Research On User Mobility Prediction Method And Application Based On Deep Learning

The location-based service records a large amount of user movement trajectory data during operation.These data not only include the location and time of user movement,but also contain rich semantic information generated by user activities.With the development of the mobile Internet,various applications have generated huge demand for using movement trajectory data to predict user mobility.After investigation and analysis of existing research materials,deep learning has become a hot research topic in computer science in recent years.Compared with traditional prediction algorithms,the deep convolutional neural network model has a good effect on spatial modeling,and the deep recurrent neural network model has a good effect on time series modeling.This thesis uses the advantages of deep learning in time series modeling and spatial modeling to study user mobility prediction methods based on deep learning.The user mobility prediction researched in this thesis is mainly divided into two parts: individual user movement prediction and user group movement prediction.Individual user movement prediction refers to a given historical movement trajectory sequence and predicts the user's movement location at the next moment,which is of great significance for services such as point of interest recommendation and traffic route planning.User group movement prediction is to study and model the movement of the group within a certain geographic range,and to predict the movement trend of the user group at the next moment.It has a major role in the construction of smart cities,urban route planning and public activity warning.The semantic information contained in the trajectory data has a significant impact on individual user movement prediction,and the which is spatio-temporal dependence contained in the data creates a huge challenge for group prediction.Existing deep learning user mobility prediction methods have the problems that data cannot be effectively input into the model,semantic information is not fully utilized,and spatiotemporal dependence cannot be well captured.In response to the above problems,this thesis uses semantic trajectory data encoding and improved deep learning models to effectively and accurately predict user movements.For individual user prediction,this thesis proposes a Semantic and Attention Spatio-temporal Recurrent Model.The model converts the information of each movement point of the user's individual movement trajectory into(grid location,time,semantic information,user ID)quadruple,uses word embedding and other methods to encode the tuple information,and input into the improved recurrent neural network to predict the user's next location.And the model uses an attention mechanism to strengthen the influence of the historical movement locations on the user's next location.For the prediction of user groups,this thesis proposes a proposes a MixedConv spatio-temporal LSTM model.The mixed convolution module is used to capture short-term dependencies and rich spatial information,and then input it into an improved spatio-temporal LSTM to predict the movement of the group of users at the next moment.Finally,experiments on Foursqure sign-in,Beijing Taxi(Taxi BJ)and other datasets and applications on Chongqing Taxi(Taxi CQ)dataset have proved the validity and accuracy of the two models proposed in this thesis.