Zero-Shot Cross-City Trajectory Prediction Using Hypernetworks

Abstract

City-wide mobility prediction models typically rely on either training with extensive local trajectory data or applying transfer learning from data-rich cities to those with limited data. In both cases, the resulting models are specialized to a specific target city for which they require at least some trajectory data to adapt. Consequently, they cannot generalize to cities unseen during training and are inapplicable where no mobility data exist. In this work, we propose H0xtra, a novel approach that enables zero-shot trajectory prediction in entirely unseen cities. H0xtra leverages a hypernetwork to generate city-specific location embeddings from spatial distributions of points of interest, e.g., restaurants or stores. These embeddings capture city-agnostic location semantics, enabling a transformer to learn universal trajectory patterns across cities. At inference, H0xtra performs zero-shot transfer without requiring any mobility data or retraining. Adaptation requires only points of interest data, which are often publicly available, to generate location embeddings specific to the target city. Trained only on a small set of source cities, H0xtra achieves strong zero-shot generalization. In our experiments, the zero-shot performance achieves an accuracy improvement of 11.3% and an average displacement error reduction of 11.5% on average compared to state-of-the-art non-zero-shot baselines.