Transferable Human Mobility Network Reconstruction with neuroGravity

Jinming Yang, Shaoyu Huang, Zongyuan Huang, Yaohui Jin, Xiaokang Yang, Marta C. Gonzalez, Yanyan XuarXiv 2026 · arXiv

TL;DR

neuroGravity is a physics-informed deep learning framework that reconstructs city-wide human mobility networks using only publicly available population and urban facility data. By blending the classical gravity law with graph neural networks, it successfully predicts absolute travel flows with minimal local data and transfers effectively to entirely unobserved cities. Furthermore, the model uncovers that a city's spatial income segregation is the primary factor dictating how well mobility patterns transfer across different urban environments.

What Problem Does This Paper Address?

Understanding human mobility is vital for urban planning, traffic management, and epidemic control. Traditionally, cities rely on manual travel surveys or aggregated mobile phone data (like Call Detail Records) to map these flows. However, these methods are prohibitively expensive and logistically challenging, leading to severe data scarcity in undeveloped regions and the Global South.

While pure physical models (like the gravity model) generalize well, they lack accuracy because they cannot capture complex, context-specific urban structures. Conversely, standard deep learning models excel at capturing complex patterns but overfit easily and fail to generalize to new cities. Crucially, existing deep learning baselines (such as Deep Gravity) require knowing a city's total outflows in advance to generate absolute volumes, making them useless for true "zero-shot" transfer to unobserved cities. neuroGravity bridges this gap.

Methods

The authors introduce neuroGravity, a hybrid architecture that operates in a logarithmic space to preserve the multiplicative nature of physical gravity laws while making the data more amenable to neural networks.

The pipeline consists of four main stages:

  1. Feature Extraction: Compiles 52 built-environment features from OpenStreetMap (OSM) and population distributions.
  2. Connection Predictor: Uses a LightGBM binary classifier to filter out implausible origin-destination (OD) pairs, establishing a sparse base network topology.
  3. Meta-Gravity Component: A deep-parameterized physics model that uses Multi-Layer Perceptrons (MLPs) to dynamically calculate the gravitational constant ($G$) and distance decay exponent ($\alpha$) based on regional features, outputting a physically grounded base flow.
  4. Edge-Enhanced Graph Transformer: Refines the base flow by capturing complex spatial dependencies and encoding them into node and edge embeddings.

Additionally, the authors introduce a novel Spatial Income Segregation Index ($SI$) based on Bregman Information decomposition. This index quantifies the degree of income clustering within a city to mathematically predict how well a model trained in City A will perform when transferred to City B.

Datasets

The model was trained, validated, and evaluated using a diverse set of open and proprietary datasets across multiple global cities:

  • Built-Environment Data: 52 distinct features spanning land-use, buildings, points-of-interest (POIs), and road networks derived from OpenStreetMap (OSM).
  • Population Data: Sourced from the U.S. Census Bureau (at the ZIP Code Tabulation Area level) and WorldPop for international regions.
  • Mobility Ground Truth: Call Detail Records (CDR) and Location-Based Services (LBS) data for six primary cities: Boston, Los Angeles, San Francisco Bay Area, Porto, Bogotá, and Riyadh.
  • Socioeconomic Indicators: U.S. Census 2021 5-year ACS (income, education), CoolClimate (carbon footprint), and environmental data ($\text{NO}_2$ concentrations).

Key Insights and Ideas

  • The Power of Log-Space Transformation: Transforming mobility flows into a logarithmic space linearizes the power-law distribution of human travel. This allows standard MLPs with ReLU activations to easily learn patterns across vastly different flow scales.
  • Segregation Governs Transferability: The authors discovered that cross-city model transferability is governed by spatial income segregation. Models transfer remarkably well between cities with similar segregation profiles ($SI$ values). Furthermore, models trained in cities with lower segregation (more uniform income distribution, like Boston) produce more generalizable mobility representations.
  • Emergent Socioeconomic Semantics: Even though neuroGravity is never trained on socioeconomic data, its learned node embeddings implicitly cluster by income and education. Using a simple gradient boosting machine, these embeddings serve as highly accurate proxies for local wealth, carbon footprints, and average travel ranges ($R_g$).

Key Results

  • Few-Shot Reconstruction: In Boston, neuroGravity successfully reconstructed the mobility network ($R^2 = 0.77$) even when only 1% of internal origin-destination pairs were observed. Under a 10% observation ratio, it achieved an $R^2$ of 0.77 and a Common Part of Commuters (CPC) of 0.73, significantly outperforming the classical gravity model ($R^2 = 0.59$, CPC = 0.63).
  • Zero-Shot Cross-City Transfer: A model trained purely on Boston data achieved impressive zero-shot performance when tested on unseen cities: $R^2 = 0.69$ in Los Angeles and $R^2 = 0.61$ in the SF Bay Area. On average, neuroGravity nearly doubled (+99%) the $R^2$ performance of the best-performing deep learning baselines.
  • Global Scalability: The authors successfully scaled neuroGravity to generate mobility flow proxies for over 1,200 cities worldwide, validating its efficacy against real-world travel surveys in data-scarce sub-Saharan African regions.

Limitations and Future Work

  • OSM Data Dependencies: The model relies heavily on OpenStreetMap data. While neuroGravity remains stable up to a 30% feature missing ratio, extreme data scarcity or severe spatial biases in OSM coverage (common in remote parts of the Global South) can still degrade estimation accuracy.
  • Multimodal Fusion: Integrating satellite imagery (via Google AlphaEarth embeddings) currently yields only marginal performance gains. The authors note that more sophisticated multimodal fusion strategies are required to better capture visual urban semantics.
  • Urban Structure Heterogeneity: The regional embeddings are highly effective at predicting socioeconomic status in monocentric cities (like Boston) but show diminishing returns in polycentric regions (like LA and the SF Bay Area) where travel patterns are naturally more dispersed and mixed.

Summary updated

Abstract

Accurate modeling of human mobility is critical for tackling urban planning and public health challenges. In undeveloped regions, the absence of comprehensive travel surveys necessitates reconstructing mobility networks from publicly available data. Here we develop neuroGravity, a physics-informed deep learning model that reliably reconstructs mobility flows from limited observations and transfers to unobserved cities. Using only urban facility and population distributions, we find that neuroGravity's regional representations strongly correlate with socioeconomic and livability status, offering scalable proxies for costly surveys. Furthermore, we uncover that spatial income segregation plays a key role in model transferability: mobility networks are most reliably reconstructed when target cities share similar segregation levels with the source. We design an index to quantify this segregation and accurately predict transferability. Finally, we generate mobility flow proxies for over 1,200 cities worldwide, highlighting neuroGravity's potential to mitigate critical data shortages in resource-limited, underdeveloped areas.
BibTeX
@misc{yang2026transferable,
  author = {Jinming Yang and Shaoyu Huang and Zongyuan Huang and Yaohui Jin and Xiaokang Yang and Marta C. Gonzalez and Yanyan Xu},
  title = {{Transferable Human Mobility Network Reconstruction with neuroGravity}},
  year = {2026},
  eprint = {2604.23678},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2604.23678},
}

Loading related content...

Figure