Official Review · ICLR 2025

This paper proposes Spatio-Temporal Diffusion Model (STDM) which redesigns the diffusion forward process for capturing correlations in time series data, and can be seamlessly integrated into current diffusion models to improve their performance in time series analysis tasks. Experiments explore the performence of STDM in time series anomaly detection and forecasting tasks. - Novel research perspective. As far as the reviewer is aware, this is the first paper to improve the performance on time series analysis tasks by redesigning the diffusion forward process. - The proposed method is flexible and extensible, and can be seamlessly integrated into the time series diffusion model to improve performance. - Motivation is not clear and does not tell a coherent story. I do not understand the motivation and significance of paying attention to the temporal patterns in the diffusion forward process of adding noise. It appears that the temporal correlations introduced by the noise in the forward process may enable the model to effectively consider and learn these correlations for denoising in the reverse process. However, the writing of the paper does not clearly explain this. - In Section 3, the author mentioned that "our methodology innovatively manipulates the forward process. This adjustment facilitates faster convergence during training and enhances robustness during inference". Nevertheless, the mechanism by which STDM accelerates training and improves inference robustness are not sufficiently explained, and both theoretical analysis and lacks empirical evidence to support this assertion. - The experiment results only evaluate the DiffusionAE and TimeGrad models, which are not enough to support the effectivenenss of the proposed method. And there is a notable absence of baselines for time series forecasting and anomaly detection, which limits the comprehensiveness of the evaluation. - The writing and charts are extremely crude and rudimentary. - Is convolution kernel $H$ trainable? What is the extra training cost of this design in diffusion forward process? - How does the convolution kernel capture spatio-temporal correlations? In my opinion, kernel $H$ seems to be able to capture only **temporal** pattern correlations within a series, but the author claims that STDM captures **spatial** correlations (for example, the contribution section). The method cannot capture spatio-temporal correlations of multivariate series anyway, I don't understand why the author named it **Spatio-Temporal** diffusion model.