Joint prediction and understanding of multimodal traffic flow with a bidirectional temporal dynamic spatial hypergraph neural network model

Traffic flow prediction plays a vital role in urban planning, traffic management and control. Graph convolutional networks have driven substantial advances in forecasting for a single transportation mode, yet they fall short when applied to modern multimodal networks because they do not account for interactions among coexisting modes. Although recent efforts have explored multimodal traffic prediction using multiple graph structures to extract pairwise, local spatial dependencies either within or across modes, these methods tend to lack the flexibility to capture high-order, global spatial correlations among multiple modes or functionally similar areas. In addition, multimodal traffic data often suffer from sparsity and random fluctuations caused by distributional differences, posing persistent challenges to cooperative prediction. To address these issues, this paper introduces a Bidirectional Temporal Dynamic Spatial Hypergraph Neural Network (BiT-DSHGNN). First, we construct a static hypergraph based on clusters of administrative functional areas and apply hypergraph convolution to capture intrinsic global spatial correlations among functionally related regions. We then design dynamic semantic hypergraphs that evolve over time, enabling the model to learn time-varying high-order spatial dependencies across modes through a dedicated dynamic hypergraph neural network module. This facilitates cross-modal information sharing, allowing high-density mode nodes to enrich the contextual representation of low-density nodes. Additionally, we propose a bidirectional temporal feature extraction module, named Bidirectional Temporal Gated Network (BTGN), which combines a Bidirectional Temporal Convolutional Network (BiTCN) with a Bidirectional Gated Recurrent Unit (BiGRU). This module integrates both past and future contextual information, further mitigating the impact of random fluctuations. Extensive experiments conduct on four real-world datasets (NYC-Taxi, NYC-Bike, CHI-Taxi, and CHI-Bike) demonstrate that our model consistently outperforms existing methods, achieving state-of-the-art prediction accuracy.