Frequency-Conditioned Diffusion Models for Time Series Generation

Seungwoo Jeong, Junghyo Sohn, Jaehyun Jeon, Yeajin Shon, Heung-Il Suk

Keywords: Diffusion models, Time series generation, Power spectrum, Fourier transformation

Abstract: Time series data, commonly used in fields like climate studies, finance, and healthcare, usually faces challenges such as missing data and privacy concerns. Recently, diffusion models have emerged as effective tools for generating high-quality data, but applying them to time series is still difficult, especially for capturing long-range dependencies and complex information. In this paper, we introduce a new diffusion model that uses frequency domain information to improve time series data generation. In particular, we apply Fourier analysis to adaptively separate low-frequency global trends from high-frequency details, which helps the model better understand important patterns during the denoising process. Finally, our approach uses a specialized frequency encoder to integrate this information, enhancing the model's ability to capture both global and local features. Through exhaustive experiments on various public datasets, our model shows an impressive performance in generating time series data for diverse tasks like forecasting and imputation, outperforming existing methods in accuracy and flexibility.

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: 8963