Analysis For Imputing Location Distribution Series Based On Bidirectional Recurrent Networks

With the popularity of mobile Internet and the ubiquity of mobile devices to obtain geographic coordinates,we can obtain massive user's geographic coordinates data.With these geographic coordinates data,we can reveal the area of user's daily activity.However,due to the randomness of the user's behavior in reporting position coordinates,the obtained geographic coordinate data is discontinuous and sparse,which provides a great challenge for us to analyze the position distribution of users in different periods.At the same time,most of the research based on trajectory is aimed at the prediction of a certain location point,which is usually the point of interest on the map,such as scenic spots,parks or business districts.However,it is impossible to solve the problem of missing trajectory by predicting a certain position point.There are a variety of possibilities for the location of the user during a certain missing trajectory period,so this paper adopts the location distribution instead of the position point to impute the trajectory in the missing trajectory time period.This paper proposes a trajectory imputing method for missing trajectories caused by sparse trajectories.The work done is as follows:1.This paper proposes to use Google S2 method to discretize a large number of user coordinates.Meanwhile,according to the population distribution of the city,this paper aggregate the low frequency areas which can achieve discretization with different levels of Google S2 coding and reduce the size of location coding.2.Different from most trajectory studies in which embedding is adaptive training with model training,this paper designs an embedding model and adds the hierarchical relationship between location points to learn the potential representation between locations.3.In this paper,an improved bidirectional recurrent model is proposed to capture the inherent time changes in the location distribution,so as to impute the trajectory.4.The experiment data in this paper is more than 10 million check-ins,which verified the effectiveness of the model.The experimental results show that the discrete representation based on Google S2 can reduce the calculation scale of the model effectively and thus reduce the time cost on the model.Trajectory embedding based on hierarchical relationship and trajectory data can capture the relationship between location points.Finally,the improved bidirectional recurrent network can impute the missing trajectory.