MGTST: Multi-scale and Cross-channel Gated Transformer for Multivariate long-term time-series forecasting
Primary Area: representation learning for 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.
Keywords: Transformer, multivariate time-series forecasting, deep learning
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: We propose MGTST, a Transformer-based model that explicitly utilizes cross-channel dependency and multi-scale dependency for multivariate time-series forecasting.
Abstract: Transformer-based models have emerged as popular choices for the multivariate long-term time-series forecasting problem due to their ability to capture long-term dependencies. However, current transformer-based models either overlook crucial mutual dependencies among channels or fail to capture various temporal patterns across different scales. To fill the gap, we propose a novel model called MGTST (Multi-scale and cross-channel Gated Time-Series Transformer). In this model, we introduce three innovative designs, including Parallel Multi-Scale Architecture (PMSA), Temporal Embedding with Representation Tokens (TERT), and Cross-Channel Attention and Gated Mechanism (CCAGM). In addition, we introduce Channel Grouping (CG) to mitigate channel interaction redundancy for datasets with a large number of channels. The experimental results demonstrate that our model outperforms both channel-dependent (CD) and channel-independent (CI) baseline models on seven widely used benchmark datasets, with performance improvement ranging from 1.5 percent to 41.9 percent compared to the state-of-the-art in terms of forecasting accuracy.
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: 3124
Loading