Reproduce “Cross Space and Time: A Spatio-Temporal Unitized Model for Traffic Flow Forecasting” on OpenResearch

About this paper

Weilin Ruan, Wenzhuo Wang, Siru Zhong, Wei Chen, Li Liu, Yuxuan Liang

The Spatio-Temporal Unitized Model (STUM), developed at the Hong Kong University of Science and Technology (Guangzhou), introduces a unified framework for traffic flow forecasting that concurrently captures spatial and temporal dependencies. It consistently improves prediction accuracy across various baseline models and datasets with minimal computational overhead, addressing limitations of existing spatio-temporal graph neural networks.

Key insights

• Unifying spatial and temporal information processing within a single adaptive cell, rather than separating modules, leads to more effective and efficient spatio-temporal representation learning.
• Low-rank matrix factorization provides an efficient mechanism to adaptively capture spatio-temporal dependencies, handling data heterogeneity while maintaining stable computational costs.
• A modular and generalizable framework can serve as a plug-and-play enhancement, consistently boosting the performance of diverse state-of-the-art spatio-temporal graph neural networks across various datasets and metrics.

Summary by alphaXiv.