Spatiotemporal Backward Inconsistency Learning Gives STGNNs Icing on the Cake
Keywords: spatiotemporal learning; time series learning; graph neraul network
Abstract: Spatiotemporal prediction models facilitate various smart-city applications across various domains,such as traffic and climate. While current advancements in these models emphasize leveraging cutting-edge technologies to enhance spatiotemporal learning, they often operate under the implicit assumption of spatiotemporal feature consistency between inputs and labels, overlooking the critical issue of input-label inconsistency. In this study, we introduce a universal spatiotemporal backward inconsistency learning module capable of seamless integration into a variety of models, offering a notable performance boost by explicitly modeling label features to address input-label inconsistency. Our approach includes the development of a spatiotemporal residual theory, advocating for a holistic spatiotemporal learning that encompasses both forward spatiotemporal learning to capture input data’s spatiotemporal features for generating base predictions, akin to existing STNNs, and a backward process to learn residuals that rectify input-label inconsistency, thereby refining the base predictions. Based on this theory, we design the Spatio-Temporal Backward Inconsistency Learning Module (STBIM) for this backward correction process, comprising a residual learning module for decoupling inconsistency information from input representations and label representations, and a residual propagation module for smoothing residual terms to facilitate stable learning. The generated prediction correction term is used to enhance the prediction accuracy. Experimental results on 11 datasets from the traffic and atmospheric domains, combined with 15 spatiotemporal prediction models, demonstrate the broad positive impact of the proposed STBIM. The code is available at https://anonymous.4open.science/r/ICLR2025-2598.
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: 2598
Loading