With the development of mobile devices and GPS positioning technology,more and more Location-Based Social Network(LBSN)applications have been designed and produced.They have accumulated a large amount of check-in data,providing the possibility to perform recommended tasks using deep learning techniques.Point-of-Interest(POI)recommendation is one of the most important service in LBSN.It can not only improve the user experience,but also brings commercial value to suppliers.However,the POI recommendation is more challenging than the traditional item recommendation due to data sparsity and the lack of negative feedback,too many influencing factors and dynamic variability or periodicity of user's preference.Most of the traditional POI recommendation models focused on modeling the frequency and time of user's check-in,or recreated the recommendation method combining time and space context.People rarely pay attention to the user behavior association in the check-in sequence of POI.Besides,they didn't make progress capturing the dynamic variability or periodicity of user preference.Our study found out some implicit relevance between POI check-in sequence modeling and natural language processing,and considered the POI recommendation problem as a part of the sequence-aware recommendation.Firstly,in order to better understand sequence features like users' preference in long-terms or short-terms,the user's check-in data was divided into sequences by using sequence data splitting rules.After that,we applied recurrent neural network to model the POI check-in sequence.We used context information,attention mechanism,ranking loss function and appropriate negative sampling method to make recurrent neural network adapted to the scenes of POI recommendation.Then we proposed two POI recommendation models,named Context-Aware Point-of-Interest Recommendation Model(DTCPS)and Long Short-Attention Point-of-Interest Recommendation Model(LSAPR),and carried out extensive off-line experiments on two real world datasets.Finally,we applied these two models to real project and develop an Android application,TourGuide.Our experimental results show that our two models both outperform state-of-the-art baselines on both datasets.These two models have been deployed in TourGuide successfully,and the realtime personalized recommendation of POI has been achieved in the application. |