Light attention-based neural networks for traffic flow prediction

Abstract

Spatial–temporal traffic patterns in transportation significantly influence the design of prediction models, which require both high accuracy and computational efficiency. This paper introduces the Light Attention-based Spatial-Temporal Neural Networks (Light-ASTNN), a lightweight traffic prediction model designed for higher prediction accuracy. The model integrates network topology information from a transportation network into a spatial attention to enhance the attention mechanism’s capacity. The effectiveness of the proposed model is validated through comparable experiments with a previous model, using 5 real-world traffic graph network-based datasets. The experimental results show that the proposed model can achieve a better performance in both the accuracy and computational efficiency, despite the fewer parameters. Furthermore, the experiments further highlight the critical role of network topology information in computing spatial correlations using the attention mechanism.