A-STFGCN kills propagation delay to boost traffic predictions
Mask-based attention removes delay errors and beats 8 baselines on 5 datasets.
Current graph convolution-based models for traffic flow prediction suffer from two major flaws: they ignore the varying propagation delays between adjacent nodes when extracting spatial-temporal features, and they often stack complex structures that slow down training, compromising timeliness. To solve this, researchers propose A-STFGCN (Attention-Based Spatial-Temporal Fusion Graph Convolution Network). Its core is a spatial-temporal fusion block that removes propagation delay errors using a mask matrix within a multi-head self-attention mechanism. This allows the network to jointly capture short-term fluctuations and long-term trends in traffic data without the static bias introduced by traditional convolution.
Extensive experiments on five real-world traffic datasets show A-STFGCN achieves the best overall performance compared to eight baseline methods, including classic models like STGCN and DCRNN. It also demonstrates strong computation and data utilization efficiency — crucial for real-time deployment in smart city systems. By eliminating propagation delay, the model produces more accurate predictions of traffic flow, enabling better route optimization, congestion management, and urban mobility planning. This work, submitted to arXiv in July 2026, represents a significant step toward practical, low-latency traffic forecasting.
- Addresses propagation delays between adjacent nodes that previous graph convolution models ignore.
- Uses a mask-based multi-head self-attention mechanism in a spatial-temporal fusion block to remove delay errors.
- Achieves best overall accuracy across five real-world datasets while maintaining high computation and data efficiency.
Why It Matters
Real-time, delay-free traffic predictions can power smarter urban mobility systems and reduce congestion globally.