Time-Frequency Conditioned Diffusion for Multivariate Time Series Imputation

Time series data underpin critical applications in domains including healthcare and meteorology, yet their utility is frequently compromised by pervasive missing values. While diffusion-based generative models offer promise for imputation, existing approaches are constrained by their limited ability to effectively capturing dynamic shifts or variations in inherent periodic patterns present within the observed portions of the multivariate time series data. To address this gap, we propose a novel time-frequency conditioned diffusion called CDTI for multivariate time series imputation, integrating time-frequency representations as structured spectral priors to explicitly guiding the denoising process. CDTI transforms observed data into a spectrogram, extracts deep features of periodic dynamics via a dedicated time-frequency feature learning (TFFL) module, and injects these features via a cross-view interaction (CVI) mechanism at each denoising step. This synergistic fusion of temporal and spectral domains enables CDTI to perceive complex periodic shifts and abrupt changes, constraining the reverse process for high-fidelity reconstruction. Extensive evaluations on five benchmark datasets show that CDTI achieves consistently competitive performance across a wide range of missing data scenarios and missing rates, demonstrating robust imputation capability compared with existing state-of-the-art methods. The source code is available at URL(https://github.com/Cofeesy/CDTI).