Multi-modal traffic flow joint prediction with a multi-source information knowledge graph

Modern transportation systems are multimodal and influenced by diverse external factors, such as weather, POIs, and traffic incidents. Although existing multimodal traffic prediction methods use multi-graph structures to model cross-mode correlations, they often overlook heterogeneous couplings between external factors and different traffic modes. To address this limitation, we propose KG-BTHN, a bidirectional temporal-spatial hypergraph neural network with a traffic knowledge graph for multimodal traffic prediction. The knowledge graph integrates multimodal traffic data with external factors via attribute-enhanced graph representation learning. A gated fusion module combines graph embeddings with raw traffic features, while hypergraph convolution and bidirectional temporal convolution are employed to capture static and dynamic dependencies, respectively. Extensive experiments on a real-world New York City dataset demonstrate that KG-BTHN outperforms state-of-the-art baselines across multiple metrics.