Cross-region zero-shot bike-sharing demand prediction with transitive transfer learning using multi-modal public transit data

Predicting bike-sharing demand in regions without historical data is essential for early-stage fleet allocation and rebalancing. However, conventional supervised forecasting models rely on region-specific historical observations and often generalize poorly across heterogeneous urban contexts. We propose CZBTTL, a cross-region zero-shot framework that uses multi-modal public-transit demand as an intermediate domain for transitive transfer learning. Equal-capacity Constrained K-Means with CORAL distance partitions source-region transit sequences into maximum-difference sub-domains to increase training heterogeneity. A two-stage procedure then combines transit-to-bike mapping pre-training with time-aware adversarial alignment to learn transferable temporal representations. GRU- and Transformer-based implementations were evaluated across six district-level transfer tasks in two scenarios: transfer among similarly developed districts and transfer from core to emerging districts. Across both scenarios, the CZBTTL implementations collectively achieved the lowest RMSE and MAE for every target. Ablation results supported the contributions of maximum-difference partitioning and the two-stage learning strategy. By enabling demand estimation without historical target-region bike-sharing observations, CZBTTL can support early-stage fleet deployment, parking-area planning, and hourly rebalancing in newly served urban areas.