Mixture-of-Diffusers: Dual-Stage Diffusion Model for Improved Time Series Generation
Keywords: Time Series Generation; Time Series Analysis; Diffusion Models; Mixture-of-Experts
Abstract: Synthetic Time Series Generation (TSG) is a crucial task for data augmentation and various downstream applications. While TSG has advanced, its effectiveness often relies on the availability of extensive training datasets, posing challenges in data-scarce scenarios. Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs) have shown promise, but they frequently struggle to capture the complex temporal dynamics and interdependencies inherent in time series data. To address these limitations, we propose a novel generative framework, Mixture-of-Diffusers (MoD). This approach decomposes the diffusion process into a collection of specialized diffusers, each designed to model specific patterns at distinct noise levels. Early-stage diffusers focus on capturing overarching global and coarse patterns, while late-stage diffusers specialize in capturing fine-grained details as the noise level diminishes. This hierarchical decomposition empowers MoD to learn robust representations and generate realistic time series samples. The model is trained using a combination of multi-objective loss functions, ensuring both temporal consistency and alignment with the true data distribution. Extensive experiments on a diverse range of real-world and simulated time series datasets demonstrate the superior performance of MoD compared to state-of-the-art TSG generative models. Furthermore, rigorous evaluations incorporating both qualitative and quantitative metrics, coupled with assessments of downstream task performance on long-term generation and scarce time series data (see Figure 1), collectively validate the efficacy of our proposed approach.
Supplementary Material: zip
Primary Area: learning on time series and dynamical systems
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.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
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: 14026
Loading