T-GCN: A Temporal Graph Convolutional Network for Traffic Prediction

Zhao, Ling; Song, Yujiao; Zhang, Chao; Liu, Yu; Wang, Pu; Lin, Tao; Deng, Min; Li, Haifeng · 2019 · OpenAlex-citations

DOI: 10.1109/tits.2019.2935152

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of accurate, real-time traffic forecasting, a critical component of Intelligent Traffic Systems for urban planning and management. The authors identify that existing methods often fail to simultaneously capture the complex spatial and temporal dependencies inherent in traffic data. Spatial dependence is governed by the topological structure of road networks, where traffic states influence adjacent roads, while temporal dependence involves dynamic changes such as periodicity and trends. Traditional convolutional neural networks (CNNs) are ill-suited for non-Euclidean graph structures like road networks, and many time-series models ignore spatial constraints. To resolve this, the authors propose the Temporal Graph Convolutional Network (T-GCN), a novel neural network architecture designed to model both dependencies concurrently. The T-GCN model integrates a Graph Convolutional Network (GCN) with a Gated Recurrent Unit (GRU). The GCN component processes the road network’s adjacency matrix and traffic feature matrices to learn complex topological structures, thereby capturing spatial dependence. The GRU component processes the resulting spatial features over time to capture dynamic changes and temporal dependence, utilizing update and reset gates to manage long-term information retention more efficiently than Long Short-Term Memory (LSTM) networks. The model predicts future traffic information (e.g., speed, flow, density) by learning a mapping function from historical data and network topology. The training objective minimizes the error between predicted and actual values using a loss function that includes L2 regularization to prevent overfitting. The authors evaluate T-GCN on two real-world datasets: the SZ-taxi dataset, comprising taxi trajectory speeds from 156 roads in Shenzhen’s Luohu District aggregated every 15 minutes, and the Los-loop dataset, containing speed data from 207 sensors in Los Angeles aggregated every 5 minutes. The experiments compare T-GCN against several baselines, including History Average, ARIMA, Kalman Filtering, Support Vector Regression, and various deep learning models like DCNN and DCRNN. Performance was measured using Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), Accuracy, Coefficient of Determination (R²), and Explained Variance Score across prediction horizons of 15, 30, 45, and 60 minutes. The results demonstrate that T-GCN significantly outperforms state-of-the-art baselines, reducing prediction errors by approximately 1.5% to 57.8%. The model exhibits stable performance across different prediction horizons, indicating its effectiveness for both short-term and long-term forecasting tasks. The study concludes that combining GCN for spatial modeling and GRU for temporal modeling provides a superior framework for capturing spatio-temporal correlations in traffic data. This approach offers a robust solution for traffic forecasting that can be applied to other spatio-temporal prediction tasks involving graph-structured data.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• anticipation