[Paper Review] MobilityGPT: Enhanced Human Mobility Modeling with a GPT model

Ammar Haydari, Dongjie Chen+3 more|arXiv (Cornell University)|Feb 5, 2024

Human Mobility and Location-Based Analysis4 citations

TL;DR

MobilityGPT proposes a GPT-based generative model for human mobility that enhances trajectory realism by integrating gravity-aware sampling, road connectivity matrix (RCM) masking, and reinforcement learning fine-tuning (RLTF). It outperforms state-of-the-art methods by 9% on average in trajectory similarity metrics, achieving superior fidelity in origin-destination patterns, trip length, and spatial distribution.

ABSTRACT

Generative models have shown promising results in capturing human mobility characteristics and generating synthetic trajectories. However, it remains challenging to ensure that the generated geospatial mobility data is semantically realistic, including consistent location sequences, and reflects real-world characteristics, such as constraining on geospatial limits. We reformat human mobility modeling as an autoregressive generation task to address these issues, leveraging the Generative Pre-trained Transformer (GPT) architecture. To ensure its controllable generation to alleviate the above challenges, we propose a geospatially-aware generative model, MobilityGPT. We propose a gravity-based sampling method to train a transformer for semantic sequence similarity. Then, we constrained the training process via a road connectivity matrix that provides the connectivity of sequences in trajectory generation, thereby keeping generated trajectories in geospatial limits. Lastly, we proposed to construct a preference dataset for fine-tuning MobilityGPT via Reinforcement Learning from Trajectory Feedback (RLTF) mechanism, which minimizes the travel distance between training and the synthetically generated trajectories. Experiments on real-world datasets demonstrate MobilityGPT's superior performance over state-of-the-art methods in generating high-quality mobility trajectories that are closest to real data in terms of origin-destination similarity, trip length, travel radius, link, and gravity distributions. We release the source code and reference links to datasets at https://github.com/ammarhydr/MobilityGPT.

Motivation & Objective

• Address the lack of semantic realism and geospatial consistency in synthetic human mobility trajectories generated by existing models.
• Overcome limitations of GANs, VAEs, and LSTMs in preserving spatial-temporal dependencies and trajectory continuity.
• Develop a controllable, GPT-based framework that maintains realistic location sequences and adheres to road network constraints.
• Enable high-quality synthetic trajectory generation without human feedback, using automated fine-tuning via trajectory feedback.
• Improve similarity between synthetic and real trajectories across key mobility metrics such as OD pairs, trip length, and gravity distribution.

Proposed method

• Reformulate human mobility modeling as an autoregressive sequence generation task using a GPT backbone with self-attention mechanisms.
• Introduce gravity-aware sampling during pretraining to align generated trajectories with real-world origin-destination interaction patterns.
• Apply road connectivity matrix (RCM) masking to the model’s logits, ensuring only physically connected road segments are predicted, thus preserving trajectory continuity.
• Develop a Reinforcement Learning from Trajectory Feedback (RLTF) fine-tuning method that minimizes travel distance differences between real and synthetic trajectories.
• Use Jensen-Shannon Divergence (JSD) to quantitatively evaluate similarity across multiple mobility metrics during training and inference.
• Leverage temperature scheduling during inference to balance diversity and quality, showing improved performance at higher temperatures.

Experimental results

Research questions

• RQ1 Can a GPT-based model generate human mobility trajectories that are semantically realistic and spatially consistent?
• RQ2 How does incorporating gravity modeling and road connectivity constraints improve the realism of synthetic trajectories?
• RQ3 To what extent does RL-based fine-tuning without human feedback enhance trajectory similarity compared to supervised or preference-based methods?
• RQ4 How do individual components (gravity, RCM, RLTF) contribute to the overall performance of the MobilityGPT model?
• RQ5 Does temperature tuning in inference improve the diversity and quality of generated mobility sequences?

Key findings

• MobilityGPT outperforms all baseline models by an average of 9% in trajectory similarity across metrics, with the highest improvement in origin-destination (OD) and trip length similarity.
• The full MobilityGPT model achieves a JSD of 0.114 for OD similarity and 0.124 for trip length, significantly outperforming ablated variants.
• The ablation study shows that removing any of the three core components—gravity model, RCM, or RLTF—results in a measurable performance drop, with RCM and gravity contributing most to OD similarity.
• RLTF fine-tuning outperforms both supervised fine-tuning (SF) and direct preference optimization (DPO), achieving the lowest JSD (0.114 for OD, 0.124 for trip length).
• Higher temperature settings during inference improve performance, indicating that controlled randomness enhances the diversity and realism of generated trajectories.
• Spatial density maps of generated trajectories in Porto show that MobilityGPT closely replicates real-world spatial patterns, outperforming other models in visual and statistical similarity.