Identifiability Guarantees For Time Series Representation via Contrastive Sparsity-inducing

Khalid Oublal, Said Ladjal, David Benhaiem, Emmanuel LE BORGNE, François Roueff

Keywords: Time Series Representations Learning, Generalization, Disentangled Representations Learning, Source Separation

TL;DR: We empirically demonstrate and prove that Contrastive Sparsity-inducing guarantees disentangled representations, which improve compositional generalization and interpretability in shifted distributions.

Abstract: Time series representations learned from high-dimensional data, often referred to as ”disentanglement” are generally expected to be more robust and better at generalizing to new and potentially out-of-distribution (OOD) scenarios. Yet, this is not always the case, as variations in unseen data or prior assumptions may insufficiently constrain the posterior probability distribution, leading to an unstable model and non disentangled representations, which in turn lessens generalization and prediction accuracy. While identifiability and disentangled representations for time series are often said to be beneficial for generalizing downstream tasks, the current empirical and theoretical understanding remains limited. In this work, we provide results on identifiability that guarantee complete disentangled representations via Contrastive Sparsity-inducing Learning, which improves generalization and interpretability. Motivated by this result, we propose the TimeCSL framework to learn a disentangled representation that generalizes and maintains compositionality. We conduct a large-scale study on time series source separation, investigating whether sufficiently disentangled representations enhance the ability to generalize to OOD downstream tasks. Our results show that sufficient identifiability in time series representations leads to improved performance under shifted distributions. Our code is available at https://anonymous.4open.science/r/TimeCSL-4320.

Primary Area: other topics in machine learning (i.e., none of the above)

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: 11200