Fusion Mamba with Mixed Graph Learner for Long-term Traffic Flow Prediction

Long-term prediction as a vital task of multivariate traffic flow time series remains challenging due to spatial-temporal heterogeneity and non-stationarity. To address this, we propose Fusion Mamba with Mixed Graph Learner (FMMGL), a novel framework that integrates the graph learners with a cross-fusion Mamba. Specifically, the Mixed Graph Learner reconstructs the traffic data by memorizing typical features in input sequence to explicitly disentangles the heterogeneity in space, and Fusion Mamba further enlarges the receptive field to be robust to temporal dependencies from normal to non-stationarity. Together, these components enable FMMGL to effectively tackle the any traffic scenarios, delivering improved performance with optimized computational efficiency. Comprehensive experiments on four real-world traffic datasets (PeMS03, PeMS04, PeMS07 and PeMS08) demonstrate the superiority of FMMGL outperforms several state-of-the-art methods.