STAGformer: A Spatio-temporal Agent Graph Transformer for Micro Mobility Demand Forecasting

Short summary

STAGformer is a new Spatio-Temporal Agent Graph Transformer for station-level bike-sharing demand forecasting that reduces standard self-attention complexity from quadratic to linear O(NT) via a two-step agent attention mechanism using learnable spatial and temporal tokens. The model integrates spatio-temporal encoding, graph propagation, temporal convolution, and agent attention, validated on NYC Citi-Bike and Chicago Divvy-Bike datasets where it outperforms state-of-the-art baselines in RMSE and MAE. Ablation studies confirm that the agent attention module is the most critical component for capturing global spatio-temporal dependencies.

• •Introduces STAGformer, a graph transformer with linear-complexity agent attention for bike-sharing demand forecasting
• •Outperforms SOTA baselines on NYC Citi-Bike and Chicago Divvy-Bike datasets across multiple prediction horizons
• •Ablation studies show agent attention mechanism is critical for modeling long-range spatio-temporal dependencies

Generated with AI, which can make mistakes.