ScholarWorks@성균관대학교

초록

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 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 two-phase framework that adaptively aligns spatial dependencies across domains. Specifically, an input-aware graph generation module and graph neural network (GNN)-based DG layer, together with the domain generalization loss, align the spatial dependencies learned in the second phase with those established in the first phase, thereby enabling more precise and consistent cross-domain alignment. We additionally incorporate a two-view regularization strategy for sensor independence and temporal consistency, respectively. Our theoretical analysis demonstrates that ASAM contributes to robustness across diverse distributions. Extensive evaluations of the four real-world datasets show ASAM outperforms thirteen recent baselines. To the best of our knowledge, this work is among the first to focus on extensive spatial dependency alignment, opening new directions for DG in MTS data. Our code is available at https://github.com/learndatalab/ASAM.

키워드