High-fidelity human mobility data—such as individual GPS trajectories, cellular signaling logs, and transit smart card records—is vital for modern urban planning, epidemic forecasting, and smart city design. However, as noted by Xiangjie Kong et al. in their survey on mobility trajectory generation, collecting real-world trajectories faces severe legal, ethical, and financial hurdles. Location history inherently reveals sensitive personal routines, making raw datasets highly restricted under modern privacy regulations.
Furthermore, much of this data is locked within corporate silos. Yuanshao Zhu et al. point out in SynMob that high-quality mobility logs are typically the proprietary property of ride-hailing or telecommunications companies, creating a "data desert" for independent researchers. This lack of open-access data directly limits the validation of epidemiological models (Brian Mac Namee et al. in An Open Data Driven Epidemiological Agent-Based Model for Irish Towns) and stalls urban representation learning in data-scarce regions (Fengze Sun et al. in FlexiReg).
To bypass these barriers, a powerful research direction has emerged: overcoming mobility data scarcity by synthesizing realistic trajectories and modeling movement through non-sensitive open proxies.
The Core Intuition: Invariants of Human Movement
If we cannot safely share or collect raw coordinates, how can we accurately simulate or infer urban movement? The core insight of this field is that human mobility is not random noise. Instead, it is governed by a combination of physical infrastructure (road networks, building layouts), temporal rhythms (work hours, weekends), and semantic intent (shopping, commuting, returning home).
Because these underlying drivers are highly structured, models do not need to memorize exact real-world trajectories. Instead, they can capture the statistical invariants of human behavior. By modeling how physical space and human intent interact, researchers can either generate entirely synthetic trajectories that mimic real-world distributions or reverse-engineer movement patterns using non-sensitive, publicly available data streams.
Generative Trajectory Simulation
The first major paradigm involves training generative machine learning models to dream up realistic, privacy-preserving trajectories from scratch. Early methods relied on rigid, rule-based physics models, but modern approaches leverage deep generative architectures and large language models (LLMs) to capture complex behavioral nuances.
For instance, the framework proposed by Yuanshao Zhu et al. in SynMob utilizes probabilistic diffusion models to generate high-fidelity synthetic GPS trajectories. These synthetic paths preserve macro-level traffic demand and micro-level travel times, allowing urban planners to test traffic management systems without exposing any individual's true location history.
To make these trajectories truly realistic, however, models must also capture human decision-making. Chen Ju et al. introduced TrajLLM, a framework that uses LLMs as autonomous agents to simulate the "intent" behind movement. By combining an LLM's semantic reasoning (e.g., "I need to buy groceries, then pick up my child") with spatial constraints, the framework generates highly interpretable and realistic agent behaviors.
[Semantic Intent (LLM Agent)] ──> [Spatial Constraints (Road Network)] ──> [Realistic Synthetic Trajectory]
A key challenge in generative modeling is zero-shot generalization: can a model simulate traffic in a city where no trajectory data exists? Jingyuan Wang et al. addressed this in GTG by incorporating "Space Syntax" theory. By disentangling city-specific spatial features from invariant structural mobility patterns, their model accurately simulates human movement in entirely new urban environments using only the local road network topology.
Cross-Modal Inference and Open Proxies
The second major paradigm bypasses trajectory generation entirely by extracting mobility patterns from non-sensitive, publicly available data streams. This approach treats public data as a proxy for hidden physical movement.
Consider the challenge of mapping urban commuting patterns (Origin-Destination flows). Traditionally, this requires intrusive surveys or sensitive mobile tracking logs. To circumvent this, Can Rong et al. developed GlODGen, a framework that generates commuting flows using only public satellite imagery and population density data. By training vision-language foundation models to recognize the functional layout of a city from space, the authors achieve accuracy competitive with models built on proprietary socioeconomic data.
Similarly, public infrastructure data can be reverse-engineered to democratize mobility insights. Lakmal Meegahapola demonstrated in Inferring Accurate Bus Trajectories from Noisy Estimated Arrival Time Records that individual transit trajectories can be reconstructed using only noisy, public Estimated Time of Arrival (ETA) feeds. This eliminates the need for highly restricted smart card transaction logs or private GPS feeds from transit authorities.
Even coarse, aggregated historical records can substitute for real-time tracking in critical scenarios. In public health, tracking the spread of infectious diseases across borders normally demands high-resolution mobile phone roaming data. However, Nick Ruktanonchai et al. showed in Census-derived migration data as a tool for informing malaria elimination policy that widely available, static census migration data can serve as an effective proxy for mapping "parasite highways," allowing officials to coordinate disease elimination without invasive surveillance.
By treating human mobility as a predictable function of geography, intent, and public infrastructure, these paradigms effectively decouple urban computing from its dependency on invasive, proprietary tracking.
Key Foundational Work
Towards Geo-Social Intelligence: Mining, Analyzing, and Leveraging Geospatial Footprints in Social Media (2013) — contributed 1 concept elements to this direction.
Extracting urban patterns from location-based social networks (2011) — contributed 1 concept elements to this direction.