Denoising diffusion models have emerged as a formidable method, consistently surpassing previous state-of-the-art benchmarks. However, a notable challenge in time series-related tasks like anomaly detection and forecasting is the conditioning for models to reconstruct inputs accurately or generate samples based on past time steps rather than producing entirely new samples. To address this, we introduce a novel technique that enhances the sampling capabilities of denoising diffusion models for time series analysis, namely Spatio-Temporal Diffusion Models (STDM). While recent methods fall short of mapping contextual neighborhood dependencies directly into the sampling of a noisy sample, we focus on guiding the forward process of the diffusion model. The degeneration of a sample is based on the idea that values of neighboring time steps are highly correlated. We benefit from this assumption by presenting a diffusion step-dependent convolutional kernel to capture spatial relations and a combined, correlated noise to degenerate the input. Our method can be integrated seamlessly into various existing time series diffusion models. We compare the results of anomaly detection and forecasting when using the traditional and our novel forward process. In our experiments on synthetic and real-world datasets, we show that an adaption of the forward process can be beneficial, as our approach outperforms diffusion models with the ordinary forward process in task-specific metrics, underscoring the efficacy of our strategy in enhancing time series analysis through advanced diffusion techniques.
Keywords: Diffusion Models, Time Series Analysis, Anomaly Detection, Forecasting
TL;DR: A novel approach for manipulating the forward process of time series diffusion models to benefit from temporal correlations
Abstract:
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: 11133
Loading