Where to Go Next: Enhancing Zero-Shot Capability for Cross-City Mobility Prediction
Next location prediction models support applications such as personalized recommendations and demand forecasting. In practice, deploying these models in a new city is expensive, as they often require substantial historical mobility data and costly retraining. A more scalable setting is zero-shot cross-city prediction: train in one city and deploy in another without collecting target-city trajectories. However, existing methods are largely city-bound, where their inputs and outputs are tied to city-specific identifiers, so performance collapses when the city changes. To address this issue, in this paper, we propose ReLoX, a framework for enhancing zero-shot capability for cross-city mobility prediction. ReLoX combines an anchor-centered relative trajectory representation with a local egocentric image that captures nearby spatial relations and semantics. A dual-encoder architecture extracts sequential and visual cues and fuses them via a gated hierarchical module for accurate prediction. For within-city use, ReLoX further adds a lightweight global context branch to handle long-range moves better, while cross-city zero-shot inference uses the local-window predictor without any target-city adaptation. Experiments on large-scale mobility data across three cities show that ReLoX remains competitive within cities and achieves substantial gains under cross-city zero-shot protocols, where city-bound baselines largely fail.