DROSIA: Decoupled Representation on Sequential Information Aggregation for Time Series Forecasting

Kaitao Zhang, Yueyi Xu

Keywords: decoupled representation, sequence modeling, time series forecasting, representation learning

TL;DR: We propose DROSIA to integrate temporal relationships as additional representation for each time point, which achieves the sequential information aggregation in a decoupled manner.

Abstract: Time series forecasting is crucial in various fields, including finance, energy consumption, weather, transportation, and network traffic. It necessitates effective and efficient sequence modeling to encapsulate intricate temporal relationships. However, conventional methods often aggregate sequential information into representations of each time point by considering other points in the sequence, thereby ignoring the intra-individual information and suffering from inefficiency. To address these challenges, we introduce a novel approach, DROSIA: Decoupled Representation On Sequential Information Aggregation, which only integrates temporal relationships once as an additional representation for each point, achieving sequential information aggregation in a decoupled fashion. Thus balancing between individual and sequential information, along with a reduction in computational complexity. We select several widely used time series forecasting datasets, and previously top-performing models and baselines, for a comprehensive comparison. The experimental results validate the effectiveness and efficiency of DROSIA, which achieves state-of-the-art performance with only linear complexity. When provided with a fair length of input data, the channel-independent DROSIA even outperforms the current best channel-dependent model, highlighting its proficiency in sequence modeling and capturing long-distance dependencies. Our code will be made open-source in the subsequent version of this paper.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

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: 5368