MobiDiff discrete diffusion: synthetic human mobility generator

MobiDiff is an end-to-end discrete diffusion framework for synthetic human mobility data, submitted to arXiv on 9 Jul 2026. The model directly denoises multi-channel semantic skeletons that decompose each check-in into spatial, activity, and temporal channels, avoiding interpolation or latent-trace pipelines.

How does MobiDiff work?

MobiDiff constructs mobility entries as multi-channel semantic skeletons and applies discrete diffusion to denoise them, with explicit spatial, activity, and temporal channels. The paper says the method decomposes each human check-in event into spatial, activity, and temporal channels, then uses structured event-, group-, and channel-level masking so the model captures trajectory-level patterns and within-event dependencies.

The pipeline skips costly steps common in prior diffusion approaches: it does not require interpolation, latent trace construction, or coarse-to-fine realization. Instead it operates directly on the discrete semantic representation and learns to remove noise across channels and masked structures, producing synthetic check-in sequences as final output.

How does it perform and compare?

MobiDiff preserves trajectory length and temporal interval distributions while remaining competitive on broader mobility statistics, and it is substantially faster at inference than a prior state-of-the-art. The authors evaluate generation fidelity, privacy-preserving behavior, and efficiency on three large-scale real-world datasets from Atlanta, Boston, and Seattle. Results show MobiDiff effectively preserves trajectory length and temporal interval distributions and is "5.3× faster than GeoGen on average during inference."

Beyond the single speed metric, the paper positions discrete diffusion as an interpretable framework that natively models discrete semantic events with explicit region, activity, time, and interval structures. The structured masking at event, group, and channel levels is the mechanism the paper highlights for jointly capturing trajectory-level mobility patterns and within-event dependencies.

Why it matters

Synthetic mobility data is expensive to collect and hard to share because of privacy concerns. MobiDiff offers a way to generate realistic, discrete semantic mobility traces without reconstructing continuous latent trajectories, which reduces engineering complexity. The reported 5.3× average inference speed advantage over GeoGen implies lower compute cost for producing large synthetic datasets, which matters for teams that need many realizations for planning or privacy testing.

By modeling explicit activity and temporal channels and using structured masking, MobiDiff also makes the generation process more interpretable than approaches that rely on opaque continuous latent traces. That interpretability can help practitioners check whether synthetic datasets preserve specific statistical properties like trajectory length or temporal interval distributions, two properties the paper says the method preserves.

What to watch

Check for code and data releases linked to the arXiv submission and for independent benchmarks that reproduce the paper's evaluation on the Atlanta, Boston, and Seattle datasets. Also watch for follow-up comparisons against GeoGen and other diffusion-based mobility models that report metrics beyond inference speed, such as privacy leakage tests and downstream task performance.