简介:Thewidespreadoflocation-basedsocialnetworksbringsaboutahugevolumeofusercheck-indata,whichfacilitatestherecommendationofpointsofinterest(POIs).Recentadvancesondistributedrepresentationshedlightonlearninglowdimensionaldensevectorstoalleviatethedatasparsityproblem.CurrentstudiesonrepresentationlearningforPOIrecommendationembedbothusersandPOIsinacommonlatentspace,andusers'preferenceisinferredbasedonthedistance/similaritybetweenauserandaPOI.SuchanapproachisnotinaccordancewiththesemanticsofusersandPOIsastheyareinherentlydifferentobjects.Inthispaper,wepresentanoveltranslation-based,timeandlocationaware(TransTL)representation,whichmodelsthespatialandtemporalinformationasarelationshipconnectingusersandPOIs.Ourmodelgeneralizestherecentadvancesinknowledgegraphembedding.Thebasicideaisthattheembeddingofa〈time,location〉paircorrespondstoatranslationfromembeddingsofuserstoPOIs.SincethePOIembeddingshouldbeclosetotheuserembeddingplustherelationshipvector,therecommendationcanbeperformedbyselectingthetop-kPOIssimilartothetranslatedPOI,whichareallofthesametypeofobjects.Weconductextensiveexperimentsontworeal-worlddata.sets.TheresultsdemonstratethatourTransTLmodelachievesthestate-of-the-artperformance.Itisalsomuchmorerobusttodatasparsitythanthebaselines.