GCMT: a graph-contextualized multitask spatio-temporal joint prediction model for cellular trajectories

Abstract

Spatio-temporal joint prediction aims to simultaneously predict the next location and the corresponding switch time for a cellular trajectory. An accuracy prediction requires not only sequential information but also spatio-temporal context information. Although existing methods can utilize trajectory modeling to support the joint prediction, they fail to learn the complicated geographical influence, temporal dependencies and various context information. To this end, we propose a graph-contextualized multitask learning method for spatio-temporal joint prediction. Specially, to model each location’s spatio-temporal dependencies, a graph embedding module is adopted to jointly capture the geographical influence and temporal cyclic effect by embedding three relational graphs (i.e., location-location, location-region, and location-time) into a shared low dimensional space. Moreover, considering the impact of traffic-related contexts on trajectory movement, we design a traffic encoder to model the dynamic of traffic flows, which comprises several spatio-temporal blocks combining temporal gated CNN with spatial graph convolution. In addition, a context-attention layer is proposed to fuse trajectory sequential information and traffic information based on various background factors. Finally, GCMT is evaluated on two real-world datasets to demonstrate its advantages.