STOP! A Out-of-Distribution Processor with Robust Spatiotemporal Interaction
Keywords: Spatiotemporal learning; out-of-distribution learning; spatiotemporal prediction
Abstract: Recently, spatiotemporal graph convolutional networks have attained significant success in spatiotemporal prediction tasks. However, they encounter out-of-distribution (OOD) challenges due to the sensitivity of node-to-node messaging mechanism to spatiotemporal shifts, leading to suboptimal generalization in unknown environments. To tackle these issues, we introduce the **S**patio-**T**emporal **O**OD **P**rocessor (STOP), which leverages spatiotemporal MLP channel mixing as its backbone, separately incorporating temporal and spatial elements for prediction. To bolster resilience against spatiotemporal shifts, STOP integrates robust interaction including a centralized messaging mechanism and a graph perturbation mechanism. Specifically, centralized messaging mechanism configures Context Aware Units (ConAU) to capture generalizable context features, constraining nodes to interact solely with ConAU for spatiotemporal feature interaction. The graph perturbation mechanism uses Generalized Perturbation Units (GenPU) to disrupt this interaction process, generating diverse training environments that compel the model to extract invariant context features from these settings. Finally, we customized a spatiotemporal distributionally robust optimization (DRO) to enhance generalization by exposing the model to challenging environments. Through evaluations on six datasets, STOP showcases competitive generalization and inductive learning. The code is available at https://anonymous.4open.science/r/ICLR2025-STOP.
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: 116
Loading