HyFAD Model Boosts Time Series Imputation with Hybrid Diffusion

A new paper from researchers at multiple institutions introduces HyFAD, a hybrid time-frequency diffusion model designed to tackle the challenging problem of time series imputation—filling in missing values in sequential data. Existing diffusion-based methods struggle with frequency-sensitive denoising, often losing high-frequency details or failing to balance global trends with local fluctuations. HyFAD addresses this by extending the standard DDPM (Denoising Diffusion Probabilistic Model) paradigm with a coupled time-frequency framework. The reverse denoising process proceeds sequentially: first in the time domain to capture low-frequency global trends, then in the frequency domain to refine high-frequency spectral components. This coarse-to-fine generation ensures that both broad patterns and fine-grained details are preserved. Additionally, the model introduces a frequency-aware step embedding that leverages the relationship between diffusion steps and spectral components, providing step-dependent guidance for more accurate band-wise reconstruction.

Extensive experiments on multiple benchmark datasets demonstrate that HyFAD achieves state-of-the-art performance in time series imputation, outperforming previous methods that operate solely in the time or frequency domain. The source code is publicly available on GitHub, enabling reproducibility and further research. For professionals working with sensor data, financial time series, or IoT monitoring, HyFAD represents a significant advancement in handling missing data. By combining time-domain trend learning with frequency-domain detail refinement, the model offers a practical tool for improving the quality of real-world datasets, leading to more reliable downstream analysis and decision-making in domains like finance, healthcare, and industrial monitoring.