Source-linked AI summary
T-GCN: A Temporal Graph ConvolutionalNetwork for Traffic Prediction
Ling Zhao, Yujiao Song, Chao Zhang, Yu Liu, Pu Wang, Tao Lin, Min Deng, Haifeng Li
TL;DR
Traffic forecasting must account for both urban road-network topology and traffic dynamics over time, yet existing approaches may omit spatial dependence or inadequately represent complex topologies. The paper proposes T-GCN, combining GCN for spatial dependence with GRU for temporal dependence, and reports lower prediction error than baselines on real-world datasets. The method also shows steady performance across prediction horizons, while errors remain when traffic records are absent or small values produce large relative differences.
Problem
Traffic forecasting is challenging because urban road topology and time-varying traffic create simultaneous spatial and temporal dependences.
Method
T-GCN combines GCN to model road-network topology and spatial dependence with GRU to model dynamic traffic changes and temporal dependence.
Results
1.5%-57.8%: T-GCN reduces prediction error compared with all baseline methods on Shenzhen taxi-speed and Los-loop datasets.
Takeaways & Limitations
T-GCN obtains the best prediction performance across all reported evaluation metrics and prediction horizons in the traffic-forecasting experiments.
Takeaways & Limitations
Prediction errors arise when no taxis produce zero-record conditions and when small traffic values make small differences yield large relative errors.
Abstract
from arXiv · showhide
Accurate and real-time traffic forecasting plays an important role in the Intelligent Traffic System and is of great significance for urban traffic planning, traffic management, and traffic control. However, traffic forecasting has always been considered an open scientific issue, owing to the constraints of urban road network topological structure and the law of dynamic change with time, namely, spatial dependence and temporal dependence. To capture the spatial and temporal dependence simultaneously, we propose a novel neural network-based traffic forecasting method, the temporal graph convolutional network (T-GCN) model, which is in combination with the graph convolutional network (GCN) and gated recurrent unit (GRU). Specifically, the GCN is used to learn complex topological structures to capture spatial dependence and the gated recurrent unit is used to learn dynamic changes of traffic data to capture temporal dependence. Then, the T-GCN model is employed to traffic forecasting based on the urban road network. Experiments demonstrate that our T-GCN model can obtain the spatio-temporal correlation from traffic data and the predictions outperform state-of-art baselines on real-world traffic datasets. Our tensorflow implementation of the T-GCN is available at https://github.com/lehaifeng/T-GCN.
1 INTRODUCTION
Traffic forecasting is difficult because urban road topology shapes spatial dependence while traffic changes over time through periodicity and trend. T-GCN combines GCN and GRU to model these dependencies, with reported improvements over baselines.
- Motivation: Traffic forecasting supports traffic management, planning, control, congestion response, and route selection, but remains challenging because traffic has complex spatial and temporal dependences.Traffic data include flow, speed, and density, whose patterns are analyzed to predict road conditions.
- Motivation: Road-network topology creates spatial dependence: upstream and downstream traffic states influence one another through transfer and feedback effects.Strong interactions between adjacent roads can alter short-term traffic-flow similarity across network positions.
- Motivation: Traffic volume exhibits temporal dependence through weekly periodicity and within-day trends, while current conditions can reflect earlier traffic states.The temporal patterns include both recurring changes and evolving conditions over different time scales.
- Research gap: Existing temporal methods often ignore spatial dependence, while conventional CNNs cannot represent complex urban road topologies because they are designed for Euclidean data.Ignoring the road network can reduce the accuracy of traffic-state prediction.
- Approach: T-GCN combines GCN for road-network topology and spatial dependence with GRU for dynamic traffic changes and temporal dependence.The model is designed for traffic forecasting on urban road networks and may also apply to other spatio-temporal forecasting tasks.
- Results: 1.5%-57.8%: T-GCN reduces prediction error compared with all baseline methods on Shenzhen Luohu taxi-speed and Los-loop datasets.The paper also reports a steady state across different prediction horizons, supporting both short-term and long-term traffic prediction tasks.
2 RELATED WORK
Traffic-forecasting research includes model-driven and data-driven approaches, but conventional methods have limitations in representing nonlinear, uncertain, and spatially structured traffic dynamics. Deep-learning approaches increasingly combine temporal and spatial modeling, although CNN-based methods remain constrained by complex road-network topology.
- Approaches: Traffic-forecasting methods include model-driven approaches based on prior system knowledge and data-driven approaches that learn statistical regularities from historical data.Model-driven methods describe relationships among traffic volume, speed, and density, whereas data-driven methods emphasize flexibility without modeling physical properties directly.
- Traditional methods: Parametric methods such as time-series, linear-regression, and Kalman-filtering models use specified relationships to predict future traffic data.ARIMA and its variants are presented as prominent time-series approaches in traffic forecasting.
- Traditional methods: Traditional parametric models are simple and convenient but assume stationarity, cannot reflect traffic nonlinearity and uncertainty, and are vulnerable to random events.Nonparametric models address these issues by learning statistical regularities from sufficient historical data.
- Deep learning: Deep neural networks gained attention because they capture dynamic traffic characteristics, with models distinguished by whether they include spatial dependence.Some neural approaches model temporal dependence alone, while others combine spatial and temporal features.
- Deep learning: CNN-LSTM and related models improve spatio-temporal forecasting, but CNNs remain limited because traffic networks have complex topologies rather than Euclidean grid structure.The related work includes models using CNNs for spatial dependence and LSTMs for temporal dynamics or variability.
- Research direction: The paper proposes a neural-network approach intended to capture complex temporal and spatial traffic features for forecasting on urban road networks.This approach addresses the related-work emphasis on using both spatial and temporal dependence.
3 METHODOLOGY
T-GCN forecasts traffic information from historical road data by combining graph convolution for spatial dependence with GRU-based temporal modeling. The method represents the road network as a graph, extracts spatial features from topology, models dynamic changes over time, and trains predictions with regularized loss.
- 3.1 Problem Definition: Traffic forecasting predicts future road traffic information from historical observations represented on an urban road-network graph.Traffic information may include speed, flow, or density; the experiments use traffic speed as an example.
- 3.3.1 Spatial Dependence Modeling: The graph convolutional network captures spatial dependence by learning topological relationships among connected roads and their attributes.The road network is modeled with an adjacency matrix, and graph convolution operates on nodes and their first-order neighborhoods.
- 3.3.2 Temporal Dependence Modeling: The gated recurrent unit captures temporal dependence from the dynamically changing traffic sequence while using fewer parameters and faster training than LSTM.GRU gates regulate information from previous states, supporting temporal modeling for traffic data.
- 3.3.3 Temporal Graph Convolutional Network: T-GCN combines graph convolution and gated recurrent units so spatial and temporal dependencies are modeled together for traffic prediction.Historical time-series data first pass through graph convolution for spatial features, then through GRU processing for dynamic changes.
- 3.3.4 Loss Function: The model is trained by minimizing prediction error between real and predicted traffic speed with an added L2 regularization term.The regularization coefficient λ controls the contribution of the L2 term, which helps avoid overfitting.
4 EXPERIMENTS
The experiments evaluate T-GCN on SZ-taxi and Los-loop using multiple metrics, hidden-unit settings, baselines, prediction horizons, and perturbations. T-GCN achieves strong spatio-temporal forecasting performance, remains effective across horizons, and shows robustness to noise, while errors remain at peaks and low traffic values.
- Model parameters: 100 hidden units are optimal for SZ-taxi, whereas 64 hidden units produce the highest precision and lowest error for Los-loop.For SZ-taxi, increasing hidden units beyond the optimum reduces precision as model complexity and computational difficulty increase.
- Prediction results: T-GCN obtains the best prediction performance across all evaluation metrics and prediction horizons on both SZ-taxi and Los-loop.The comparison covers 15-, 30-, 45-, and 60-minute horizons.
- Spatio-temporal capability: 57.8% lower RMSE than GCN at 15 minutes and 57.6% lower RMSE at 30 minutes demonstrate stronger spatial-feature modeling.GCN considers spatial features only.
- Long-term prediction: T-GCN predictions remain relatively stable as the prediction horizon changes and achieve the best baseline comparison regardless of horizon.The reported horizon analysis covers both prediction error and precision.
- Robustness: Prediction performance changes little under Gaussian and Poisson noise, indicating robustness to high-noise conditions.The perturbation analysis covers both SZ-taxi and Los-loop.
- Model interpretation: The model predicts poorly at peaks, and missing records for roads without taxis plus large relative errors at low traffic values contribute to remaining discrepancies.The paper attributes peak smoothing to the GCN’s smooth Fourier-domain filter.
5 CONCLUSION
The paper concludes that T-GCN combines graph convolution and recurrent modeling to capture spatial and temporal traffic dependence. On two real-world datasets, it achieves the best results across prediction horizons and shows robustness in perturbation analysis.
- Conclusion: T-GCN combines GCN and GRU components to model spatial dependence from road topology and temporal dependence from dynamic node attributes.Roads are represented as graph nodes, connections as edges, and traffic information as node attributes.
- Conclusion: T-GCN achieves the best prediction results across different horizons on the SZ-taxi and Los-loop datasets compared with five baseline models.The baselines are HA, ARIMA, SVR, GCN, and GRU.
- Conclusion: Perturbation analysis illustrates the robustness of the proposed approach.
- Conclusion: The authors state that T-GCN successfully captures spatial and temporal features for spatio-temporal traffic forecasting.