Research On Prediction Of Individual Movement Behavior At Urban Scale Based On Deep Learning

With the rapid development of mobile Internet technology,spatial positioning technology,and communication technology,it enables relevant departments to accurately locate and track the location and place of mobile users,and thus can collect a large amount of users' spatio-temporal track data.These data can not only record users' daily activities,but also lay a solid foundation for the development of social network services and spatial positioning technology.In the context of the current rapid development of big data technology,more and more researchers use these spatio-temporal trajectory data for human movement trajectory prediction research,which makes these data have important practical significance and academic value in the fields of social network services,intelligent traffic management and smart city construction.In recent years,artificial intelligence,machine learning,deep learning and other technologies have been developed rapidly,bringing new research ideas for human movement trajectory prediction.For example,deep learning algorithms are used to predict the trajectories of pedestrians and vehicles,or to make location recommendations for pedestrian locations within a certain range,etc.However,it is very difficult to model hundreds of millions of trajectory data,and there are still many challenges to be tackled,such as spatio-temporal trajectories with spatio-temporal information heterogeneity,sparse trajectory data,sparse representation in the model,long-term dependency of trajectory data,etc.To address these series of problems,this paper takes the prediction of individual movement behavior at urban scale as the starting point and spatio-temporal trajectory data as the research object of this paper,and carries out the following research work,with the following details.1.In order to consider the impact of both users' personalized preferences and group preferences on trajectory prediction,a novel next location prediction model based on personalized preferences and group preferences is proposed in this paper.This model considers several key factors affecting group movement behavior,including geographic environment,distance decay and individual spatial activity characteristics,and designs a module that combines a priori statistical information and recurrent neural networks,i.e.,a dynamic spatio-temporal dependence module,which learns the influence of group preferences on users' movement patterns.To learn the personalized preferences of each user,this paper proposes a module based on a bidirectional long-and short-term memory network and an attention mechanism to capture the personalized dynamic preferences of individuals.In addition,this paper proposes a novel graph embedding method to represent each location and its category,which can effectively learn the sequential relationship between visited locations.2.In order to improve the problem that the training of trajectory data based on recurrent neural network is very time-consuming and the spatio-temporal information is not fully exploited,a new method called San Move is proposed in this paper,which is based on self-attentive network to predict the next visit location of users.The specific idea of this model is that firstly,the check-in data of each user in a period of time is obtained and its corresponding whole trajectory is generated,then the whole trajectory of each user is divided into historical trajectory and current trajectory according to certain division rules,and after that the model embeds the historical trajectory and current trajectory into a dense representation.Finally,two self-attention-based modules are applied to capture users' long-term and short-term preferences,respectively.In this model,we design a spatio-temporally guided non-intrusive self-attention(STNOVA)module,which combines non-intrusive self-attention(NOVA)with spatio-temporal information to learn spatio-temporal higher-order information in trajectory data.The research on this model provides a new solution and technical support for individual mobile trajectory prediction and recommendation,and also lays the foundation for the current development in the field of spatio-temporal data mining and spatio-temporal information processing.3.In order to solve the problem that all existing models fail to consider the overall movement trend of users,this paper proposes a next location prediction model integrating spatial clustering information and context-awareness,which takes geographic area information into account and is used to reflect the overall movement trend of users.To address the problems that existing human trajectory prediction algorithms can hardly fully explore the overall movement trend of users and fail to fully utilize the spatial region information of the city,this paper adopts the spatial clustering technique Hi Spatial Cluster to divide the whole city into several geographic regions of different sizes and takes the regional information as the input information of the model.At the same time,the model adopts an attention mechanism to learn users' long-term and short-term preferences respectively,and is used to select the most relevant historical records for the current trajectory.Meanwhile,in order to better capture the influence of other factors in the model on the trajectory prediction,the model introduces multi-task components,such as time prediction and region location prediction to enhance the generality of the model framework and make the prediction results more accurate and better applied.