Diffusion Models for Time Series Applications: A Survey

Lequan Lin,Zhengkun Li,Ruikun Li,Xuliang Li,Junbin Gao

2023-05-01

Abstract:Diffusion models, a family of generative models based on deep learning, have become increasingly prominent in cutting-edge machine learning research. With a distinguished performance in generating samples that resemble the observed data, diffusion models are widely used in image, video, and text synthesis nowadays. In recent years, the concept of diffusion has been extended to time series applications, and many powerful models have been developed. Considering the deficiency of a methodical summary and discourse on these models, we provide this survey as an elementary resource for new researchers in this area and also an inspiration to motivate future research. For better understanding, we include an introduction about the basics of diffusion models. Except for this, we primarily focus on diffusion-based methods for time series forecasting, imputation, and generation, and present them respectively in three individual sections. We also compare different methods for the same application and highlight their connections if applicable. Lastly, we conclude the common limitation of diffusion-based methods and highlight potential future research directions.

Machine Learning

What problem does this paper attempt to address?

### Problems the Paper Attempts to Solve

This paper primarily explores the latest advancements in the application of diffusion models in time series and provides a systematic review. Specifically, the paper aims to:

1. **Fill Methodological Gaps**: Due to the current lack of systematic summaries and discussions on these models, this paper provides a foundational resource to help new researchers entering this field and to inspire future research.

2. **Introduce Basic Knowledge of Diffusion Models**: The paper first introduces the basic concepts and working principles of diffusion models to help readers better understand the subsequent content.

3. **Focus on Time Series Tasks**: It focuses on time series prediction, interpolation, and generation methods based on diffusion models, and provides detailed explanations under these three themes.

4. **Compare Different Methods**: It conducts a comparative analysis of different methods for the same application scenario and points out their connections (if applicable).

5. **Summarize Limitations and Future Directions**: Finally, it points out the common limitations of existing diffusion model methods and proposes potential future research directions.

Through the above points, the paper hopes to provide researchers in this field with a comprehensive and self-contained literature review, thereby promoting further development in this emerging field.