TL;DR
BIGCity is a pioneering foundation model for urban spatiotemporal (ST) data that unifies individual-level trajectories and population-level traffic states. By representing city dynamics through a single, prompt-based architecture, it achieves state-of-the-art performance across eight diverse urban analysis tasks.
Introduction: Bridging the Urban Data Divide
Urban spatiotemporal data analysis has traditionally been divided into two distinct silos: individual-level models (trajectories) and population-level models (traffic states). Trajectory models focus on predicting the movement of single agents, while traffic state models predict aggregate metrics like flow or speed across a road network.
In reality, these modalities are deeply interconnected: macro-level congestion arises from micro-level individual decisions, and those same conditions dictate future individual routes. However, existing AI models are typically "solo-modality," requiring specialized architectures for each task. This lack of unification leads to fragmented insights and high development overhead. BIGCity addresses this by creating a universal representation for all urban movement data.
Methodology: The BIGCity Architecture
BIGCity bridges the gap between trajectories and traffic states through three core innovations:
- The ST-unit: To unify disparate data types, the authors define a basic spatiotemporal unit—a triple consisting of a road segment, its dynamic traffic state, and a timestamp. This allows both trajectories (sequences of segments) and traffic states (sequences of conditions) to be expressed in a shared mathematical format.
- Spatiotemporal Tokenizer: This module converts raw ST-units into high-dimensional tokens. It utilizes Graph Attention Networks (GATs) to capture the spatial layout of the road network and MLPs to integrate temporal features.
- VMTP (Versatile Model with Task-oriented Prompt): BIGCity employs a GPT-2 backbone enhanced with Low-Rank Adaptation (LoRA). Instead of structural changes for different tasks, it uses textual prompts and placeholders (
[CLAS]for classification,[REG]for regression) to guide the model's output.
The model undergoes a two-stage training process: a self-supervised Masked Reconstruction phase to learn the underlying physics of city movement, followed by Multi-Task Prompt Tuning where the model is co-trained on specific downstream tasks simultaneously.
Evaluation and Key Results
The model was evaluated using real-world datasets from Beijing (taxi trajectories), Xi’an, and Chengdu (online car-hailing), with road networks extracted from OpenStreetMap.
BIGCity was tested against 18 specialized baselines across 8 distinct tasks. The results demonstrate a clear "Modality Synergy": by training on both trajectories and traffic states, the model learns a more robust representation of urban dynamics than specialized models.
Key performance highlights include:
- Trajectory Tasks: BIGCity outperformed SOTA models like START and JGRM in Next-Hop Prediction and Travel Time Estimation (TTE).
- Traffic State Tasks: For multi-step traffic prediction, it achieved an average improvement of ~2.83% over established models like GWNET and ST-Norm.
- Generalization: A model pre-trained on Beijing data transferred effectively to Xi'an and Chengdu, often outperforming baselines trained specifically on those target cities.
Conclusion and Future Directions
BIGCity successfully applies the prompting paradigm from NLP to physical-world urban data, handling diverse output types—from discrete road IDs to continuous time values—within a single framework. While currently focused on road segments, future iterations aim to incorporate heterogeneous spatial elements like Points of Interest (POIs) and grid-based population heatmaps to create a truly universal model for urban science.
Summary updated