Trajectory Linking Across Heterogeneous Trajectory Datasets

Nowadays,with the rapid development of various smart devices and mobile In-ternet,our daily activities are being perceived and recorded through different means at all times.These hardware and software equipments record the spatial information of users at specific times using different methods,thus resulting in an enormous amount of human trajectory data.The human trajectory data is of great value.It is possible to determine the mobility characteristics of groups of people as well as individuals through the analysis of human trajectory data and such information can be used to improve our quality of life.However,the effect of trajectory data mining depends on the quality of the original trajectory data,such as the density and size of information.Trajectory data drawn from a single source is limited by its acquisition method,and usually con-tains various flaws.Multiple sources can be combined to provide better data support for trajectory data mining tasks and promote the effect of such tasks.However,the het-erogeneous trajectories of a user are usually collected by different hardware or software equipments,leading to the identity isolation between them.Therefore,researchers have proposed the trajectory linking problem,with the aim of finding multiple trajectories of users from heterogeneous trajectory datasets through data mining.After conducting an in-depth and thorough investigation,this paper divides the problems of linking trajectories into two broad categories according to whether there exists known trajectories:multi-source trajectory linking problem and unsupervised multi-source trajectory linking problem.And further divided it into three levels.Re-search work had been carried out gradually based on these three levels.Trajectory linking across two trajectory sources.This is a problem concerning the special case of trajectory linking,in which each user has only two heterogeneous trajectories.Therefore,the goal of the dual-source trajectory linking is to link the het-erogeneous trajectory pairs of each person in different trajectory datasets.Through this study,we aim to understand the correlation between heterogeneous trajectories of each individual.In this paper,we propose a new method called one-to-one constraint trajec-tory linking with multi-dimensional information(OCTL)that links the corresponding trajectories of one person in different sets in a one-to-one manner.Firstly,OCTL ex-tracts relevant multi-dimensional features from different trajectory datasets,including spatial,temporal and spatio-temporal information,to predict corresponding relation-ships between trajectories.Using these features and the labeled data,OCTL calculates the corresponding probabilities between each pair of trajectories.This is followed by the formulation of link inference to conduct bipartite graph matching in order to link one trajectory to another.The advantages of OCTL are empirically verified on real-world trajectory sets with convincing results.Trajectory linking across multiple trajectory sources.This is a problem con-cerning the generalized trajectory linking.In this problem,each user has an indefinite number of heterogeneous trajectories.Some trajectories are known to belong to the user and are known as linked trajectories,while the rest are unlinked trajectories.To solve this problem,this paper introduces a new concept known as the movement pattern,to describe the inherent drive of trajectory formation,such as the user’s occupation,hob-bies,location preferences etc.Based on such a concept,we design a dual-objective neural network model to integrate the sequential dependency and temporal regularity of trajectories in order to learn the movement patterns of users and their trajectories simultaneously.Thereafter,the corresponding relationship between an unlinked trajec-tory and a user is herein determined by the similarity of their movement patterns.The experimental results on real-world trajectory datasets demonstrate the effectiveness of this solution.Unsupervised trajectory linking problem.This is a problem concerning the spe-cial scenarios for trajectory linking.In this problem,each user has an indefinite number of heterogeneous trajectories,but all of them are unlinked trajectories.The goal of this problem is to have all unlinked trajectories divided into multiple clusters.All tra-jectories within each cluster belong to the same user.Therefore,this problem can be formulated as a special trajectory-clustering problem.Inspired by the movement pat-terns,we propose a dual-output neural-network based clustering model to resolve the problem of unsupervised trajectory linking.The model consists of two parts:The first part is an auto-encoder model,which learns the mapping that embeds trajectories into the latent feature space using the trajectories themselves;The second part is a clustering neural network,which combines the trained encoder from the auto-encoder model with an iterative clustering process to iteratively optimize the trajectory clustering results.The training process of our model is a self-supervised learning one,which can solve the problem of unsupervised trajectory linking without using labeled data.This feature has been proven in the final experiment.