Attention Based Spatial-Temporal Graph Convolutional Networks for Traffic Flow Forecasting

Guo, Shengnan; Lin, Youfang; Feng, Ning; Song, Chao; Wan, Huaiyu · 2019 · OpenAlex-citations

DOI: 10.1609/aaai.v33i01.3301922

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 traffic flow forecasting, a critical component of Intelligent Transportation Systems. Existing methods often fail to capture the complex, nonlinear, and dynamic spatial-temporal correlations inherent in traffic data. Traditional statistical and machine learning models struggle with high-dimensional data and require extensive feature engineering, while many deep learning approaches are limited to grid-based inputs or fail to model dynamic correlations effectively. To overcome these limitations, the authors propose the Attention-based Spatial-Temporal Graph Convolutional Network (ASTGCN), a novel deep learning model designed to process traffic data directly on its original graph structure. The ASTGCN model consists of three independent components that model recent, daily-periodic, and weekly-periodic dependencies, respectively. Each component processes specific historical time segments relevant to the prediction window. The core of each component is a spatial-temporal block containing two modules: a spatial-temporal attention mechanism and a spatial-temporal convolution module. The attention mechanism dynamically captures correlations by applying spatial attention to model varying influences between different locations and temporal attention to model dependencies between different time slices. The convolution module combines graph convolutions, based on spectral graph theory and Chebyshev polynomial approximation, to extract spatial features from the network topology, with standard convolutions to capture temporal features. The outputs of the three components are fused via a fully connected layer to generate the final prediction. Experiments were conducted on two real-world datasets from the Caltrans Performance Measurement System (PeMS). The results demonstrate that ASTGCN outperforms state-of-the-art baselines in traffic flow prediction accuracy. The model’s ability to simultaneously model dynamic spatial and temporal correlations allows it to effectively handle the complex patterns of highway traffic, such as daily routines and weekly variations, without relying on manual feature engineering or grid-based data transformations. The significance of this work lies in its effective integration of graph convolutions and attention mechanisms for spatial-temporal forecasting. By modeling the traffic network as a graph and dynamically adjusting the importance of spatial and temporal neighbors, ASTGCN provides a robust framework for predicting traffic flows. This approach advances the field by offering a method that can directly leverage the topological structure of transportation networks, leading to more accurate and efficient traffic management solutions.

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