Predicting Human Mobility Across Cities with CityCondBERT

What is it about?

Many cities have limited data for forecasting where people will go next, making it hard to build reliable mobility prediction models for transportation and urban planning. CityCondBERT tackles this by training a shared Transformer on trajectory sequences from multiple cities and then adapting to each city using lightweight city-conditioned modules (FiLM and residual adapters), rather than training a separate model from scratch for every city. To better match the HuMob Challenge evaluation (GEO-BLEU), we also introduce GEO-BLEU-Sinkhorn, a differentiable sequence-level loss that rewards spatial proximity and n-gram trajectory continuity via entropy-regularized optimal transport. Across four benchmark cities in the SIGSPATIAL GISCUP 2025 HuMob Challenge, our two-stage pretraining + city-wise fine-tuning strategy improves GEO-BLEU by about 25% over per-city BERT baselines.

Featured Image

Photo by Yoav Aziz on Unsplash

Photo by Yoav Aziz on Unsplash

Why is it important?

This work addresses a key practical issue in human mobility prediction: models trained per city often struggle to generalize when data availability differs across cities. CityCondBERT improves cross-city transfer by keeping the Transformer encoder shared while injecting city-specific information through post-encoder FiLM conditioning and residual adapters. In addition, GEO-BLEU-Sinkhorn provides a differentiable surrogate that better aligns training with the GEO-BLEU evaluation metric by capturing both spatial proximity and trajectory-level n-gram continuity using Sinkhorn-based optimal transport. In the HuMob Challenge setting, this combination yields consistent gains, reaching a GEO-BLEU of 0.1516 on the organizers’ evaluation split and about a 25% relative improvement over per-city BERT baselines trained from scratch.

Perspectives