Event-Coupled Network for Traffic Conversion Prediction in Weather Emergencies

Capturing traffic conversion between multi-agent travel patterns during weather emergencies is critical for improving urban transportation services and enhancing travel efficiency for citizens. However, existing studies have not fully explored the peculiarities of severe weather events, such as their rapid onset, temporal inhomogeneity, and data sparsity, which challenges conventional prediction models. In this paper, we introduce an Event-Coupled Field-aware Factorization Network(EC-FFN) that characterizes these accompanying events through triples and employs continuous-valued embeddings to address the data inhomogeneity. A Conv-FFN structure, which combines sparse feature interactions with field-aware factorization mechanisms, is designed to capture both local conversion features and global traffic dependencies. Additionally, we consider semantics and hierarchical associations from a spatial evolutionary perspective, enabling a combinatorial analysis of the endogenous causes of traffic conversion. Experiments on New York taxi and bike-share trip datasets demonstrate the model's effectiveness in predicting traffic conversion under weather emergencies.