ShuffleMTM: Learning Cross-channel Dependence in Multivariate Time Series from Shuffled Patches
Keywords: Self-supervised learning, masked modeling, multivariate time series, cross-channel dependence
TL;DR: We propose ShuffleMTM, a simple yet effective masked time-series modeling framework learning cross-channel dependence from shuffled patches.
Abstract: Masked time-series modeling has widely gained attention as a self-supervised pre-training method for multivariate time series (MTS). Recent studies adopt a channel-independent (CI) strategy to enhance the temporal modeling capacity. Despite the effectiveness and performance of this strategy, the CI methods inherently overlook cross-channel dependence, which is inherent and crucial in MTS data in various domains. To fill this gap, we propose ShuffleMTM, a simple yet effective masked time-series modeling framework to learn cross-channel dependence from shuffled patches. Technically, ShuffleMTM proposes to shuffle the unmasked patches from masked series across different channels, positioned at the same index. Then, Siamese encoders learn two views of masked patch representations from original and shuffled masked series, simultaneously capturing the temporal dependence within a channel as well as spatial dependence across different channels. ShuffleMTM pre-trains the Siamese encoders to reconstruct the original series by incorporating cross-channel information with intra-channel cross-time information. Our proposed method consistently achieves superior performance in various experiments, compared to advanced CI pre-training methods and channel-dependent methods in both time series forecasting and classification tasks.
Supplementary Material: zip
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.
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: 5726
Loading