Trajectory Representation Modeling Method Integrating Spatio-temporal Features

Trajectory is a mainly composed of dynamic geographic information.Trajectory modeling and prediction research has always been a hot direction in geographic information science and is also one of the important topics in the field of spatiotemporal data mining.It is of great significance to interpret human behavior patterns,solve urban traffic problems,and explore self-driving technology.In recent years,with the development of the "data-intensive" scientific paradigm,deep learning methods have been adopted by many trajectory modeling studies to help obtain vector representations from irregular trajectory data that can reflect the intrinsic laws of trajectory data,and to improve the accuracy of trajectory classification or prediction.Some achievements have been made in traffic congestion identification,classification of urban functional areas,individual location prediction,etc.However,the existing deep learning methods pay less attention to the dynamic geographic data affected by the law of spatial and temporal changes.Unlike image and voice data,trajectory data always changes with time and space.On the one hand,it is difficult to directly learn from the original data of the trajectory from the deep-level rules of distant neighboring dependencies between the trajectory nodes.On the other hand,the spatio-temporal evolution process of trajectory is also affected by a variety of spatio-temporal features.Studying the fusion of these features and trajectory representation model will not only help to improve the prediction ability of the model,but also help to examine and explore the geospatial laws presented by trajectory representation from a geographical perspective,and provide a new scientific research method for the understanding of the spatio-temporal process of trajectory.In this paper,trajectory in road network space is studied.Based on the current situation of insufficient research on trajectory representation modeling,this paper studies the composition of spatio-temporal features of trajectory,including the sequence of trajectory coordinates,the distant neighboring dependencies under different scales,and the prior spatio-temporal features in different aspects,including time,space,motion and environment.On this basis,the trajectory representation modeling method considering distant neighboring dependencies and the trajectory representation modeling method integrating the prior spatio-temporal features are studied.Based on the typical taxi trajectory data,relevant experiments are designed to verify the validity and applicability of the models.The main research contents and conclusions are as follows:(1)The theoretical basis of trajectory representation modeling is studied.This paper discusses the basic concept and definition of trajectory,the process of spatiotemporal change of trajectory,and the composition of spatio-temporal features.Then,the problem definition of trajectory representation modeling is given,and the definition,acquisition method and evaluation method of trajectory representation vector are described.On this basis,the method of obtaining the distant neighboring dependencies between the trajectory nodes is proposed,and the calculation methods of the temporal features,spatial features,motional features and environmental feature of the trajectories are discussed.(2)The modeling method of trajectory representation considering distant neighboring dependences is studied.Based on the distant neighboring dependencies between the trajectory nodes,a modeling method of trajectory representation considering distant neighboring dependencies is proposed.The experimental results show that:(1)The trajectory representation model considering distant neighboring dependencies is superior to other models in the comparison of multiple evaluation metrics,and it can capture the distant dependence and the neighboring dependence at the same time.(2)Through the trajectory sequence representation vector obtained by the model,we can learn the geographical meaning represented by the two dependencies,and the similarity of the vectors can be used to approximate the spatial relationship between geographical locations.(3)The modeling method of trajectory representation integrating prior features is studied.This paper obtains the trajectory sequence representation vector through three models: the learning model of the trajectory node's prior spatio-temporal feature,the learning model of the trajectory global prior spatio-temporal feature,and the joint training model integrating the above features.The experimental results show that the joint training model has the best performance and more stable advantages,and in the performance of feature expression and prediction,it embodies the combination of node and global features.In addition,the trajectory sequence representation vector can represent the spatial clustering of the trajectory,road and route information,as well as the deeper combination and rule between features.(4)The applicability evaluation of trajectory representation vector is studied.In order to verify the effectiveness and applicability of the method in this paper,the trajectory representation vector is input to the simplest three-layer neural network for prediction.This paper discusses the generalization ability and spatial applicability of the method by analyzing the differences in the trajectory sample characteristics and spatial distribution of the error,accuracy,and stability of the prediction results.Through clustering and spatial statistics of trajectory representation vectors,the clustering characteristics of trajectory representation vectors in geographic space,road space,and OD positions are analyzed.The results of a series of experiments show that:(1)An incomplete pre-order trajectory can be converted into a TSV containing certain complete trajectory information,and can be used as input data of other models.And it can achieve better results on a smaller sample set;(2)TSV's representation ability is better in trajectories with longer length or higher completion;(3)TSV has a wider applicability of trajectory types and spatial applicability;(4)The global prior spatiotemporal feature is helpful for the generalization of trajectory representation vector in space,and the node spatiotemporal feature is helpful for the enhancement of applicability of trajectory representation in the region with dense trajectory.In this paper,by introducing the theory of representation modeling,integrating geographical features and deep learning methods,the ability of capturing and expressing the two spatio-temporal features of distant neighboring dependencies and prior spatio-temporal features is improved.In addition,the applicability of trajectory sequence representation vector is analyzed through experiments.From the perspective of geography,this paper combines the spatio-temporal features of trajectories with representation learning methods.The research results are expected to provide a new scientific research method for revealing and elaborating the internal relationship between trajectory representation and spatio-temporal changes.