A Multi‐View Spatio‐Temporal Learning Model for the Co‐Prediction of Multimodal Origin–Destination Travel Demand

ABSTRACT

Accurately predicting multimodal travel demand at the origin–destination (OD) dyadic level is essential for intelligent transportation systems and sustainable urban planning while related modeling capability is still limited. This study proposes a novel multi‐view spatiotemporal graph neural network, CoM‐STGNN, which can co‐predict OD flows for multiple transportation modes through intra‐ and inter‐modal learning modules. A flow dual graph is innovatively employed to represent OD flows as nodes and its spatio‐temporal relationships as edges. Three spatio‐temporal complementary relationships of OD flows are considered and modeled via multi‐view graphs: spatial adjacency, mobility features, and the semantic similarity with encoding of built‐environment features and multimodal flow series. The CoM‐STGNN model further incorporates graph convolutional networks (GCNs) and temporal convolutional networks (TCNs) to extract intra‐modal flow features. In addition, a self‐attention mechanism is introduced to summarize inter‐modal interactions and fuse complementary patterns across transport modes to generate the final prediction. Empirical results in Wuhan and New York show that CoM‐STGNN outperforms baselines across multiple evaluation metrics, including RMSE, MAE, and SMAPE. Ablation studies further validate the effectiveness of the environment‐flow embedding and the attention‐based multimodal learning approach.