Multivariate Time-series Forecasting with SPACE: Series Prediction Augmented by Causality Estimation
Keywords: Time Series Forecasting, Causal Learning, Transfer Entropy, Graph Based Learning
Abstract: The analysis of multivariate time series (MTS) presents a complex yet crucial task with substantial applications in areas such as weather forecasting, policy formulation, and stock market prediction. It is important to highlight three key characteristics of MTS that contribute to the challenging and multifaceted nature of their analysis: (i) their interrelationships are represented through causal relationships rather than mere similarities; (ii) they convey information across multiple independent factors; and (iii) their dynamics often arise from inherent temporal dependencies. While conventional time series analysis frameworks often fail to capture one or more of these aspects, resulting in incomplete or even misleading conclusions, we propose an end-to-end trainable $\textbf{S}$eries $\textbf{P}$rediction model $\textbf{A}$ugmented by $\textbf{C}$ausality $\textbf{E}$stimation (SPACE) to address these limitations. This model effectively incorporates temporal dependencies and causal relationships, featuring a temporal embedding and a transfer entropy-based Cross-TE module designed to enhance predictions through causality-augmented mechanisms. Experiments demonstrate that SPACE achieves state-of-the-art results on challenging real-world time series prediction tasks, showing its effectiveness and versatility.
Primary Area: causal reasoning
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: 13805
Loading