The authors propose a Hybrid Time-Varying Graph Neural Network (HTVGNN) to improve traffic flow predictions. Traditional methods based on graph neural networks (GNNs) have limitations in capturing spatial correlations in traffic data. HTVGNN overcomes these drawbacks by using a time-aware multi-attention mechanism and a novel graph learning strategy to learn both static and dynamic spatial associations between traffic nodes. The effectiveness of HTVGNN is demonstrated with four real datasets, showing higher prediction accuracy than the latest space-time graph neural network model.
Publication date: 19 Jan 2024
Project Page: https://github.com/DaibenaoZSTU/HTVGNN
Paper: https://arxiv.org/pdf/2401.10155
