FreqMamba: Frequency-aware multi-graph fusion with Mamba for traffic flow prediction

Accurate traffic flow prediction is fundamental to intelligent transportation systems, yet existing spatiotemporal graph neural networks face three critical limitations: inadequate exploitation of multi-scale temporal patterns, suboptimal graph construction that fails to capture heterogeneous spatial dependencies, and computational inefficiency in long-sequence modeling. To address these challenges, we propose FreqMamba, a frequency-aware multi-graph fusion framework integrated with Mamba-based temporal modeling for robust traffic flow forecasting. Specifically, our approach introduces three key innovations: First, the discrete wavelet transform explicitly decomposes traffic signals into low-frequency trend components and high-frequency event components, enabling decoupled learning of persistent patterns and transient anomalies. Second, we construct specialized event graphs, trend graphs, and critical-node graphs that capture pattern-specific spatial relationships, with attention-based fusion mechanisms adaptively integrating complementary information across graph structures. Third, we incorporate a selective state space model to achieve efficient long-range temporal dependency modeling with linear computational complexity, overcoming the quadratic bottleneck of Transformer-based approaches. Extensive experiments on four real-world datasets demonstrate that FreqMamba consistently outperforms 20 state-of-the-art baselines, achieving average improvements of 0.65%–5.33% in MAE while reducing training time by up to 12.40% and memory consumption by up to 17.59% compared to competitive methods. Moreover, generalization experiments on four additional datasets demonstrate its consistent effectiveness. Ablation studies, efficiency analyses, and case analyses further validate the effectiveness of each architectural component, confirming FreqMamba's ability to simultaneously achieve high prediction accuracy, computational efficiency, and robustness across diverse traffic scenarios.