Incorporating Reachability Knowledge into a Multi-Spatial Graph Convolution Based Seq2Seq Model for Traffic Forecasting

Jiexia Ye, Furong Zheng, Juanjuan Zhao, Kejiang Ye, Chengzhong Xu

Abstract

Accurate traffic state prediction is the foundation of transportation control and guidance. It is very challenging due to the complex spatiotemporal dependencies in traffic data. Existing works cannot perform well for multi-step traffic prediction that involves long future time period. The spatiotemporal information dilution becomes severe when the time gap between input step and predicted step is large, especially when traffic data is not sufficient or noisy. To address this issue, we propose a multi-spatial graph convolution based Seq2Seq model. Our main novelties are three aspects: (1) We enrich the spatiotemporal information of model inputs by fusing multi-view features (time, location and traffic states) (2) We build multiple kinds of spatial correlations based on both prior knowledge and data-driven knowledge to improve model performance especially in insufficient or noisy data cases. (3) A spatiotemporal attention mechanism based on reachability knowledge is novelly designed to produce high-level features fed into decoder of Seq2Seq directly to ease information dilution. Our model is evaluated on two real world traffic datasets and achieves better performance than other competitors.

Introduction

Traffic forecasting is a classic spatiotemporal challenge: what happens at one sensor is influenced by its own history and the state of distant sensors. However, existing models struggle with two main issues:

  1. Information Dilution: In standard Sequence-to-Sequence (Seq2Seq) models, as the "gap" between the input time step and the predicted future step increases, the signal from the encoder becomes increasingly diluted. This leads to poor long-term (multi-step) accuracy.
  2. Data Quality: Real-world traffic data is often "dirty"—noisy, sparse, or missing. Purely data-driven models fail when data is poor, while models relying only on fixed prior knowledge are too rigid to capture complex traffic patterns.

Methodology

The proposed MSGC-Seq2Seq architecture addresses these challenges through a three-stage approach.

Spatiotemporal Attribute Fusion

Instead of relying solely on raw traffic flow or speed, the model enriches inputs by fusing them with spatial and temporal embeddings. Spatial embeddings are generated using node2vec on the road network, while temporal embeddings are created using DeepWalk on a "temporal line graph" that captures the sequential nature of time slots.

Multi-view Spatial Graph Convolutions (GCNs)

The model utilizes three distinct matrices to capture complex spatial relationships:

  • Adjacency Matrix: Based on geographical proximity and historical trend similarity.
  • Semantic Matrix: A dynamic, data-driven matrix that identifies hidden correlations between regions with similar flow patterns, regardless of physical distance.
  • Reachability Matrix: A novel cross-step matrix based on average travel time ($M_{ij}$) between nodes. It models the physical process of traffic flow, recognizing that an event at one point may only affect another point after a specific time lag.
The MSGC-Seq2Seq architecture integrates multi-view spatial correlations and a reachability-aware decoder to combat information dilution.

Reachability-Aware Seq2Seq

To combat information dilution, high-level features extracted from the reachability GCN are fed directly into the decoder at each output step. This architectural "shortcut" ensures the decoder maintains fresh spatial context for every future step it predicts, bypassing the typical RNN bottleneck.

Experimental Results

The model was evaluated on two real-world highway traffic datasets: METR-LA (207 sensors) and PeMS (325 sensors).

Performance and Long-Term Accuracy

MSGC-Seq2Seq consistently outperformed state-of-the-art baselines, including STGCN, DCRNN, and GMAN, across all metrics (MAE, RMSE, MAPE). Notably, the model's advantage becomes more pronounced as the prediction horizon increases. For 60-minute (12-step) forecasts on METR-LA, the model achieved a MAPE of 8.70%, significantly outperforming STGCN (11.73%) and DCRNN (9.59%).

Performance comparison (MAE, RMSE, MAPE) across prediction steps on METR and PEMS datasets, demonstrating the model's superior long-term accuracy.

Robustness and Efficiency

In fault tolerance tests where up to 90% of data was replaced with noise, MSGC-Seq2Seq showed the slowest rate of performance deterioration. This robustness stems from its integration of reachability and prior knowledge. Furthermore, the model is computationally efficient, training nearly twice as fast as GMAN per epoch by utilizing pre-computed reachability matrices.

Conclusion

MSGC-Seq2Seq effectively addresses the information dilution problem in multi-step traffic forecasting by incorporating physical reachability knowledge. By fusing multi-view spatial correlations and employing a reachability-aware decoder, the model achieves superior accuracy and robustness. Future work may explore dynamic reachability matrices that adapt to real-time congestion levels and the explicit modeling of external factors like weather and holidays.

Summary updated

Abstract

Accurate traffic state prediction is the foundation of transportation control and guidance. It is very challenging due to the complex spatiotemporal dependencies in traffic data. Existing works cannot perform well for multi-step traffic prediction that involves long future time period. The spatiotemporal information dilution becomes serve when the time gap between input step and predicted step is large, especially when traffic data is not sufficient or noisy. To address this issue, we propose a multi-spatial graph convolution based Seq2Seq model. Our main novelties are three aspects: (1) We enrich the spatiotemporal information of model inputs by fusing multi-view features (time, location and traffic states) (2) We build multiple kinds of spatial correlations based on both prior knowledge and data-driven knowledge to improve model performance especially in insufficient or noisy data cases. (3) A spatiotemporal attention mechanism based on reachability knowledge is novelly designed to produce high-level features fed into decoder of Seq2Seq directly to ease information dilution. Our model is evaluated on two real world traffic datasets and achieves better performance than other competitors.
BibTeX
@misc{ye2021incorporating,
  author = {Jiexia Ye and Furong Zheng and Juanjuan Zhao and Kejiang Ye and Chengzhong Xu},
  title = {{Incorporating Reachability Knowledge into a Multi-Spatial Graph Convolution Based Seq2Seq Model for Traffic Forecasting}},
  year = {2021},
  eprint = {2107.01528},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2107.01528},
}
Figure