Research On Spatio-temporal Prediction Method For Cellular Trajectory

With the rapid development of 5G communication technology,smartphone users have a growing demand for unimpeded and ubiquitous high-quality communication services.To provide users with better experience and communication services,it has become a hot research topic to mine mobility patterns and user preferences in spatio-temporal cellular trajectories and make accurate trajectory predictions,so as to realize timely and effective base station scheduling.However,most existing studies on spatio-temporal prediction for cellular trajectories ignore the effective information implied in the trajectory data,and are difficult to deal with the few-shot problem of personalized prediction.In order to solve the above problems,this paper focuses on the spatio-temporal prediction for cellular trajectory.The main contents can be summarized as follows:(1)To make use of the effective information implied in cellular trajectory data,this paper proposes a spatio-temporal multi-task learning based cellular trajectory prediction model.The model adopts a multi-task learning framework,which takes trajectory location prediction and arrival time prediction as main tasks,and travel intention prediction as an auxiliary task to mine deep knowledge and provide long-term intention information.The model designs a distribution-aware loss function to accurately capture the intention information with complex spatial associations between different regions,and adopts a gating mechanism to realize the effective fusion of sequence information and intention information.In addition,considering that trajectory movement is affected by contextual information,the model combines a bipartite graph embedding module and traffic encoder to capture geographic influence,time cycle effect and traffic-related context information respectively.(2)To solve the few-shot problem of personalized trajectory prediction,this paper proposes an adaptive meta-learning model,which can realize rapid personalization with a small amount of task-related data on the basis of a generalized initial network.Considering that the same network initialization will ignore users’ different preferences,the model further constructs a memory mechanism on the basis of meta-learning.It combines task clustering and cluster-aware parameter memory to provide personalized network initialization parameters for each task,so as to ensure knowledge sharing among similar tasks.Moreover,in order to make effective use of task-related data for fine-tuning,the model adopts a trajectory sampling method to generate a support set combining both user and trajectory related knowledge,and designs a weight generator to adaptively assign a reasonable weight to each trajectory.This paper conducts extensive comparative experiments on real-world datasets,and the proposed models are compared with state-of-the-art methods.The experimental results demonstrate that the models proposed in this paper are helpful to improve the accuracy of spatio-temporal prediction for cellular trajectories in both general scenarios and few-shot scenarios of personalized prediction.