Transfer Learning for Human Mobility Modeling in New Cities

Citizen mobility trajectories are crucial for urban planning, transportation, and commercial analysis, yet collecting large‑scale data for a new city is difficult. At WWW 2020, JD Urban released a paper titled “What is the Human Mobility in a New City: Transfer Mobility Knowledge Across Cities,” which investigates how to infer a city’s mobility distribution using transfer learning based on POI, road‑network, and traffic information.

1. Mobility Intention Generation The authors observe that explicit travel patterns differ across cities (e.g., short home‑to‑metro trips in Beijing vs. their absence in smaller cities). They therefore learn a shared latent “Mobility Intention Space” via domain generalization: spatial context features (POI distribution, topological OD features, distance to transit stations) are extracted from source cities, a mapping function G is trained to minimize maximum mean discrepancy between cities, and a generative model is built on the aligned intention distribution.

2. Target‑City Origin‑Destination (OD) Generation Using the learned intention space, the method treats OD generation as a similarity search. All candidate OD pairs within 6 km are enumerated, their spatial features are mapped through G , and the most similar candidate to a generated intention vector is selected via a KD‑Tree index.

3. Route Generation For each generated OD pair, a set of candidate routes is created by enumerating the top‑m shortest non‑overlapping paths. Overlap is measured with the wJCD index and filtered by a threshold θ. A listwise ranking network Gu (a multi‑layer perceptron) scores each candidate; scores are converted to probabilities with Softmax and trained using cross‑entropy against real trajectory distributions.

4. Experiments The approach is evaluated on four regions—Beijing Chaoyang, Beijing Haidian, Chengdu, and Hefei—covering first‑, second‑, and third‑tier cities. OD distribution accuracy is measured with nMMD, and route preference accuracy with KL divergence. Results show that domain‑generalized models significantly reduce MMD and improve route selection, with m = 5 and θ = 0.7 offering the best trade‑off between coverage (~90 % of real trajectories) and computational cost.

5. Case Study A deployment on Xiong’an Rongcheng, using models trained on Beijing and Chengdu, generates trajectories that closely match on‑site observations, confirming the method’s practical applicability.