Reproduce “UrbanMind: Urban Dynamics Prediction with Multifaceted Spatial-Temporal Large Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yuhang Liu, Yingxue Zhang, Xin Zhang, Ling Tian, Yanhua Li, Jun Luo
UrbanMind, a framework developed at the State University of New York at Binghamton and collaborators, leverages multifaceted spatial-temporal Large Language Models for urban dynamics prediction, including traffic speed, inflow, and travel demand. It consistently achieves lower MAE and RMSE than prior baselines in both standard and challenging zero-shot prediction tasks, notably demonstrating an 8.5% lower MAE and 9.9% lower RMSE compared to UrbanGPT in cross-city generalization.
Key insights
- Specialized masking strategies within Muffin-MAE effectively capture both inter-correlations among diverse urban dynamics and specific spatio-temporal dependencies, bridging the gap between raw data and LLMs.
- Semantic-aware prompt engineering combined with a partially frozen attention mechanism allows LLMs to adapt to spatio-temporal patterns while preserving general language understanding.
- A test-time adaptation mechanism using a shared-layer reconstructor provides a method to dynamically mitigate distributional shifts, which is crucial for robust generalization in dynamic urban environments, especially in zero-shot scenarios.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.