LSGCN: Long Short-Term Traffic Prediction with Graph Convolutional Networks

Rongzhou Huang, Chuyin Huang, Yubao Liu, Genan Dai, Weiyang KongIJCAI 2020 · DOI

TL;DR

This paper introduces LSGCN, a novel deep learning framework for traffic prediction that excels at both short-term and long-term forecasting. By integrating a new graph attention mechanism (cosAtt) with graph convolutional networks (GCN) in a "spatial gated block," LSGCN effectively captures complex spatial-temporal dependencies in traffic data, leading to more stable and accurate predictions.

What Problem Does This Paper Address?

Traffic prediction is crucial for applications like intelligent route planning and traffic management. However, accurately forecasting traffic conditions over varying time horizons (short-term: 5-30 mins, long-term: 30-60 mins) remains a challenge. Traditional methods struggle with the inherent nonlinearity and complexity of traffic data. While deep learning approaches, particularly those using graph convolutional networks (GCNs), have shown promise, existing methods often fall short in handling both short- and long-term predictions simultaneously. Some iterative prediction methods accumulate errors over time, while non-iterative methods may not adequately consider the temporal dynamics. This paper aims to bridge this gap by proposing a framework that can effectively capture both spatial and temporal features for accurate predictions across different time scales.

Methods

The core of LSGCN is its novel spatial gated block, which combines two key components:

  • Graph Convolutional Network (GCN): Used to capture the spatial location adjacency of road networks. This leverages the inherent graph structure of traffic systems.
  • cosAtt (new Graph Attention Network): This novel attention mechanism is designed to capture the similarity in road conditions between different sensors, even if they are not directly adjacent. It uses a cosine similarity measure between node features, making it more stable than traditional attention mechanisms that might rely on learned weights that can fluctuate.

These two components are integrated within the spatial gated block using a gated mechanism, allowing them to selectively leverage spatial adjacency and road condition similarity. The LSGCN architecture also incorporates Gated Linear Units (GLUs) for capturing temporal features and a convolution unified layer to merge time steps. A key design choice is the adoption of an iterative prediction mechanism, similar to some previous works, but with a simplified network structure to mitigate error accumulation.

Datasets

The paper evaluates LSGCN on three real-world traffic datasets:

  • PeMSD4: Collected from the San Francisco Bay Area, with 307 sensors.
  • PeMSD7: Collected from District 7 of California, with 228 sensors.
  • PeMSD8: Collected from San Bernardino, with 170 sensors.

These datasets are standard benchmarks in traffic prediction research, previously used in studies like STGCN and ASTGCN. The paper provides detailed statistics on the number of nodes, edges, time steps, and time spans for each dataset.

Key Insights and Ideas

  • Road condition similarity matters: The paper highlights that not just adjacent roads but also roads with similar traffic conditions (e.g., similar congestion levels) influence each other. The novel cosAtt mechanism is designed to capture this, offering a more nuanced understanding of spatial dependencies.
  • Stability through cosAtt: By using a cosine similarity and a sigmoid activation, cosAtt provides more stable attention weights compared to traditional GATs, which can be prone to instability due to learned attention functions.
  • Balancing iterative prediction with error reduction: LSGCN adopts iterative prediction for its simulation capabilities but addresses the common issue of error accumulation by using a simplified network structure with fewer layers.
  • Unified spatial-temporal feature extraction: The spatial gated block effectively fuses information from spatial adjacency (GCN) and road condition similarity (cosAtt), creating a powerful module for extracting complex spatio-temporal features.

Key Results

LSGCN demonstrates strong performance across all three datasets and evaluation metrics (MAE, MAPE, RMSE) for both short-term and long-term predictions.

  • Superior performance: LSGCN consistently outperforms baseline methods like HA, ARIMA, DCRNN, STGCN, and ASTGCN, particularly in long-term prediction tasks and on complex traffic networks like PeMSD4.
  • Stable predictions: Replacing cosAtt with a traditional graph attention mechanism (LSGCN(GAT)) resulted in less stable prediction results, underscoring the effectiveness of cosAtt.
  • Efficient training: LSGCN exhibits significantly faster training convergence compared to DCRNN and ASTGCN, requiring approximately 1,000 seconds versus 10,000-25,000 seconds, attributed to its more streamlined architecture.
  • Accurate capture of dynamic events: In extreme cases like morning and evening rush hours, LSGCN accurately captures the trend and timing of these dynamic changes.

Limitations and Future Work

  • Limitations: The paper acknowledges that while LSGCN reduces error accumulation compared to STGCN, it's still an iterative prediction method, and some error accumulation might persist, especially for very long-term predictions. The paper also notes that DCRNN performed best on PeMSD4 due to its diffusion convolution, suggesting that specific network structures might be more advantageous for certain types of traffic networks.
  • Future Work: The authors suggest extending the proposed model to other spatio-temporal structured sequence forecasting tasks, such as preference prediction in recommendation systems. They also imply that further research into optimizing the iterative prediction mechanism to further minimize error accumulation would be beneficial.

Summary updated

BibTeX
@inproceedings{huang2020lsgcn,
  author = {Rongzhou Huang and Chuyin Huang and Yubao Liu and Genan Dai and Weiyang Kong},
  title = {{LSGCN: Long Short-Term Traffic Prediction with Graph Convolutional Networks}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.24963/ijcai.2020/326},
}
Figure