Harnessing Spatial Dependency for Domain Generalization in Multivariate Time-series Sensor Data
Keywords: Domain Generalization. Multivariate Time-series. sensor and healthcare.
Abstract: Multivariate time-series (MTS) data from multiple sensors often vary across domains due to factors like sensor misalignment, reattachment, or individual differences, posing significant challenges for domain generalization (DG). Existing methods inadequately address the alignment of domain-specific spatial dependencies across different domains in MTS data, as they often assume a unified invariant spatial structure and overlook the distributional discrepancies arising from varying sensor relationships. To address this limitation, we propose ASAM (Adaptive Spatial Dependency Alignment in MTS Data for Domain Generalization), a novel framework that adaptively aligns spatial dependencies across domains. ASAM proposes a DG layer with domain generalization loss function and two-view regularization loss functions to align spatial dependencies between domains adaptively. We adopt a two-phase approach to align different sets of domains effectively. An input-aware graph generation process and a GNN-based DG layer, coupled with the domain generalization loss function, adaptively align the spatial dependencies learned in the second phase with those from the first phase, ensuring a more precise alignment. We additionally incorporate a two-view reg- ularization method to effectively capture underlying spatiotemporal information comprised of spatial decorrelation loss and Gaussian kernel loss. Our theoretical analysis demonstrates that ASAM effectively assimilates information bottleneck, ensuring robustness across diverse distributions. Extensive evaluations of the four real-world datasets show ASAM outperforms ten recent baselines. To the best of our knowledge, this work is among the first to explore DG approaches for MTS data by focusing on spatial dependency alignment. Our code is available at https://anonymous.4open.science/r/ASAM.
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: 5482
Loading