SFADNet: Spatio-temporal Fused Graph based on Attention Decoupling Network for Traffic Prediction

Mei Wu, Wenchao Weng, Jun Li, Yiqian Lin, Jing Chen, Dewen Seng

TL;DR

SFADNet is a new deep learning framework that improves traffic flow prediction by "decoupling" complex traffic into distinct sub-patterns. By generating adaptive, fused spatio-temporal graphs for each pattern, it captures the shifting relationships between road sensors far more accurately than traditional static models.

What Problem Does This Paper Address?

Predicting traffic is notoriously difficult because the relationship between two locations isn't just about physical distance—it’s dynamic. For example, two sensors near different office parks might show highly correlated traffic during morning rush hour but have zero relationship at midnight.

Most existing models rely on static spatial graphs (like a fixed map of the road network) or fail to account for how different types of traffic (commuters vs. local shoppers) mix together. The authors argue that we need a way to separate these overlapping traffic patterns and model their unique spatial and temporal dependencies independently.

Methods

The authors introduce SFADNet (Spatio-temporal Fused Graph based on Attention Decoupling Network), which consists of three main innovations:

  1. Decouple Module: Instead of treating traffic flow as a single number, this module uses node and time embeddings to split the total flow into $M$ independent "traffic patterns." This allows the model to "see" different underlying behaviors simultaneously.
  2. Dynamic Graph Generation (DGG): Rather than using a fixed roadmap, SFADNet builds an adaptive graph. It uses a cross-attention mechanism to fuse spatial features (where sensors are) and temporal features (periodic trends) into a single "fused" adjacency matrix. This matrix is asymmetric, capturing how one road's traffic might influence another's without necessarily being influenced back.
  3. Residual Graph Convolution (RGC) & TS Modules: Each decoupled traffic pattern is processed through its own RGC module to capture spatial relationships. These are then fed into a Temporal Sequence (TS) module—specifically a Gated Recurrent Unit (GRU)—to handle short-term timing and periodic dependencies.

Datasets

The model was evaluated on four standard large-scale benchmarks from the California Department of Transportation (CalTrans) Performance Measurement System (PeMS):

  • PEMS03, PEMS04, PEMS07, and PEMS08.
  • The data consists of traffic flow aggregated at 5-minute intervals.
  • The task: Predict the next 60 minutes (12 time steps) based on the previous 60 minutes of history.

Key Insights and Ideas

  • The "Aha" Moment: Traffic is a Mixture. The core insight is that "traffic" is an aggregate of different behaviors. By decoupling these behaviors, the model can apply a specific "logic" (graph) to the commuter flow and a different "logic" to the local flow.
  • Dynamic Sparsity: To keep the model efficient, the authors only retain the "Top-K" strongest relationships in their generated graphs. This prevents the model from being overwhelmed by noise from distant, unrelated sensors.
  • Fusion via Attention: Simply stacking time and space data isn't enough. Using cross-attention allows the model to dynamically decide which is more important—the physical location or the time of day—when calculating the strength of a connection between two nodes.

Key Results

SFADNet consistently outperformed state-of-the-art baselines (including popular models like GraphWaveNet and STG-NCDE) across all four datasets:

  • Accuracy: On the PEMS03 dataset, SFADNet achieved a Mean Absolute Error (MAE) of 14.66, significantly lower than the next best model (ST-AE at 15.44).
  • Pattern Impact: Ablation studies showed that increasing the number of decoupled patterns ($M$) generally improved results. For instance, on PEMS04, moving from no decoupling to 3 patterns dropped the MAE from 18.75 to 18.32.
  • Graph Quality: The "fused" graph (combining time and space) performed better than using just a spatial graph or just a temporal graph, proving that the interaction between the two is where the most predictive power lies.

Limitations and Future Work

While the paper demonstrates strong empirical results, a few areas remain for exploration:

  • Interpretability: While the model "decouples" traffic into patterns, the paper doesn't explicitly label what these patterns represent (e.g., "this is the school run pattern"). Future work could look into making these decoupled streams more interpretable for city planners.
  • Computational Overhead: Generating dynamic graphs and processing multiple decoupled streams is more computationally intensive than static methods. Although the authors use "Top-K" sparsification, scaling this to city-wide networks with tens of thousands of sensors might require further optimization.
  • External Factors: The current model relies on historical flow and embeddings. Incorporating external data like weather, holidays, or accidents could further refine the "decoupling" process.

Summary updated

Abstract

In recent years, traffic flow prediction has played a crucial role in the management of intelligent transportation systems. However, traditional prediction methods are often limited by static spatial modeling, making it difficult to accurately capture the dynamic and complex relationships between time and space, thereby affecting prediction accuracy. This paper proposes an innovative traffic flow prediction network, SFADNet, which categorizes traffic flow into multiple traffic patterns based on temporal and spatial feature matrices. For each pattern, we construct an independent adaptive spatio-temporal fusion graph based on a cross-attention mechanism, employing residual graph convolution modules and time series modules to better capture dynamic spatio-temporal relationships under different fine-grained traffic patterns. Extensive experimental results demonstrate that SFADNet outperforms current state-of-the-art baselines across four large-scale datasets.
BibTeX
@misc{wu2025sfadnet,
  author = {Mei Wu and Wenchao Weng and Jun Li and Yiqian Lin and Jing Chen and Dewen Seng},
  title = {{SFADNet: Spatio-temporal Fused Graph based on Attention Decoupling Network for Traffic Prediction}},
  year = {2025},
  eprint = {2501.04060},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2501.04060},
}
Figure