Abstract Sensor and machine learning technologies have improved the perception of traffic systems by providing detailed data about individual vehicle trajectories. Combining data from different types of sensors shows promise for comprehensive perception of global traffic, but it remains challenging. Stationary roadside units only gather sparse trajectories of passing vehicles, while crowd-sourced data records entire trajectories but only consists of a very low sample rate of vehicles. Therefore, there is a need to learn route choice behavior from crowd-sourced data to infer complete paths for the sparse trajectories. Existing route choice models assume path set enumeration or the Markovian property for simplicity, which leaves room for capturing the long sequence of choice behavior from data for added precision. Additionally, the path inference problem is often broken down into multiple independent route choice problems between any consecutive sparse observations, leaving room for exploring one-shot long-sequence inference. To address these challenges, we propose RoutesFormer, an efficient sequence-based, data-driven route choice Transformer that requires minimal assumptions due to the capacity of the model architecture. By being sequence-based, RoutesFormer unifies the route choice and path inference problems, accommodating all observations together and avoiding the need to break down the problem into separate route choices, thereby improving optimality. Experiments conducted on the Shanghai taxi dataset demonstrate that RoutesFormer has made significant improvements over six existing baseline models in various challenging path inference tasks. Specifically, RoutesFormer has achieved state-of-the-art accuracy with an average total link length accuracy of 0.914/0.870 compared to the baselines’ best average accuracy of 0.896/0.845, and it ranks first across all tasks. Additionally, the attention mechanism used in RoutesFormer is interpreted, providing a lens to study traveler’s route choice behavior in the real world.

Highlights • Innovates a sequence-based path inference model to remedy long-standing drawbacks. • Unifies path inference/route choice in an end-to-end paradigm. • Tailors two types of attention mechanisms to express complex route choice behavior. • Achieves SOTA accuracies, compared with six baselines from classic to deep models.