MobiDiff: Semantic-Aware Multi-Channel Discrete Dif... | AI Research

MobiDiff: Semantic-Aware Multi-Channel Discrete Diffusion for Human Mobility Data Generation
Human mobility data—records of where people go and what activities they perform—are vital for urban planning, transportation optimization, and disaster response. However, these datasets are difficult to share due to privacy concerns and high collection costs. While researchers have used generative models to create synthetic versions of this data, existing methods often rely on complex, multi-stage pipelines that struggle to capture the discrete, structured nature of human movement. MobiDiff introduces a more efficient, end-to-end approach that treats mobility data as a sequence of semantic events, allowing for faster and more accurate synthetic data generation.

A New Way to Model Movement

Traditional diffusion models often work with continuous data, requiring researchers to convert mobility traces into complex latent representations or perform costly interpolations. MobiDiff takes a different path by using "discrete diffusion." It views a human trajectory as a sequence of "semantic skeletons," where each check-in event is broken down into five distinct channels: macro-region, micro-region, activity category, absolute time, and the time interval between events. By modeling these as discrete tokens rather than continuous coordinates, the framework can generate realistic activity patterns in a single, streamlined process.

Structured Learning and Denoising

To ensure the generated data is realistic, MobiDiff uses a sophisticated masking strategy. During training, the model hides parts of the mobility skeleton—such as a specific time interval or a location—and learns to reconstruct the missing information based on the surrounding context. This happens at three levels: the individual event, groups of related channels (like spatial or temporal factors), and single channels. By forcing the model to solve these "fill-in-the-blank" tasks, it learns the complex dependencies between where a person is, what they are doing, and when they are doing it. The model also incorporates "numeric-aware" features, ensuring that even though the data is processed as discrete tokens, the underlying geographic and temporal logic remains intact.

Performance and Efficiency

The researchers tested MobiDiff on large-scale mobility datasets from Atlanta, Boston, and Seattle. The results demonstrate that the framework is highly effective at preserving key mobility statistics, such as trajectory length and the time intervals between events. Beyond its ability to generate high-fidelity data, MobiDiff is significantly faster than existing state-of-the-art methods; for example, it is 5.3 times faster than the GeoGen model during the inference phase.

Considerations for Future Use

While MobiDiff shows great promise in creating efficient and interpretable synthetic data, the authors note that there is still room for improvement. Specifically, while the model excels at capturing temporal and trajectory-level patterns, its spatial fidelity—the precision of the generated locations—can vary depending on the city and the specific metrics being measured. Nevertheless, the study suggests that discrete diffusion is a highly effective and scalable framework for researchers looking to generate synthetic mobility data while protecting individual privacy.