MobiDiff Generates Realistic Human Mobility Data Efficiently

Human mobility data is crucial for urban planning, transportation optimization, and resource allocation, but its collection is expensive and sharing is restricted by privacy concerns. While existing diffusion-based methods can synthesize realistic mobility patterns, they often rely on continuous or latent spatio-temporal traces, which limits their ability to model discrete semantic events like specific regions, activities, times, and intervals. To overcome this, researchers developed MobiDiff, an end-to-end discrete diffusion framework. MobiDiff directly generates mobility data by denoising multi-channel semantic skeletons, avoiding the complex and costly interpolation or latent trace construction pipelines used by other methods. The framework decomposes each human check-in event into spatial, activity, and temporal channels, using structured masking at event, group, and channel levels to capture both trajectory-level patterns and within-event dependencies. Evaluated on large datasets from Atlanta, Boston, and Seattle, MobiDiff effectively preserves key mobility statistics and is significantly faster than state-of-the-art methods, demonstrating an interpretable and efficient approach for synthetic mobility data generation.