CausalDiffusion: Causally Related Time-Series Generation through Diffusion Models
Keywords: time-series generation, causal discovery, diffusion model, benchmark, dataset, synthetic data
TL;DR: A novel pipeline capable of generating realistic time-series along with a ground truth causal graph that is generalizable to different fields.
Abstract: Understanding the intrinsic causal structure of time-series data is crucial for effective real-world interventions and decision-making.
While several studies address the Time-Series Causal Discovery (TSCD) problem, the lack of high-quality datasets may limit the progress and evaluation of new methodologies.
Many available datasets are derived from simplistic simulations, while real-world datasets are often limited in quantity, variety, and lack of ground-truth knowledge describing temporal causal relations.
In this paper, we propose CausalDiffusion, the first diffusion model capable of generating multiple causally related time-series alongside a ground-truth causal graph, which abstracts their mutual temporal dependencies.
CausalDiffusiom employs a causal reconstruction of the output time-series, allowing it to be trained exclusively on time-series data.
Our experiments demonstrate that CausalDiffusion outperforms state-of-the-art methods in generating realistic time-series, with causal graphs that closely resemble those of real-world phenomena.
Finally, we provide a benchmark of widely used TSCD algorithms, highlighting the benefits of our synthetic data with respect to existing solutions.
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: 10672
Loading