Heterogeneous fusion graph neural network for traffic prediction

Accurate urban traffic management relies on precise traffic predictions. Spatial-temporal graph neural networks, combining graph neural networks with time series processing, have gained popularity for traffic prediction. However, traditional graph neural networks fall short in modeling high-order interactions among multiple nodes. To address these challenges, we present a novel approach, the Heterogeneous Fusion Graph Neural Network (HFGNN). In our model, we integrate hypergraph convolution which can model higher-order relationships between nodes with graph convolution to capture more complex spatial dependencies in traffic network structure. Experiments on real-world traffic data confirm the superiority of our model over state-of-the-art methods.