Deep Generative Model for Human Mobility Behavior

TL;DR

MobilityGen introduces a diffusion-based, context-aware generative model to simulate long-horizon human mobility with multifaceted behavioral attributes. By embedding activity attributes and environmental context within a transformer-based encoder–decoder and guiding generation from observed sequences, it captures both location-level patterns and richer spatio-temporal interactions, including mode-specific space usage and social co-presence. The approach outperforms classical and neural baselines across location metrics, mobility motifs, and unseen-location flows, and reveals interpretable embeddings that connect travel modes, timing, and geography. This framework enables large-scale, fine-grained mobility simulations with implications for urban design, transport policy, and public health, while supporting analyses of social exposure and segregation.

Abstract

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 present MobilityGen, a deep generative model that produces realistic mobility trajectories spanning 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, plausible, and novel mobility patterns consistent with the built environment. Beyond standard validation, MobilityGen yields insights not attainable with earlier models, including how access to urban space varies across travel modes and how co-presence dynamics shape social exposure and segregation. Our work establishes a new framework for mobility simulation, paving the way for fine-grained, data-driven studies of human behavior and its societal implications.

Figures (15)

• Figure 1: Modeling individual mobility with a deep generative framework. a. We conceptualize individual mobility as a sequence of chronologically ordered activity events, where each event has associated activity attributes (location, time use, and travel mode) and contextual information (coordinates and nearby POIs). b. Raw activity events are encoded into latent mobility embeddings through dedicated embedding modules. c. During training, the diffusion process adds Gaussian noise to the target sequence embedding $\mathbf{z}_0$. A transformer-based decoder learns to reverse this process and reconstruct a denoised embedding $\hat{\mathbf{z}}_0$. Features from the traveled sequence are extracted by an encoder and provided as guidance to the decoder. Final activity attributes are predicted via linear output heads. The model is trained by minimizing reconstruction and consistency losses between the generated and original sequences (see Methods). d. During inference, we sample a Gaussian noise vector $\mathbf{z}_T \sim \mathcal{N}(0, \mathbf{I})$ and iteratively refine it using the decoder, guided by encoder-extracted features. The denoised embedding $\mathbf{z}_0$ is then decoded into the sequence of activity events. Travel mode icons adapted from Flaticon.com.
• Figure 1: Comparison of spatial visitation patterns across models. Aggregated visit frequencies from real data (a) are compared with those generated by three baseline mobility models: EPR (b), Container (c), and MobilityGen (d). While all models recover broad national-scale trends, MobilityGen more accurately reproduces both high-density urban visitation and spatial dispersion into surrounding regions.
• Figure 2: Evaluating microscopic mobility models with location metrics. a. Rank-frequency distribution of visited locations. b. Median radius of gyration as a function of the number of displacements. c. Distribution of temporal mobility entropy across individuals. All metrics are compared between real data (black), MobilityGen (red), EPR (blue), and Container (green). MobilityGen aligns most closely with empirical distributions across all metrics.
• Figure 2: Urban-level comparison of spatial visitation patterns across models. Real and simulated visit frequencies are compared across Bern (top row), Zurich (middle row), and Lucerne (bottom row). Consistent with national-level patterns, MobilityGen best captures fine-grained spatial visitation within dense urban cores and across peripheral areas.
• Figure 3: MobilityGen generates realistic spatial patterns of location visits across individual, urban, and national levels. We compare the spatial distribution of activity locations in real data (a, c, e, g, and i) with those generated by MobilityGen (b, d, f, h, and j). a and b show visits and transitions for a sample individual, where edge width is proportional to transition frequency. The generated sequence reflects routinely visited locations while also capturing exploratory behavior. c and d display aggregated visit frequencies across all individuals in Switzerland, with warmer colors indicating higher visitation intensity on a log scale. Panels e–j provide zoomed-in views for Bern (e, f), Zurich (g, h), and Lucerne (i, j). The generated sequences closely align with observed spatial patterns while exhibiting slightly greater diversity in the locations visited.