[Paper Review] Dynamic Graph Convolutional Recurrent Network for Traffic Prediction: Benchmark and Solution

Fuxian Li, Jie Feng+4 more|arXiv (Cornell University)|Apr 30, 2021

Traffic Prediction and Management Techniques4 citations

TL;DR

This paper proposes Dynamic Graph Convolutional Recurrent Network (DGCRN), a novel framework that generates dynamic graph topologies at each time step using hyper-networks to model evolving spatial dependencies in traffic networks. By fusing learned dynamic graphs with static graphs and optimizing RNN training efficiency, DGCRN achieves state-of-the-art performance on three benchmark datasets, outperforming 15 baselines with a standardized benchmark and new dataset for fair evaluation.

ABSTRACT

Traffic prediction is the cornerstone of an intelligent transportation system. Accurate traffic forecasting is essential for the applications of smart cities, i.e., intelligent traffic management and urban planning. Although various methods are proposed for spatio-temporal modeling, they ignore the dynamic characteristics of correlations among locations on road networks. Meanwhile, most Recurrent Neural Network (RNN) based works are not efficient enough due to their recurrent operations. Additionally, there is a severe lack of fair comparison among different methods on the same datasets. To address the above challenges, in this paper, we propose a novel traffic prediction framework, named Dynamic Graph Convolutional Recurrent Network (DGCRN). In DGCRN, hyper-networks are designed to leverage and extract dynamic characteristics from node attributes, while the parameters of dynamic filters are generated at each time step. We filter the node embeddings and then use them to generate a dynamic graph, which is integrated with a pre-defined static graph. As far as we know, we are the first to employ a generation method to model fine topology of dynamic graph at each time step. Further, to enhance efficiency and performance, we employ a training strategy for DGCRN by restricting the iteration number of decoder during forward and backward propagation. Finally, a reproducible standardized benchmark and a brand new representative traffic dataset are opened for fair comparison and further research. Extensive experiments on three datasets demonstrate that our model outperforms 15 baselines consistently.

Motivation & Objective

• To address the limitation of static or adaptive adjacency matrices in capturing dynamic spatial correlations in traffic networks.
• To improve the efficiency and performance of RNN-based traffic prediction models by restricting decoder iterations during training.
• To establish a reproducible, standardized benchmark with a new representative traffic dataset for fair comparison across methods.
• To integrate both static distance-based graphs and dynamic attribute-based graphs for enhanced spatial dependency modeling.
• To demonstrate the effectiveness of dynamic graph generation in improving long-range and short-range traffic prediction accuracy.

Proposed method

• Hyper-networks generate dynamic graph filter parameters at each time step based on node attributes, enabling time-varying spatial modeling.
• Dynamic graph construction is performed by filtering node embeddings and integrating them with a pre-defined static graph to form a hybrid topology.
• A restricted iteration strategy during forward and backward propagation enhances training efficiency without sacrificing performance.
• The model uses a sequence-to-sequence architecture with gated recurrent units (GRUs) for temporal modeling, improved via scheduled sampling.
• Node embeddings are learned from spatial and temporal features, with GNNs in the hyper-network capturing higher-order spatial dependencies.
• The dynamic graph generator is trained end-to-end to jointly optimize for both spatial and temporal representation learning.

Experimental results

Research questions

• RQ1 Can dynamic graph construction at each time step improve traffic prediction accuracy compared to static or adaptive graphs?
• RQ2 How does the fusion of dynamic and static graphs affect performance in capturing complex spatio-temporal dependencies?
• RQ3 What is the impact of hyper-network-based dynamic filter generation on modeling real-time traffic correlation changes?
• RQ4 How does restricting RNN decoder iterations during training affect efficiency and prediction performance?
• RQ5 To what extent does the proposed benchmark and dataset enable fair and reproducible evaluation across different models?

Key findings

• DGCRN outperforms 15 state-of-the-art baselines across three datasets, demonstrating consistent superiority in MAE and RMSE metrics.
• The dynamic graph generation mechanism significantly improves long-range prediction, especially when combined with static graphs.
• The optimal hidden state dimension for DGCRN is around 64, with overfitting occurring at higher values.
• On the NE-BJ dataset, a node embedding dimension of 80–100 yields the best performance, indicating higher complexity requires richer representations.
• The dimension of GNN in the hyper-network should be balanced—too high leads to overfitting, while 16–32 provides optimal gains.
• The scheduled sampling decay step of 4000 on METR-LA yields the best performance, confirming the importance of hyperparameter tuning.