Decoupling Variable and Temporal Dependencies: A Novel Approach for Multivariate Time Series Forecasting
Keywords: Time Series Forecasting, transformer, multivariate time series forecasting
Abstract: In multivariate time series forecasting using the Transformer architecture, capturing temporal dependencies and modeling inter-variable relationships are crucial for improving performance. However, overemphasizing temporal dependencies can destabilize the model, increasing its sensitivity to noise, overfitting, and weakening its ability to capture inter-variable relationships. We propose a new approach called the Temporal-Variable Decoupling Network (TVDN) to address this challenge. This method decouples the modeling of variable dependencies from temporal dependencies and further separates temporal dependencies into historical and predictive sequence dependencies, allowing for a more effective capture of both. Specifically, the simultaneous learning of time-related and variable-related patterns can lead to harmful interference between the two. TVDN first extracts variable dependencies from historical data through a permutation-invariant model and then captures temporal dependencies using a permutation-equivariant model. By decoupling variable and temporal dependencies and historical and predictive sequence dependencies, this approach minimizes interference and allows for complementary extraction of both. Our method provides a concise and innovative approach to enhancing the utilization of temporal features. Experiments on multiple real-world datasets demonstrate that TVDN achieves state-of-the-art performance.
Primary Area: applications to computer vision, audio, language, and other modalities
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: 8571
Loading