Mixed Channel Dependency Diffusion Model with Retrieval Guidance for Time Series Forecasting

Yu-Hao Huang, Wu-Jun Li

Keywords: Time series forecasting, Diffusion Model, Retrieval augmented generation

Abstract:

Recent advancements in deep learning techniques have improved the performance of time series forecasting, especially with state-of-the-art generative models.
Despite making progress in capturing conditional time-series patterns with uncertainty, existing time series generative models face reliability and computational challenges in long-term forecasting, especially when the number of variate is large.
Moreover, the maximum likelihood objective of generative modeling is prone to underestimation for low-density region of the data manifold, therefore leading to sub-optimal conditional sampling quality.
In this paper, we propose a mixed channel dependency diffusion model with retrieval guidance (MiDDiR) to address these challenges.
In MiDDiR, we employ a novel mixed channel dependency method on time series diffusion model, encoding historical time series in a channel-dependent manner to obtain informative historical representation while denoising in a channel-independent manner to decrease modeling complexity.
During inference, we retrieve similar history occurrence for explicitly tilting the score estimation as retrieval guidance to enhance forecasting quality.
Extensive experiments demonstrate the effectiveness of \rgdiff, outperforming baselines in a variety of real-world time series datasets.

Primary Area: learning on time series and dynamical systems

Submission Number: 23811