Abstract
Generative models have shown promising results in capturing human mobilitycharacteristics and generating synthetic trajectories. However, it remainschallenging to ensure that the generated geospatial mobility data issemantically realistic, including consistent location sequences, and reflectsreal-world characteristics, such as constraining on geospatial limits. Wereformat human mobility modeling as an autoregressive generation task toaddress these issues, leveraging the Generative Pre-trained Transformer (GPT)architecture. To ensure its controllable generation to alleviate the abovechallenges, we propose a geospatially-aware generative model, MobilityGPT. Wepropose a gravity-based sampling method to train a transformer for semanticsequence similarity. Then, we constrained the training process via a roadconnectivity matrix that provides the connectivity of sequences in trajectorygeneration, thereby keeping generated trajectories in geospatial limits.Lastly, we proposed to construct a preference dataset for fine-tuningMobilityGPT via Reinforcement Learning from Trajectory Feedback (RLTF)mechanism, which minimizes the travel distance between training and thesynthetically generated trajectories. Experiments on real-world datasetsdemonstrate MobilityGPT's superior performance over state-of-the-art methods ingenerating high-quality mobility trajectories that are closest to real data interms of origin-destination similarity, trip length, travel radius, link, andgravity distributions.