Spatio-Temporal Graph Convolutional Networks: A Deep Learning Framework for Traffic Forecasting

Yu, Bing; Yin, Haoteng; Zhu, Zhanxing · 2018 · OpenAlex-citations

DOI: 10.24963/ijcai.2018/505

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 mid- and long-term traffic forecasting, a critical component of intelligent transportation systems. Traditional statistical methods, such as linear regression and ARIMA, often fail to capture the high nonlinearity and complex spatio-temporal dependencies of traffic flow, particularly over longer prediction horizons. While recent deep learning approaches have improved accuracy, many rely on recurrent neural networks (RNNs) or standard convolutions that are computationally expensive, prone to error accumulation, or limited to grid-structured data. To overcome these limitations, the authors propose Spatio-Temporal Graph Convolutional Networks (STGCN), a novel deep learning framework designed to model traffic networks as graphs and extract spatio-temporal features using a fully convolutional architecture. The STGCN framework formulates traffic prediction as a time-series problem on graph-structured data, where nodes represent sensor stations and edges represent connectivity. The model consists of stacked spatio-temporal convolutional blocks, each employing a "sandwich" structure: two temporal gated convolution layers surround a central spatial graph convolution layer. The spatial component utilizes graph convolutions based on spectral theory, employing either Chebyshev polynomial approximation or a first-order approximation to efficiently capture spatial dependencies without the high computational cost of standard spectral methods. The temporal component replaces recurrent units with 1-D causal convolutions and gated linear units (GLU), enabling parallel training and faster convergence. This design allows the model to jointly process spatial topology and temporal dynamics while maintaining parameter efficiency. The proposed model was evaluated on two real-world datasets: BJER4, comprising traffic data from Beijing’s East Ring No. 4 route, and PeMSD7, derived from California’s Performance Measurement System with both medium (228 stations) and large (1,026 stations) network scales. Experiments compared STGCN against baselines including Historical Average, Linear Support Vector Regression, ARIMA, Feed-Forward Neural Networks, Fully-Connected LSTM, and Graph Convolutional GRU. Performance was measured using Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and Root Mean Squared Error (RMSE) for prediction horizons of 15, 30, and 45 minutes. Results demonstrated that STGCN consistently outperformed all baselines across both datasets and all metrics. The advantage was particularly pronounced on the more complex PeMSD7 dataset, highlighting the model's ability to leverage spatial topology for improved accuracy. The significance of this work lies in demonstrating that purely convolutional structures can effectively replace recurrent networks for spatio-temporal forecasting, offering faster training speeds and fewer parameters. By explicitly modeling the graph structure of traffic networks, STGCN captures comprehensive spatial correlations that previous methods neglected. The study concludes that this framework is not only superior for traffic prediction but also serves as a universal approach for processing structured time series in other domains requiring simultaneous spatial and temporal feature extraction.

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