Channel Independence Improves Out-of-Distribution Generalisation in Multivariate Time Series Classification

Tom Ryder, Xi Chen

Keywords: time series classification, OOD generalization, domain generalization

TL;DR: Learning with channel independence significantly improves robustness to distribution shift in multivariate time series classification.

Abstract: Robustness to distribution shift is a necessary property of machine learning models for their safe and effective deployment. However, deep learning models are susceptible to learning spurious features of the in-distribution (ID) training data that fail to generalise to out-of-distribution (OOD) data. Domain generalisation algorithms aim to tackle this problem, but recent studies have demonstrated that their improvement over standard empirical risk minimisation is marginal. We address this problem for multivariate time series classification (TSC), where it is standard practise to use feature extractor architectures that learn with channel dependence (CD), enabling cross-channel patterns to be learned. Inspired by recent success in time series forecasting, we investigate how channel independence (CI) impacts OOD generalisation in TSC. Our experiments on six time series datasets reveal that ID and OOD features exhibit significantly greater distributional divergence when learned with CD compared to CI. As a consequence, models that learn with CI are more robust to distribution shift, evidenced by smaller generalisation gaps (the difference between ID and OOD performance) across datasets. On datasets that have a stronger shift, OOD accuracy is substantially higher for CI than CD.

Primary Area: learning on time series and dynamical systems

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: 9897