Transfer-Mamba: Selective state space models with spatio-temporal knowledge transfer for few-shot traffic prediction across cities

Abstract

Spatio-temporal traffic forecasting significantly impacts the development of smart cities. Owing to uneven development levels and the substantial costs of data collection, many cities often encounter the challenge of limited data availability when undertaking traffic prediction tasks. This paper introduces Transfer-Mamba, a method that employs Spatio-temporal selective state space models with transfer learning for few-shot traffic prediction across multiple cities. Transfer-Mamba features a Spatio-temporal graph pre-training process, which incorporates an encoder–decoder architecture with a Mamba module and an adaptive graph convolutional network. This process enhances the feature representation and transferability of traffic data from multiple source cities by capturing Spatio-temporal correlations. To distinguish unique traffic patterns and data distributions in source cities, an unsupervised learning algorithm groups similar traffic characteristics during the Spatio-temporal knowledge clustering phase. These clustered patterns are then retrieved through a Spatio-temporal knowledge querying process, which extracts traffic meta-knowledge to guide the few-shot prediction phase for the target city. Extensive experiments conducted on four real-world traffic datasets demonstrate that Transfer-Mamba outperforms the existing mainstream baselines, which provides valuable insight for optimizing road traffic management.