Abstract
Vehicle-based travel time estimation is crucial for many travel scheduling and city planning applications in intelligent transportation systems. Since trajectories in different trips are affected by evolutionary Spatio-temporal dynamics (e.g., evolving travel patterns for different days of the week and varying road networks affected by traffic accidents or temporary restrictions, etc.), it is substantial to investigate these dynamics for accurate estimation. In this paper, we propose a novel deep learning model which fuses location features, distance features, and temporal features with meta learning-based neural networks, to implicitly learn path representations for evolving travel patterns in different days of week and road networks. Specifically, we utilize the meta learning-based optimization method to transfer the shared meta knowledge across trajectories in distinct Evolving-Tasks (i.e., a limited amount of trajectory data on different days of the week), which facilitates generalizing rapidly on evolving travel patterns for different days of the week. In addition, road network information is obviated in our model, which makes it a natural solution to tolerate evolving road networks while mitigating the computation burden contemporaneously. Comprehensive experiments on three real-world datasets demonstrate the superiority of our proposed model.