Mamba4Traffic: A Mamba-Based Model for Traffic Prediction on Multiple Channels

Intelligent transportation systems require accurate multivariate traffic forecasting to enable real-time vehicle coordination and dynamic traffic management in connected vehicular environments. Traffic data consists of multiple interrelated variables, including speed, flow, and occupancy, yet most deeplearning approaches either predict only one variable or require separate models for each, preventing unified modeling of all traffic features. To address this, we introduce Mamba4Traffic, a forecasting framework built on the Mamba selective statespace architecture that natively handles multivariate inputs and simultaneously models all traffic variables. The proposed model employs a dual-path structure with separate temporal and spatial processing streams. The temporal path combines Mamba blocks with Long Short-Term Memory (LSTM) layers to capture both long-range dependencies and short-term temporal dynamics. Additionally, Mamba4Traffic uses Feature-wise Linear Modulation (FiLM) in its embedding strategy to integrate learned and explicit embedding vectors, enabling dynamic feature representation without increasing model complexity. To encourage more comprehensive evaluation of multivariate traffic data, we introduce vehicular variants of standard traffic prediction evaluation metrics that utilize normalization and averaging to provide a single metric that reflects model performance across all traffic variables, unaffected by variable ranges or distributions. Our results demonstrate that Mamba4Traffic achieves strong predictive performance with efficient parameter usage compared to baseline models.