M-STAR: Multi-Scale Spatiotemporal Autoregression for Human Mobility Modeling

arXiv ID: 2512.07314v1

Published: 2025-12-08

Authors: Yuxiao Luo, Songming Zhang, Sijie Ruan, Siran Chen, Kang Liu, Yang Xu, Yu Zheng, Ling Yin

Categories: cs.AI, cs.LG

Relevance Score: 0.90 / 1.00

View on arXiv Download PDF

Summary

M-STAR (Multi-Scale Spatio-Temporal AutoRegression) is a novel framework designed to generate long-term human mobility trajectories, overcoming the limitations of previous AI-generated content (AIGC) methods regarding inefficiency and lack of explicit multi-scale modeling. It achieves this through a coarse-to-fine spatiotemporal prediction process, leveraging a Multi-scale Spatiotemporal Tokenizer and a Transformer-based decoder. Experiments on real-world datasets demonstrate M-STAR's superior fidelity and significantly improved generation speed compared to existing methods.

Medical Relevance

Accurate and efficient modeling of human mobility is fundamentally important for public health, serving as a critical input for epidemic modeling to predict disease spread, understand transmission dynamics, and inform public health interventions during outbreaks.

AI Health Application

M-STAR, an AI model for multi-scale spatiotemporal autoregression, can be applied in health by providing enhanced capabilities for epidemic modeling. By generating more accurate and long-term human mobility trajectories, it can significantly improve the prediction of infectious disease spread, inform public health strategies for containment and resource allocation, and strengthen biosecurity measures against biological threats. It's an AI tool that feeds into health-related simulation and predictive models.

Key Points

Addresses critical limitations in current AIGC methods for synthetic trajectory generation, specifically their inefficiency in long-term generation (e.g., weekly) and absence of explicit spatiotemporal multi-scale modeling.
Introduces M-STAR, a new framework capable of generating comprehensive long-term human mobility trajectories.
Employs a unique coarse-to-fine spatiotemporal prediction process to effectively model mobility across different temporal and spatial granularities.
Integrates a Multi-scale Spatiotemporal Tokenizer, which is designed to encode hierarchical mobility patterns inherent in human movement data.
Utilizes a Transformer-based decoder for next-scale autoregressive prediction, allowing for sequential and detailed trajectory generation.
Demonstrated superior fidelity in generated trajectories compared to existing state-of-the-art methods.
Achieved a significant improvement in the speed of long-term trajectory generation, making it more practical for real-world applications.

Methodology

M-STAR implements a coarse-to-fine spatiotemporal prediction process. This involves a Multi-scale Spatiotemporal Tokenizer, which hierarchically encodes mobility patterns at various spatial and temporal scales into tokens. These tokens are then fed into a Transformer-based decoder that performs autoregressive prediction, generating the subsequent, finer-grained mobility patterns and ultimately constructing detailed long-term trajectories.

Key Findings

The study found that M-STAR significantly outperforms existing methods in two key areas: it achieves higher fidelity in the generation of synthetic human mobility trajectories, indicating a more realistic representation of movement patterns. Furthermore, M-STAR substantially improves the speed of long-term trajectory generation, addressing a critical bottleneck in prior approaches, as validated on two real-world datasets.

Clinical Impact

The ability of M-STAR to accurately and rapidly generate long-term human mobility trajectories has profound clinical and public health implications. It can enhance the precision and timeliness of epidemic models, enabling public health officials to more effectively predict the trajectory of infectious disease outbreaks, optimize resource allocation (e.g., vaccine distribution, hospital beds), plan targeted interventions (e.g., social distancing measures), and assess the impact of mobility restrictions, ultimately contributing to more effective disease control and mitigation strategies.

Limitations

The abstract primarily highlights the limitations of existing AIGC methods (inefficiency in long-term generation and lack of explicit spatiotemporal multi-scale modeling) that M-STAR aims to overcome. It does not explicitly state any inherent limitations or caveats pertaining to the M-STAR framework itself.

Future Directions

The abstract does not explicitly mention specific future research directions for the M-STAR framework.

Medical Domains

Public Health Epidemiology Infectious Disease Modeling Disease Surveillance Emergency Preparedness

Keywords

Human mobility Spatiotemporal modeling Autoregression Multi-scale Trajectory generation Transformer AIGC Epidemic modeling

Abstract

Modeling human mobility is vital for extensive applications such as transportation planning and epidemic modeling. With the rise of the Artificial Intelligence Generated Content (AIGC) paradigm, recent works explore synthetic trajectory generation using autoregressive and diffusion models. While these methods show promise for generating single-day trajectories, they remain limited by inefficiencies in long-term generation (e.g., weekly trajectories) and a lack of explicit spatiotemporal multi-scale modeling. This study proposes Multi-Scale Spatio-Temporal AutoRegression (M-STAR), a new framework that generates long-term trajectories through a coarse-to-fine spatiotemporal prediction process. M-STAR combines a Multi-scale Spatiotemporal Tokenizer that encodes hierarchical mobility patterns with a Transformer-based decoder for next-scale autoregressive prediction. Experiments on two real-world datasets show that M-STAR outperforms existing methods in fidelity and significantly improves generation speed. The data and codes are available at https://github.com/YuxiaoLuo0013/M-STAR.