User Location Prediction Based On Similar Living Patterns

The rapid development of the fifth-generation technology and the increasing popularity of various mobile intelligent devices have greatly promoted the development and application of various Location-Based Services(LBS),and also have generated a myriad of user movement trajectories depicting users’ daily behaviors.Among them,user check-ins,due to its the unique characteristics,enables researchers to study and analyze users’ mobile behavior from multiple dimensions,making user location prediction based on user check-in data gradually become a current research hotspot.Since user check-in is a spontaneous behavior with great randomness,it leads to strong sparsity of user check-in data,which is one of the reasons for the low accuracy of current location prediction.In addition,related research shows that there is a significant periodicity of user movement behavior and this periodicity is multi-scale,how to use it effectively is an open subject.To address the above problems,two location prediction methods are proposed in this thesis,specifically,the research work in this paper is as follows:First,in order to alleviate the sparsity of user check-ins and improve the accuracy of location prediction,this thesis proposes a user’s location prediction method based on the pattern of similar living patterns.At first,living pattern representation of each user is extracted according to user check-ins,and different user clusters are obtained by clustering users with similar living patterns.Then,the Points of Interest(POI)embedded learning is performed on the check-ins of each cluster,and the time information is integrated into POI type and POI location representation vector by the improved Huffman tree structure.Finally,input to the Gate Recurrent Unit(GRU)based activity prediction and location prediction model to obtain the final prediction results.Secondly,considering the limitation of the impact of similar user check-ins on location prediction results and the drawbacks of an unidirectional GRU only capturing the previous part of the user’s current trajectory,we utilize the bidirectional GRU structures to capture the context information of the current trajectory,introduce an attention mechanism to capture the periodic movement behavior in the user’s historical check-in track simultaneously,and then integrate the results of the two modules for location prediction.Experimental validations on multiple real user check-in datasets on Foursquare show that both methods proposed in this paper can improve the performance of location prediction compared with existing location prediction methods.