When Self-attention and Topological Structure Make a Difference: Trajectory Modeling in Road Networks.

The ubiquitous GPS-enabled devices (e.g., vehicles and mobile phones) have led to the unexpected growth in trajectory data that can be well utilized for intelligent city management, such as traffic monitoring and diversion. As a building block of the smart-mobility initiative, trajectory modeling has received increasing attention recently. Despite the great contributions made by existing studies, they still suffer from the following problems. (1) The topological structure of a road network is underutilized. (2) The existing methods cannot characterize the stopping probability of a trajectory. To this end, we develop a novel model entitled TMRN (Trajectory Modeling in Road Networks), which is composed of the following three modules. (1) Road2Vec: the module is developed to learn the representations of road segments by fully utilizing the topology information of a road network. (2) LWA: the lightweight attention-based module is designed to capture the long-term regularity of trajectories. (3) MOP: a novel matching operation is proposed to calculate the transition probability of the next segment for the current path. The extensive experiments conducted on two real-world datasets demonstrate the superiority of TMRN compared with state-of-the-art methods.