MambaTraffic: Contrastive Traffic Flow Prediction With Spatial–Temporal State Space Models

Traffic flow prediction is crucial for traffic management and travel optimization but remains challenging due to the spatial–temporal variability of traffic patterns. Existing methods, often based on convolutional networks or transformers, struggle to balance global feature extraction with computational efficiency and typically rely on a single learning paradigm. To address these issues, we propose MambaTraffic, a novel approach that integrates the state space model with contrastive learning. MambaTraffic employs a two-branch spatial–temporal network, combining Mamba and a spatial–temporal convolutional network, with a contrastive learning paradigm to enhance feature extraction. Extensive experiments on publicly available datasets demonstrate significant improvements, with mean absolute percentage errors for inflow and outflow reduced by 6.4% and 9.7%, respectively, on the NYCTaxi dataset, setting a new state-of-the-art benchmark.