EUVIMEDEUVIMED entdecken
Crossref

Taming Spatial Heterophily and Temporal Irregularity: A Curriculum Learning Approach for Traffic Forecasting

Hongjun Wang, Zhiwen Zhang, Jiyuan Chen, Zipei Fan, Renhe Jiang, Wei Yuan, Ryosuke Shibasaki, Xuan Song

IEEE Transactions on Intelligent Transportation Systems · 2026-08 · 27 (8) · 9475-9491

DOI: 10.1109/tits.2026.3685281 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Transactions on Intelligent Transportation Systems, 2026-08. Verfasst von Hongjun Wang und 7 weiteren Autor:innen. Die Arbeit nennt 48 Literaturangaben. Gefördert von: National Science Foundation of China, Jilin Provincial International Cooperation Key Laboratory for Super Smart City, Jilin Provincial Key Laboratory of Intelligent Policing.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

The advance in graph neural networks (GCNs) has propelled the development of spatial-temporal (ST) graph modeling (e.g., traffic flow prediction, and incident prediction) that intends to capture the inter-dependency between connected nodes and time sequence self-correlation simultaneously from dynamic node-level inputs. However, the existing literature trains their models with a random order, ignoring that the forecasting difficulty can vary significantly among node’s time sequences. Intuitively, the difficult samples indicate those patterns (e.g., abnormal patterns) are hard to learn by model directly. Thus, we consider that a curriculum with easy-to-difficult order, similar to the understanding procedure of humans, can improve the method’s performance in ST-graph modeling. Such an idea has a well-known name: curriculum learning (CL), which firstly designs the difficulty measurer for training samples and then feeds into the model with easy-to-difficult order according to the measurer. This paper explores the difficult time sequence whose temporal pattern is inconsistent with neighbors or has a poor auto-correlation. We design a spatial-temporal curriculum learning framework (STCL) trained by two additional ST tasks to obtain the vital power in ST imputation and measure each node’s signal difficulty corresponding to the performance in the auxiliary model in different ST scenarios. The ST-models are learned from scratch with explicit spatial affinities (subgraph) and easy temporal patterns, then progressively enrolled the nodes into the training dataset. Extensive experiments have been conducted in two public traffic flow datasets and two incident datasets, demonstrating our curriculum schedule’s superiority over without CL and other CL baselines.

Förderung

  • National Science Foundation of China
  • Jilin Provincial International Cooperation Key Laboratory for Super Smart City
  • Jilin Provincial Key Laboratory of Intelligent Policing

Zitieren

APA

Wang, H., Zhang, Z., Chen, J., Fan, Z., Jiang, R., Yuan, W., Shibasaki, R., & Song, X. (2026). Taming Spatial Heterophily and Temporal Irregularity: A Curriculum Learning Approach for Traffic Forecasting. IEEE Transactions on Intelligent Transportation Systems, 27(8), 9475-9491. https://doi.org/10.1109/tits.2026.3685281

Vancouver

Wang H, Zhang Z, Chen J, Fan Z, Jiang R, Yuan W, et al. Taming Spatial Heterophily and Temporal Irregularity: A Curriculum Learning Approach for Traffic Forecasting. IEEE Transactions on Intelligent Transportation Systems. 2026;27(8):9475-9491. doi:10.1109/tits.2026.3685281

BibTeX
@article{Wang2026,
  author = {Wang, Hongjun and Zhang, Zhiwen and Chen, Jiyuan and Fan, Zipei and Jiang, Renhe and Yuan, Wei and Shibasaki, Ryosuke and Song, Xuan},
  title = {Taming Spatial Heterophily and Temporal Irregularity: A Curriculum Learning Approach for Traffic Forecasting},
  journal = {IEEE Transactions on Intelligent Transportation Systems},
  year = {2026},
  volume = {27},
  number = {8},
  pages = {9475--9491},
  doi = {10.1109/tits.2026.3685281}
}
Kostenlos starten →