CrossBag: A Bag of Tricks for Cross-City Mobility Prediction

Access to large-scale human trajectory data has significantly advanced the understanding of human mobility (HuMob) behavior for urban planning. However, these data are often concentrated in major cities, leaving smaller or less-monitored areas with limited information, undermining the performance of data-hungry machine learning models for HuMob prediction. This imbalance poses a challenge for cross-city mobility prediction, as many existing models are designed for single-city settings. To address this, we present CrossBag, a set of simple yet effective techniques to boost cross-city prediction. These techniques include context-aware spatiotemporal embeddings, masking types, and a progressive knowledge transfer method to incrementally adapt the target model while preserving useful patterns from the source model for stable cross-city transfer. Additionally, we propose a test-time trajectory refinement method using top-K guided beam search to prevent predictors from getting stuck in repetitive location predictions. We validate CrossBag on the large-scale multi-city dataset from the HuMob Challenge 2024, achieving a top-10 placement out of over 100 participating teams.