核心概念
The authors propose a theoretical solution named Disentangled Contextual Adjustment (DCA) to address the out-of-distribution (OOD) issue in traffic data forecasting. They introduce the Spatio-Temporal sElf-superVised dEconfounding (STEVE) framework to improve generalization ability.
要約
The content discusses the challenges of OOD traffic forecasting due to spatio-temporal shifts and proposes a novel approach, STEVE, to address these issues. The model incorporates causal inference theory and self-supervised tasks for improved performance across various scenarios.
The authors highlight the importance of removing spurious correlations caused by external factors in traffic data forecasting. They introduce a theoretical scheme called Disentangled Contextual Adjustment (DCA) and instantiate it as the STEVE framework for better OOD generalization. The proposed model shows superior performance compared to state-of-the-art baselines in comprehensive experiments on real-world datasets.
統計
Extensive experiments on four large-scale benchmark datasets demonstrate that our STEVE consistently outperforms the state-of-the-art baselines across various ST OOD scenarios.
For example, NYCBike1 dataset shows an average MAE of 5.03 and an average MAPE of 24.40 with STEVE.
In comparison, other methods like AGCRN, ST-Norm, AdaRNN, COST, CIGA, STNSCM, and CauST have varying performances across different datasets.
引用
"The failure of prior arts in OOD traffic data is due to ST contexts acting as a confounder."
"Our proposed DCA can estimate PΘ(Y |do(X)) via PΘ(Y |do(X)) = P(C = CI)PΘ(Y |X, C = CI) + P(C = CV )PΘ(Y |X, C = CV )."