[Paper Review] Traffic Graph Convolutional Recurrent Neural Network: A Deep Learning Framework for Network-Scale Traffic Learning and Forecasting

Zhiyong Cui, Kristian Henrickson+3 more|arXiv (Cornell University)|Feb 20, 2018

Traffic Prediction and Management Techniques48 references100 citations

TL;DR

The paper proposes a Traffic Graph Convolutional Long Short-Term Memory (TGC-LSTM) framework that models road networks as graphs to learn spatiotemporal traffic patterns and forecast network-wide traffic states, with graph-convolution weights and features regularized for interpretability.

ABSTRACT

Traffic forecasting is a particularly challenging application of spatiotemporal forecasting, due to the time-varying traffic patterns and the complicated spatial dependencies on road networks. To address this challenge, we learn the traffic network as a graph and propose a novel deep learning framework, Traffic Graph Convolutional Long Short-Term Memory Neural Network (TGC-LSTM), to learn the interactions between roadways in the traffic network and forecast the network-wide traffic state. We define the traffic graph convolution based on the physical network topology. The relationship between the proposed traffic graph convolution and the spectral graph convolution is also discussed. An L1-norm on graph convolution weights and an L2-norm on graph convolution features are added to the model's loss function to enhance the interpretability of the proposed model. Experimental results show that the proposed model outperforms baseline methods on two real-world traffic state datasets. The visualization of the graph convolution weights indicates that the proposed framework can recognize the most influential road segments in real-world traffic networks.

Motivation & Objective

• Motivate and address the challenge of accurate network-scale traffic forecasting given time-varying patterns and complex spatial dependencies on road networks.
• Propose a graph-based deep learning framework that captures interactions between roadways to forecast network-wide traffic states.
• Define a traffic graph convolution grounded in physical network topology and relate it to spectral graph convolution.
• Enhance interpretability of the model by regularizing graph convolution weights (L1) and features (L2).

Proposed method

• Represent the traffic network as a graph to model road interactions.
• Introduce Traffic Graph Convolutional Long Short-Term Memory (TGC-LSTM) to learn spatiotemporal dependencies.
• Define a traffic graph convolution aligned with the physical topology and discuss its relation to spectral graph convolution.
• Apply an L1-norm on graph convolution weights and an L2-norm on graph convolution features in the loss to improve interpretability.
• Evaluate the model on two real-world traffic state datasets and compare to baseline methods.

Experimental results

Research questions

• RQ1 Can a graph-convolutional recurrent architecture effectively capture both spatial and temporal dependencies in network-scale traffic data?
• RQ2 Does incorporating topology-informed graph convolution improve forecasting accuracy over baselines?
• RQ3 Does regularization of graph convolution weights and features enhance model interpretability without sacrificing performance?
• RQ4 How does the proposed framework perform on real-world traffic datasets relative to baseline methods?

Key findings

• The proposed TGC-LSTM framework outperforms baseline methods on two real-world traffic state datasets.
• Graph convolution weights learned by the model align with influential road segments, indicating meaningful interpretability.
• The model captures interactions between roadways to forecast network-wide traffic states with improved accuracy.
• Regularization on weights and features aids interpretability while maintaining predictive performance.