Spatial-Temporal Interactive Graph Neural Network for Traffic Forecasting

Traffic flow forecasting is crucial for efficient traffic management and congestion avoidance. Traditional methods are mainly based on statistical methods, which fail to capture the complex spatial-temporal correlations among various factors that affect traffic flow. In recent years, graph neural networks (GNNs) have emerged as a potent tool for modeling and predicting traffic flows. However, the existing GNNs-based methods have limitations in fully extracting the dynamic spatial and temporal correlation of traffic flow, thereby reducing their efficacy in capturing the implicit interactive spatial-temporal relationship. In this paper, we propose an spatial-temporal interactive GNN (STIGNN) method to extract both spatial and temporal correlations and model their interactive relationship. Specifically, we employ an interactive dynamic graph convolutional network (IDGCN) to capture the dynamic spatial correlations and introduce an interactive temporal convolutional network (ITCN) to expand the field of perception in the time dimension. Extensive experiments on two real-world traffic flow datasets demonstrate that the proposed STIGNN outperforms the current state-of-the-art baselines for traffic forecasting.