Simple Yet Effective Spatio-Temporal Prompt Learning

Qianru Zhang, Lianghao Xia, Zhonghang Li, Siu Ming Yiu, Chao Huang

Keywords: Spatio-Temporal Prompt Learning, Traffci prediction, Crime Prediction

Abstract: Accurate spatio-temporal prediction is pivotal for optimizing transportation systems and enhancing urban management. However, the practical application of cutting-edge graph neural network (GNN)-based methods for these tasks encounters challenges, particularly regarding their ability to generalize. GNN-based approaches have shown promise in capturing intricate spatial and temporal dependencies found in traffic and crime data. They utilize graph structures to model relationships between locations or entities, enabling the prediction of traffic patterns and crime incidents. Nonetheless, a key challenge involves ensuring that these models can effectively generalize to unseen scenarios and adapt to varying spatio-temporal data distributions. To tackle this challenge, we present a lightweight and effective prompt learning paradigm called as PromptST. This framework serves as an adaptation of pretrained spatio-temporal prediction models, specifically designed to handle the dynamics of spatial and temporal distributions. In the context of spatio-temporal prediction, our prompt tuning incorporates a simple prompt network into the pretrained model. By automatically learning informative prompt contexts that encapsulate the underlying spatial and temporal patterns from unseen data, the spatio-temporal prompt network guides the pretrained model to successfully adapt and learn from new data distributions. Our proposed prompt learning framework has been extensively evaluated on various spatio-temporal datasets, and the results demonstrate its effectiveness. Across multiple spatio-temporal prediction tasks, our PromptST achieves state-of-the-art prediction accuracy while maintaining computational efficiency, showcasing its superiority in capturing complex dependencies and adapting to varying data distributions across time and space.

Primary Area: general machine learning (i.e., none of the above)

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/2024/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: 550