Deep Temporal Deaggregation: Large-Scale Spatio-Temporal Generative Models
Keywords: generative learning, mobility data, denoising diffusion probablistic models
Abstract: Access to spatio-temporal trajectory data is essential for improving infrastructure, preventing the spread of disease and for building autonomous vehicles. However, it remains underutilized due to limited availability, as it cannot be shared publicly due privacy concerns or other sensitive attributes. Generative time-series models have shown promise in generating non-sensitive data, but show poor performance for large-scale and complex environments. In this paper we propose a spatio-temporal generative model for trajectories, TDDPM, which outperforms and scales substantially better than state-of-the-art. The focus is primarily on trajectories of peoples' movement in cities. We propose a conditional distribution approach which unlock out-of-distribution generalization, such as to city-areas not trained on, from a spatial aggregate prior. We also show that data can be generated in a privacy-preserving manner using $k$-anonymity. Further, we propose a new comprehensive benchmark across several standard datasets, and evaluation measures, considering key distribution properties.
Supplementary Material: zip
Primary Area: generative models
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 7080
Loading