Xu","given":"Fanjiang","non-dropping-particle":"","parse-names":false,"suffix":""},{"dropping-particle":"","family":"Chen","given":"Weitong","non-dropping-particle":"","parse-names":false,"suffix":""},{"dropping-particle":"","family":"Wang","given":"Sen","non-dropping-particle":"","parse-names":false,"suffix":""}],"container-title":"InternationalConferenceonInformationandKnowledgeManagement,Proceedings","id":"ITEM-1","issued":{"date-parts":[["2016"]]},"note":"GE\n兴趣点图嵌入的方法\n通过嵌入四个图，poi-poi，poi-region，poi-time，poi-word空间。","page":"15-24","title":"Learninggraph-basedpoiembeddingforlocation-basedrecommendation","type":"paper-conference","volume":"24-28-Octo"},"uris":["/documents/?uuid=350c5ac7-4566-4d6f-9437-47fdb13e97a9"]}],"mendeley":{"formattedCitation":"^[43]","plainTextFormattedCitation":"[43]"},"properties":{"noteIndex":0},"schema":"/citation-style-language/schema/raw/master/csl-citation.json"}[18]针对因子分解方法无法对时空上下文和序列信息建模问题，提出了一种基于图嵌入的模型，叫做GE，该模型能够同时捕获序列影响，地理影响，时空周期影响以及语义信息。它通过一种统一的方式将四对相应的关系图（兴趣点-兴趣点，兴趣点-区域，兴趣点-时间，兴趣点-单词）嵌入到共享的低维空间中。之后根据经过训练后的空间兴趣点嵌入向量信息和时间衰减的影响计算用户访问不同兴趣点的偏好。文献ADDINCSL_CITATION{"citationItems":[{"id":"ITEM-1","itemData":{"abstract":"Weapplyrecurrentneuralnetworks(RNN)onanewdomain,namelyrecommendersystems.Real-liferecommendersystemsoftenfacetheproblemofhavingtobaserecommendationsonlyonshortsession-baseddata(e.g.asmallsportswarewebsite)insteadoflonguserhistories(asinthecaseofNetflix).Inthissituationthefrequentlypraisedmatrixfactorizationapproachesarenotaccurate.Thisproblemisusuallyovercomeinpracticebyresortingtoitem-to-itemrecommendations,i.e.recommendingsimilaritems.Wearguethatbymodelingthewholesession,moreaccuraterecommendationscanbeprovided.WethereforeproposeanRNN-basedapproachforsession-basedrecommendations.OurapproachalsoconsiderspracticalaspectsofthetaskandintroducesseveralmodificationstoclassicRNNssuchasarankinglossfunctionthatmakeitmoreviableforthisspecificproblem.Experimentalresultsontwodata-setsshowmarkedimprovementsoverwidelyusedapproaches.","author":[{"dropping-particle":"","family":"Hidasi","given":"Balázs","non-dropping-particle":"","parse-names":false,"suffix":""},{"dropping-particle":"","family":"Karatzoglou","given":"Alexandros","non-dropping-particle":"","parse-names":false,"suffix":""},{"dropping-particle":"","family":"Baltrunas","given":"Linas","non-dropping-particle":"","parse-names":false,"suffix":""},{"dropping-particle":"","family":"Tikk","given":"Domonkos","non-dropping-particle":"","parse-names":false,"suffix":""}],"container-title":"4thInternationalConferenceonLearningRepresentations,ICLR2016-ConferenceTrackProceedings","id":"ITEM-1","issued":{"date-parts":[["2016"]]},"note":"GRU4Rec","page":"1-10","title":"Session-basedrecommendationswithrecurrentneuralnetworks","type":"article-journal"},"uris":["/documents/?uuid=9b6f6068-146e-44f4-a335-368eec8f82bf"]}],"mendeley":{"formattedCitation":"^[47]","plainTextFormattedCitation":"[47]","previouslyFormattedCitation":"^[46]"},"properties":{"noteIndex":0},"schema":"/citation-style-language/schema/raw/master/csl-citation.json"}[19]提出GRU4Rec整合序列信息进行商品模型，它是最初的尝试将GRU模型引入到推荐系统中。由于在现实生活中经常会面临短会话的问题，在基于会话的模型中，没有用户的ID存在，只有用户的一段匿名的签到记录。该模型通过GRU模型对用户的最近的一段购买记录建模。作者针对GRU模型提出了两个创新，一是通过对GRU模型训练的批次改进减少GRU模型的训练时，二是提出了TOP1损失函数，相比于传统的BPR算法具有更好的表现效果。文献ADDINCSL_CITATION{"citationItems":[{"id":"ITEM-1","itemData":{"DOI":"10.1109/TSMC.2014.2327053","ISSN":"10834427","abstract":"Withtherecentsurgeoflocationbasedsocialnetworks(LBSNs),activitydataofmillionsofusershasbecomeattainable.Thisdatacontainsnotonlyspatialandtemporalstampsofuseractivity,butalsoitssemanticinformation.LBSNscanhelptounderstandmobileusers'spatialtemporalactivitypreference(STAP),whichcanenableawiderangeofubiquitousapplications,suchaspersonalizedcontext-awarelocationrecommendationandgroup-orientedadvertisement.However,modelingsuchuser-specificSTAPneedstotacklehigh-dimensionaldata,i.e.,user-location-time-activityquadruples,whichiscomplicatedandusuallysuffersfromadatasparsityproblem.Inordertoaddressthisproblem,weproposeaSTAPmodel.Itfirstmodelsthespatialandtemporalactivitypreferenceseparately,andthenusesaprinciplewaytocombinethemforpreferenceinference.Inordertocharacterizetheimpactofspatialfeaturesonuseractivitypreference,weproposethenotionofpersonalfunctionalregionandrelatedparameterstomodelandinferuserspatialactivitypreference.InordertomodeltheusertemporalactivitypreferencewithsparseuseractivitydatainLBSNs,weproposetoexploitthetemporalactivitysimilarityamongdifferentusersandapplynonnegativetensorfactorizationtocollaborativelyinfertemporalactivitypreference.Finally,weputforwardacontextawarefusionframeworktocombinethespatialandtemporalactivitypreferencemodelsforpreferenceinference.Weevaluateourproposedapproachonthreereal-worlddatasetscollectedfromNewYorkandTokyo,andshowthatourSTAPmodelconsistentlyoutperformsthebaselineapproachesinvarioussettings.","author":[{"dropping-particle":"","family":"Yang","given":"Dingqi","non-dropping-parti