STAGformer Boosts Micro-Mobility Demand Forecasting with Agent Graph Transformer.

Accurate demand forecasting for micro-mobility services, such as bike-sharing systems, is crucial for efficient operations but is challenging due to complex spatio-temporal dependencies across large urban networks. Traditional self-attention mechanisms, while powerful, incur quadratic computational costs, making them impractical for large-scale applications. This paper introduces STAGformer, a Spatio-Temporal Agent Graph Transformer designed to overcome these limitations. STAGformer employs a unique two-step agent attention mechanism. A small set of learnable spatial and temporal agent tokens first aggregates global information from the network and then broadcasts it back to individual stations and time steps. This innovative approach effectively captures long-range interactions while reducing computational complexity from quadratic to linear. The model integrates several core modules, including a spatio-temporal encoder, a graph propagation module, and a temporal convolution module, all working in concert with the agent attention. Extensive experiments on real-world datasets (NYC Citi-Bike and Chicago Divvy-Bike) demonstrate that STAGformer consistently outperforms state-of-the-art baselines, achieving significant improvements in forecasting accuracy across various prediction horizons.