FEDNET: FREQUENCY ENHANCED DECOMPOSED NETWORK FOR OUT-OF-DISTRIBUTION TIME SERIES CLASSIFICATION

Kehan Yin, Hong Huang, Ruize Shi, Hua Zhu, Wei Zhou, Hai Jin

Keywords: out-of-distribution, time series classification, frequency

Abstract: Time series classification is a crucial task with widespread applications in various fields such as medicine and energy. Due to the non-stationary property of time series, its data distribution will change over time, which makes it challenging for models to generalize to the out-of-distribution (OOD) environment. However, limitations persist in the current research on OOD time series classification, particularly the absence of a unified consideration addressing both domain distribution shift and temporal distribution shift. To this end, we view the time series distribution shift from the frequency perspective and propose a novel method called Frequency Enhanced Decomposed Network (FEDNet) for OOD time series classification. FEDNet utilizes frequency domain information to guide the decomposition of time series and further eliminates domain shift and temporal shift, it then obtains domain-invariant features for adapting to OOD data. Finally,we provide theoretical insights of FEDNet to validate its superiority for OOD time series classification. Comprehensive results on synthetic and real-world datasets demonstrate that FEDNet achieves state-of-the-art performance in OOD time series classification tasks, surpassing previous methods by up to 7%.Our code is available at https://anonymous.4open.science/r/FEDNet-743E

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.

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: 9399