Application Research And Analysis Of Real-time Trajectory Stream Evolution

With the popularization of smartphones and the development of mobile Internet,we have witted that the acquisition of trajectory data has become more convenient and accurate,as well as the diverse scenes,such as the tracks of taxis,the routes of shared bicycles,and the check-in behaviors in social networks.These trajectory data not only record the position information and change process of the moving object,but also reflect the implicit patterns and preferences,which is wort researching.However,the traditional trajectory mining algorithm rarely focus on the dynamics brought by the spatio-temporal properties of the trajectory,so the mining results often ignore the evolution process of the trajectory.In addition,the volumn of the trajectory data is growing everyday.The methods based on the full data set can longer process such huge volumn of data,and the processing speed cannot respond to the online request immediately.To this end,we consider data flow technology as a way to solve massive data mining problems,and focus on the trajectory evolution process in streaming scenarios,and propose real-time trajectory clustering and real-time location recommendation algorithm.This thesis first studies the trajectory clustering problem in data stream scenarios and proposes an online clustering algorithm AntClu.Aiming at the temporal and spatial characteristics of the trajectory in the streaming scene,we define the ”trend” to represent the current motion pattern of the moving object,and use the ”time synchronization distance”to quantify the representation error and improve the characterization accuracy.Then we propose a density-based online clustering algorithm based on the trend structure.The evolution process of the trajectory aggregation model is explored by dynamically maintaining the online micro-clustering,and the real-time clustering request is responded to by the macro-clustering result.Finally,we demonstrate the validity of the dynamic representation structure through meticulous argumentation,and prove that the AntClu algorithm has superior performance and clustering effect compared with the traditional algorithm.The thesis studies the real-time location recommendation problem based on data stream,and proposes a geo-distributed matrix decomposition model Geo-SRMF based on pairwise ranking approach.By analyzing the check-in data set,we conclude that the user tends to visit the venues in their neighborhood,and assume that the venues in the neighborhood has a higher rank order than the non-neighbors.Therefore,we construct a matrix decomposition model based on pairwise raning appraoch and use the gradient descent to learn the parameters.At the same time,consider that the online learning model can't remember the long-term data distribution,we introduce an update strategy based on reservoir sampling,and select the sampling candidate set to improve the learning accuracy.Finally,we demonstrate through experiments that the Geo-SRMF algorithm can effectively perform real-time processing and online recommendation of check-in data,and has better accuracy than the traditional recommendation algorithm.The experiments of the thesis demonstrate the evolution of both GPS and check-in trajectory.At last,the thesis analyses the pros and cons of two algorithms and arrange the work of the next stage.