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

2026-07-07 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, quick

Summary

STAGformer, a Spatio-Temporal Agent Graph Transformer, addresses the challenge of accurate station-level demand forecasting in bike-sharing systems, which is complicated by complex spatio-temporal dependencies and large urban networks. This model achieves efficient global modeling with linear computational complexity, specifically O(NT), by introducing a two-step agent attention mechanism. This mechanism uses a small set of learnable spatial and temporal agent tokens to first aggregate global information and then broadcast it back to individual stations and time steps, effectively capturing long-range interactions while reducing the quadratic cost of standard self-attention. STAGformer integrates four core modules: a spatio-temporal encoder, a graph propagation module, a temporal convolution module, and the agent attention module. Extensive experiments on NYC Citi-Bike and Chicago Divvy-Bike datasets demonstrate consistent outperformance against state-of-the-art baselines, showing significant improvements in RMSE and MAE across multiple prediction horizons. Ablation studies confirm the agent attention mechanism's critical role in modeling global spatio-temporal dependencies.

Key takeaway

For Machine Learning Engineers developing demand forecasting models for micro-mobility or other large-scale urban networks, STAGformer presents a significant advancement. You should consider adopting its agent attention mechanism to efficiently capture long-range spatio-temporal dependencies, reducing computational complexity from quadratic to O(NT). This approach, validated on real-world bike-sharing data, can lead to more accurate predictions (improved RMSE and MAE) by effectively integrating dynamic features and external contextual factors.

Key insights

STAGformer uses agent attention to efficiently model global spatio-temporal dependencies for micro-mobility demand forecasting.

Principles

• Agent attention reduces quadratic complexity to O(NT).
• Global information aggregation improves forecasting accuracy.
• Fusing dynamic features with external factors enhances models.

Method

STAGformer employs a two-step agent attention: tokens aggregate global data, then broadcast to stations/time steps. It integrates spatio-temporal encoding, graph propagation, and temporal convolution.

In practice

• Implement agent attention for large-scale graph problems.
• Integrate weather, time, POI data into demand models.
• Apply STAGformer architecture to other urban network predictions.

Topics

• Spatio-temporal Graph Transformer
• Agent Attention
• Micro Mobility
• Demand Forecasting
• Bike-sharing Systems
• Computational Complexity

Best for: Research Scientist, AI Scientist, Machine Learning Engineer

Related on AIssential

• A Comparative Study of Graph Neural Network Layer Selection for Interaction Modelling in Driving Trajectory Prediction — ARMA, Chebyshev, and topology-aware GNN layers, combined with sum-based aggregation and multi-head attention, enhance driving trajectory prediction accuracy.
• Learning Long Range Spatio-Temporal Representations over Continuous Time Dynamic Graphs with State Space Models — CTDG-SSM unifies state-space modeling with graph topology to capture long-range spatio-temporal dependencies in continuous-time dynamic graphs.
• ASTEROID: A Spatiotemporal Information Transformer for Forecasting Multi-Step Time Series of Molecular Dynamics — ASTEROID uses a Transformer with STI Transformation to forecast molecular dynamics, improving accuracy and reducing computational cost.
• TopoBrick: Agentic Topology Sampling of Exogenous Variables for Zero-Shot Building IoT Forecasting — TopoBrick uses agentic topology sampling of knowledge graphs for zero-shot building IoT forecasting, outperforming baselines.
• Hybrid AI model combines graphs and transformers for real-time traffic forecasts — Urban congestion is a critical global issue, leading to commuter delays, economic inefficiencies, and tragically, a million deaths annually worldwide.
• Fusing Urban Structure and Semantics: A Conditional Diffusion Model for Cross-City OD Matrix Generation — Fusing urban semantics and spatial constraints via a conditional diffusion model enhances cross-city OD matrix generation accuracy and generalizability.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.