Source-linked AI summary
Dynamic Graph Convolutional Recurrent Network for Traffic Prediction: Benchmark and Solution
Fuxian Li, Jie Feng, Huan Yan, Guangyin Jin, Depeng Jin, Yong Li
TL;DR
Traffic prediction supports intelligent transportation systems, but existing methods often miss time-varying road correlations, RNNs are inefficient, and fair comparisons are limited. DGCRN generates step-wise dynamic graphs alongside static structure, uses an efficient RNN training strategy, and is evaluated through a benchmark and new dataset; it outperforms 15 baselines across three datasets.
Problem
Traffic prediction is central to intelligent transportation systems, but methods face complex spatio-temporal correlations, inefficient recurrent operations, and limited fair comparison across common datasets.
Method
DGCRN uses a hyper-network to generate step-wise dynamic adjacency information, integrates dynamic and static graphs, and applies an efficient training strategy for RNN-based forecasting.
Results
DGCRN achieves state-of-the-art prediction performance on three datasets, improving 2%∼4% on METR-LA and PEMS-BAY and around 6% on NE-BJ over baselines.
Takeaways & Limitations
The paper provides a reproducible benchmark covering 15 representative methods across three datasets and releases a new congestion-related dataset for further traffic-prediction research.
Abstract
from arXiv · showhide
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.
1 INTRODUCTION
Traffic prediction supports intelligent transportation and urban decision-making, but existing approaches face dynamic-topology, efficiency, and fair-evaluation challenges. The paper proposes DGCRN, a dynamic graph recurrent framework, alongside a standardized benchmark, a new dataset, and experiments against 15 baselines.
- Traffic prediction supports traffic management, urban planning, congestion reduction, route planning, and public-safety early warning.
- Dynamic spatial correlations vary with real-time traffic conditions and road-network topology, while temporal dependencies combine nonlinear changes with periodicity.
- Existing pre-defined and adaptive adjacency matrices remain static, limiting their ability to represent changing road-network topology.
- RNN recurrent operations restrict training speed and hinder efficient sequence-to-sequence architectures for traffic prediction.
- DGCRN generates dynamic adjacency matrices step by step with a hyper-network and integrates the dynamic graph with a pre-defined graph.
- The paper introduces an RNN training strategy, publishes a congestion-related dataset, and benchmarks 15 representative methods across three public datasets.
- Experiments report significantly reduced prediction error and state-of-the-art accuracy compared with 15 baselines.
2 LITERATURE REVIEW
Prior traffic-prediction research spans spatial topology, spatial and temporal dependency modeling, and external features. The paper distinguishes its approach through step-wise dynamic graph generation, static-graph fusion, and an RNN training strategy intended to improve efficiency and performance.
- External features: Traffic-prediction models also incorporate external features such as weather, points of interest, time of day, and day of week.
- The paper provides source code and dataset access through the Traffic-Benchmark repository.
- Spatial topology construction: Graph-based methods address non-Euclidean road networks using distance, connectivity, multi-view, or adaptive adjacency constructions.
- Spatial topology construction: Pre-defined adjacency matrices encode fixed relations, whereas adaptive adjacency matrices learn their parameters during training.
- Spatial dependency modeling: Traditional single-series methods often rely on stationarity assumptions and do not capture correlations among multiple traffic variables.
- Temporal dependency modeling: RNNs capture temporal dependencies flexibly but recurrent operations increase time and memory consumption.
- The proposed model generates graphs at each time step for fine-grained topology modeling, fuses them with static graphs, and uses an RNN strategy targeting efficiency and performance.
3 PRELIMINARIES AND PROBLEM FORMULATION
Traffic prediction is formulated as multivariate graph-signal forecasting with a weighted directed graph as auxiliary prior knowledge. Given observed graph signals and the graph, a function predicts the next sequence of graph signals.
- The prior knowledge is a weighted directed graph G = (V, E, A), where vertices represent road-network locations and A encodes inter-vertex proximity.
- Graph signals form a tensor X ∈ R^T×N×D, with T sequence length, N nodes, and D features per node.
- At each time step, the graph signal X_t contains the features of all N nodes and has shape R^N×D.
- The forecasting function maps P observed graph-signal steps and graph G to Q future graph-signal steps.
4 METHODOLODY
DGCRN generates a dynamic graph at each time step and combines it with a predefined static graph inside recurrent graph-convolution modules. It also reduces decoder iterations during training to improve efficiency while preserving long-horizon learning.
- Graph Generator: DGCRN uses separate encoder and decoder graph generators to produce dynamic adjacency matrices synchronously with recurrent operations.The generated dynamic graph and predefined static graph are then used for graph convolution.
- Graph Generator: The graph generator uses hyper-networks to create dynamic filters, filters node embeddings, and derives adjacency from pairwise node similarities.Two hyper-networks generate filters for source and target node embeddings, whose similarities define the dynamic adjacency matrix.
- Dynamic Graph Convolutional Recurrent Module: DGCRM replaces classical GRU matrix multiplications with dynamic graph convolutions to model spatial and temporal correlations simultaneously.Dual-directional graph convolution is used for directed graphs, and the convolution includes dynamic and static adjacency information.
- Graph Generator: At each time step, speed, time of day, and the previous hidden state are concatenated as hyper-network input for dynamic graph generation.This incorporates current and historical information into the dynamic node features used by the graph generator.
- Dynamic Graph Convolutional Recurrent Module: The dynamic graph convolution combines static and dynamic graph results through a weighted sum, with learnable propagation parameters and component weights.The module supports k-hop propagation and uses the resulting node states within the recurrent architecture.
- Training Strategy: Curriculum learning trains the decoder on only the first i steps initially, increasing i until the full prediction horizon is reached.This reduces early-stage forward and backward computation and GPU memory use while providing a basis for longer-range prediction.
5.1 Experiment Settings
Experiments use three real-world traffic-speed datasets, standardized preprocessing, and repeated evaluation against traditional and deep-learning baselines. The datasets differ in spatial correlations and traffic complexity.
- Datasets: Experiments use METR-LA, PEMS-BAY, and the newly published NE-BJ traffic-speed datasets.NE-BJ contains 500 road segments from congested main roads in northeast Beijing, while METR-LA and PEMS-BAY contain 207 and 325 sensors, respectively.
- Datasets: NE-BJ was introduced to represent complex urban congestion that freeway speed datasets may not fully capture.It was collected from Tencent Map navigation data on weekdays in July 2020.
- Dataset analysis: Inter-node correlations are strongest in METR-LA, weaker in PEMS-BAY, and intermediate but relatively stronger than PEMS-BAY in NE-BJ.The comparison uses Pearson correlations between all node pairs.
- Preprocessing: Data are sampled every 5 minutes, Z-Score normalized, and split into training, validation, and testing periods with dataset-specific proportions or dates.METR-LA and PEMS-BAY use 70% training, 10% validation, and 20% testing; NE-BJ uses 15, 3, and 5 days, respectively.
- Baselines and metrics: The benchmark compares DGCRN with statistical, CNN-based, GNN-based, RNN-based, and attention-based traffic-prediction models using MAE, RMSE, and MAPE.Baselines include HA, VAR, SVR, FNN, ARIMAkal, FC-LSTM, DCRNN, STGCN, Graph WaveNet, ST-MetaNet, AGCRN, GMAN, and MTGNN.
5.2 Traffic Benchmark
The benchmark standardizes comparisons across three datasets and forecasting horizons, while results and ablations show benefits from dynamic graph modeling, static-graph fusion, and the proposed training strategy. Performance varies with dataset complexity and hyperparameter choices.
- Traffic Benchmark: The open-source benchmark evaluates 15 representative methods on three datasets using 15-minute, 30-minute, and 1-hour forecasting horizons.Missing test values are masked and excluded from metrics and loss.
- Results: DGCRN achieves state-of-the-art performance across all three datasets, improving 2%–4% on METR-LA and PEMS-BAY and around 6% on NE-BJ.It generally outperforms other baselines except for certain long-range metrics where GMAN is stronger on PEMS-BAY and NE-BJ.
- Results: Prediction difficulty varies substantially across datasets: NE-BJ is more complex, while PEMS-BAY is easier for the evaluated models.ARIMA and STSGCN are not listed for NE-BJ because they break down on its complex traffic-speed series.
- Ablation Study: Ablations show that dynamic and static graphs, hyper-networks, generator inputs, and the proposed training strategy each contribute to DGCRN performance.Removing dynamic information or generator inputs degrades results, especially for long-range prediction.
- Ablation Study: Replacing the Hadamard product with matrix multiplication sharply reduces performance while increasing dynamic-filter parameters fortyfold.The matrix-multiplication variant requires reshaping the dynamic filter tensor before multiplication.
- Parameter Study: Parameter studies find dataset-dependent optima, including a hidden-state dimension around 64 and higher node-embedding dimensions for NE-BJ.The reported optimal scheduled-sampling decay on METR-LA is around 4000, while overly large hidden or hyper-network dimensions can overfit.
6 CONCLUSION
The paper concludes that step-wise dynamic graph generation can complement a pre-defined graph and improve traffic prediction. It presents DGCRN, an efficient training strategy, the NE-BJ dataset, and a standardized benchmark as resources for further research.
- DGCRN generates a step-wise dynamic graph that cooperates with a pre-defined graph for traffic prediction.The dynamic graph learns adjacency-matrix parameters and self-adjusts using dynamic information at each time step.
- The authors preliminarily show that dynamic graph generation can improve prediction performance.
- The paper uses a general RNN training method described as more efficient and beneficial for performance.
- The NE-BJ dataset is presented as representative of urban traffic conditions and valuable for further research.
- A reproducible traffic benchmark is provided to support fair comparisons and convenient model design.
- Future work will address traffic congestion and incorporate external factors such as temperature, weather, points of interest, and accidents.