POI Recommendation With Interactive Behaviors And User Preference Dynamics Embedding

Recently,Location-based Social Networks(LBSNs)have been widely used in people's daily life.People would love to use mobile apps such as Meituan,Dianping to choose Point-of-Interests(POIs)and share their locations through Weibo,Wechat and other social software.It generates the demand for more effective and efficient POI recommendation.Existing research mainly focused on the improvement of the POI recommendation performance by some additional information such as social links,geographical distance,etc.However,further studies are necessary on the integration of these complex influenced factors.Based on two public LBSN datasets Weeplaces and Foursquare,this study analyzes the check-in histories and social networks in datasets.We also discuss on how to improve the efficiency and accuracy of recommending POIs.The main work of this study is as follows:(1)We summarize the characteristics of users' check-in history and social links in Weeplaces and Foursquare and analyze the potential relations between users and POIs.We found that users,in most cases,visit limited places,while each place is only visited by a small number of users.Secondly,users who have direct friend relationships as well as so-called "friend of a friend" in social networks have high check-in similarity with the target user.Finally,the time interval between two consecutive check-ins is short in general,and users are more inclined to move in the nearby places.(2)Based on the data analysis,we propose a framework of POI recommendation which combines user interaction behaviors and user check-in sequence dynamic preference.In the user interaction module,we use the user-user homogeneous graph,POI-POI homogeneous graph and user-POI bipartite graph to learn the embedding representation vector of the user and interest location.In the user check-in sequence dynamic preference module,the convolutional neural network and user attention network are used to get the contextual vector of location and user representation vector.Furthermore,the weight parameter ? is set to combine the two modules to get the candidate recommended POIs for users.(3)Based on the proposed recommendation framework,we conduct scientific experiments on two real-world datasets.Firstly,we analyzed the influence of the weight parameter ? on the experimental results.The appropriate value of parameter ? in the subsequent experiments is selected accordingly.Secondly,we choose four classical recommendation methods as comparison methods and then perform experiments on the datasets.Finally,we analyze the performance of various versions of our model on the datasets.Extensive experiments demonstrate the effectiveness of our model on two real-world datasets.