DeepTimeGeo: Trajectory Reconstruction From Sparse Data With Transformer

Shangqing Cao, Jiaman Wu, Aparimit Kasliwal, Baoqi Chen, Giuseppe Perona, Marta C. González

Abstract: The completion of sparse Location-Based Service (LBS) data for modeling urban-scale origin-destination (OD) flow is of great importance to transportation planning applications. Sparse trajectories lack realistic human mobility patterns. Only with completed trajectories one can derive urban-scale OD flow that resembles complete travel diaries as those gathered by surveys or actively collecting phone applications. We present DeepTimeGeo (DTG), a transformer encoder-only model that reconstructs complete trajectories from sparse LBS inputs. We adopt a rank-based representation of locations to preserve individual-level heterogeneity and take a sequence-to-sequence approach to address the issue of gradient back-propagation blockage when it comes to regulating human mobility patterns. We devise human mobility distribution-based loss functions and leverage auxiliary learning to model the dynamics of exploration versus returns in users’ spatial choices. Experimental results show the superiority of DTG in trajectory reconstruction compared to other start-of-the-art generative models for human mobility trajectories. We conducted a case study with LBS data in the city of Coral Gables, Florida. The case study reveals that DTG leads to a reduction of more than 15% (1.35 vs. 1.60), when compared to the state-of-art model, in the cross-entropy loss that measures the deviation from the ground truth departure time distribution. We further demonstrate through SUMO simulation that DTG-generated trip demand captures both morning and evening rush hours, enabled by the more accurate distribution of trip departure time with important implications for traffic estimates.

External IDs:doi:10.1109/tits.2026.3657275