Deep Generative Model for Human Mobility Behavior

Understanding and modeling human mobility is central to challenges in transport planning, sustainable urban design, and public health. Despite decades of effort, simulating individual mobility remains challenging because of its complex, context-dependent, and exploratory nature. Here, we advance a unified event-level formulation of daily mobility and propose MobilityGen to generate multi-attribute event sequences over days to weeks at large spatial scales. By linking behavioral attributes with environmental context, MobilityGen reproduces key patterns such as scaling laws for location visits, activity time allocation, and the coupled evolution of travel mode and destination choices. It reflects spatio-temporal variability and generates diverse and plausible mobility patterns consistent with the built environment. Beyond standard validation, MobilityGen enables analyses that have been difficult with earlier models, including how access to urban space varies across travel modes and how co-presence dynamics shape social exposure and segregation. Together, these results support an integrated, data-driven basis for fine-grained studies of human mobility behavior and its societal implications.

💡 Research Summary

The paper introduces MobilityGen, a novel deep generative framework for simulating individual human mobility at an event‑level granularity. Rather than treating location choice, time allocation, and travel‑mode selection as separate components, the authors model daily mobility as a chronologically ordered sequence of multi‑attribute events, each defined by a tuple ⟨location ID, start time, duration, travel mode⟩ together with contextual information about the built environment (geographic coordinates, nearby points of interest, land‑use descriptors).

To learn the complex joint distribution of these sequences, MobilityGen employs a conditional denoising diffusion probabilistic model (DDPM) adapted for discrete data. Raw categorical attributes (location, mode) are mapped to continuous vectors via learnable embedding matrices, while continuous attributes (time, duration) are projected through small feed‑forward networks. The resulting event embeddings are fed into a transformer encoder that extracts global context from the observed travel trace and its environmental features; this context guides the reverse diffusion process carried out by a transformer decoder. During training, Gaussian noise is gradually added to the embeddings (forward diffusion) and the decoder learns to reconstruct the original embeddings (reverse diffusion). A rounding step converts the denoised continuous embeddings back into discrete tokens, and linear heads predict the final attribute values. The loss combines a mean‑squared reconstruction term for the embeddings with a cross‑entropy term for the discrete tokens, plus consistency penalties that enforce coherent relationships among attributes.

The authors evaluate the model on a large‑scale GNSS‑based smartphone travel survey covering the whole of Switzerland. Locations are discretized onto a hierarchical grid with a finest resolution of roughly 500 × 500 m, and each cell is enriched with POI and land‑use data. MobilityGen is trained on over one million event sequences and compared against classic mechanistic models (EPR, Container), utility‑based simulators (Equasim, OASIS), and recent deep sequence models. Results show that MobilityGen accurately reproduces known scaling laws (Pareto‑type visitation frequency, distance distributions), captures realistic daily time‑use patterns, and preserves the coupled evolution of travel mode and destination choice. Importantly, the model generates plausible multi‑day to multi‑week trajectories, maintaining both short‑term variability (e.g., rush‑hour peaks) and longer‑term shifts (e.g., weekend versus weekday behavior).

Beyond standard validation, the paper demonstrates three downstream applications that were previously difficult with existing simulators: (1) mode‑specific accessibility analysis, revealing spatial inequities in public‑transport reach; (2) co‑presence network construction, enabling quantitative assessment of social exposure and segregation; and (3) scenario testing for urban policy (e.g., adding bike lanes) by conditioning generated sequences on altered environmental features.

The main contributions are: (i) a unified event‑level formulation that jointly models location, time, duration, and mode; (ii) the adaptation of diffusion models to discrete, multi‑attribute mobility sequences via embedding‑diffusion‑rounding; (iii) incorporation of rich built‑environment context into the generative process; and (iv) empirical evidence that the approach reproduces both macro‑level statistical regularities and micro‑level behavioral dependencies.

Limitations include the reliance on high‑quality GNSS and POI data, which may not be available in many regions, and the computational cost of diffusion‑based sampling, which hampers real‑time applications. The latent embeddings are also largely opaque, making policy interpretation challenging. The authors suggest future work on cross‑regional generalization, faster diffusion variants (e.g., DDIM), and integrating graph neural networks to embed explicit spatial network structure for better interpretability.

In summary, MobilityGen represents a significant step forward in data‑driven human mobility simulation, offering a flexible, high‑fidelity tool for transport planning, urban design, public‑health modeling, and social‑science research.