Fourier dynamic hypergraph interaction network for traffic flow forecasting

Traffic flow forecasting is essential for intelligent transportation systems, yet it remains challenging to jointly model high-order spatial dependencies, periodic temporal patterns, and short-term fluctuations. This paper propos-es a Fourier Dynamic Hypergraph Interaction Network (FDHGIN) for traf-fic flow forecasting. In the spatial dimension, FDHGIN learns dynamic node–hyperedge associations to capture many-to-many traffic interactions beyond pairwise road connections. In the temporal dimension, learnable Fourier temporal embeddings (FTE) are combined with causal temporal convolutions (CTConv) to encode periodic patterns and local fluctuations while respecting temporal causality constraints. An Interactive Dynamic Hypergraph Convolution (IDHC) module further splits the temporal repre-sentation into odd- and even-indexed subsequences and performs bidirec-tional cross-branch modulation and residual refinement under dynamic hypergraph guidance. Experiments on six real-world datasets, covering eight forecasting tasks, demonstrate that FDHGIN achieves competitive forecasting performance, particularly in terms of mean absolute error (MAE) and root-mean-square error (RMSE). Ablation studies reveal that all key components contribute to prediction accuracy. Further sensitivity and efficiency analyses verify the model’s robustness to hyperparameters and its desirable balance between accuracy and computational complexity.