Reproduce “Learning to Transfer Across Modes: Towards Unified Urban Mobility Forecasting” 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
Yixuan Zhao, Man Luo
TransMod is a cross-modal urban mobility forecasting framework that transfers knowledge from data-rich bike-sharing and metro systems to zone-level ride-hailing prediction. By combining geography-constrained station-to-zone mapping, cross-modal representation alignment, and a memory of reusable temporal patterns, it achieves MAEs of 10.78 in New York and 8.30 in Chicago, outperforming the strongest reported baseline by about 13% and 21%, respectively.
Key insights
- Soft spatial assignment is central to cross-modal transfer: replacing it with hard nearest-zone mapping increases MAE by 13.0% in New York and 16.0% in Chicago.
- The memory module provides reusable temporal priors that complement conventional recurrent decoders; replacing it with a GRU raises MAE by 8.8% and 11.8% in the two cities.
- Transfer is most useful when target data are scarce: pre-training in New York and adapting with 10% of Chicago data reduces RMSE from 16.21 to 12.38, while the method is also less sensitive to coarser spatial grids and longer forecasting horizons.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.