Recurrent Prompt Learning for Spatio-Temporal Forecasting
Date
Authors
Chen, C
Liu, Y
Niu, C
Shi, K
Chen, L
Zhu, T
Supervisor
Item type
Journal Article
Degree name
Journal Title
Journal ISSN
Volume Title
Publisher
Institute of Electrical and Electronics Engineers (IEEE)
Abstract
Spatio-temporal forecasting holds great significance for various applications in the intelligent transportation system. Foundation models are revolutionizing spatio-temporal forecasting models due to their one-fits-all generalization capabilities. To reprogram the foundation models for the targeted downstream tasks, prompt learning has emerged as an effective approach by optimizing a small set of learnable input tokens while keeping the backbone intact. However, current prompt learning in the spatio-temporal domain typically suffers from two critical limitations: (1) time-agnostic, which cannot capture the temporal evolution during the prompt learning to deal with temporally dependent or sensitive scenarios. (2) input-agnostic, which remain static upon the end of the training, thus failing to fit different distributions at inference. To address these challenges, we propose a recurrent prompt learning framework named RePro to repurpose foundation models to downstream ST forecasting tasks. For the first challenge, we design a recurrent prompt network that is dynamically conditioned on the time-evolving prompts and recurrently optimized based on the historical context. This design injects time-awareness into the prompt to achieve progressive recalibration of the intermediate representations in the foundation model under varying temporal contexts. For the second challenge, we incorporate input data of the current time step into the update of each recurrent prompt state, leading to the input-conditioned prompt learning. This design effectively encapsulates distributional shifts into the prompt dynamics, improving generalization and robustness. Furthermore, two complementary modules are introduced to facilitate the effective application of the recurrent prompt to the foundation model, i.e., cross-prompt aggregation and layer-conditioned prompt adaptation. Specifically, the first module aims to unify the prompt representation and reduce the redundancy, while the second module distributes the recurrent prompts into diverse layers of the foundation model for hierarchical prompting. Extensive experiments on multiple spatio-temporal forecasting benchmarks demonstrate that RePro consistently outperforms strong state-of-the-art baselines across MAE, RMSE, and MAPE, achieving up to 8.3% reduction in MAE, with ablation studies validating the contribution of each proposed component.Description
Keywords
40 Engineering, 4008 Electrical Engineering, 4009 Electronics, Sensors and Digital Hardware, Machine Learning and Artificial Intelligence, 0906 Electrical and Electronic Engineering, 1005 Communications Technologies, Networking & Telecommunications, 4008 Electrical engineering, 4009 Electronics, sensors and digital hardware
Source
IEEE Transactions on Consumer Electronics, ISSN: 0098-3063 (Print); 1558-4127 (Online), Institute of Electrical and Electronics Engineers (IEEE), PP(99), 1-1. doi: 10.1109/TCE.2026.3675217
Publisher's version
Rights statement
This is the Author's Accepted Manuscript of an article published in IEEE Transactions on Consumer Electronics. The Version of Record will be available at DOI: 10.1109/TCE.2026.3675217