核心概念
ST traffic forecasting models need to address out-of-distribution issues caused by spatio-temporal shifts and external factors, such as time and weather variations.
摘要
The content discusses the challenges of spatio-temporal traffic forecasting, the impact of external factors on traffic data distribution, and the need to handle out-of-distribution scenarios. It introduces the Disentangled Contextual Adjustment (DCA) and the Spatio-Temporal sElf-superVised dEconfounding (STEVE) framework to improve generalization in OOD scenarios. The model incorporates causal inference theory and self-supervised tasks to enhance the robustness of traffic forecasting models.
- Introduction to the importance of ST traffic forecasting.
- Challenges faced in forecasting due to distribution shifts.
- Proposal of DCA and STEVE frameworks.
- Implementation details of the model.
- Experimental results and comparisons with baselines.
統計資料
"Comprehensive experiments on four large-scale benchmark datasets demonstrate that our STEVE consistently outperforms the state-of-the-art baselines across various ST OOD scenarios."
"The proposed causal graph among input X, output Y, and confounder C reveals the importance of addressing spurious correlations in traffic data."
引述
"We propose a theoretical solution named Disentangled Contextual Adjustment (DCA) from a causal lens."
"Our STEVE completely beats its canonical degradation STGCN, which supports the confounding assumption of ST context C."