HyperAOE: Generalizing Zero-Shot Cross-City Trajectory Prediction with Hypernetworks

Why It Matters

Zero‑shot mobility forecasting eliminates the costly data‑collection barrier for emerging cities, accelerating smart‑city initiatives and transportation planning.

Key Takeaways

• HyperAOE enables zero‑shot trajectory prediction for data‑scarce cities
• Architecture separates spatial meta‑data, city encoder, seq2seq backbone
• Supports diverse meta‑data sources and encoder designs
• Outperforms prior models in extensive cross‑city benchmarks
• Open‑source code released for reproducibility and extension

Pulse Analysis

Urban mobility forecasting has traditionally hinged on massive, city‑specific trajectory datasets, a requirement that stalls deployment in regions lacking comprehensive sensor coverage. Conventional transfer‑learning methods mitigate this gap by fine‑tuning models with limited local data, yet they still depend on at least a modest sample of trips to achieve acceptable accuracy. HyperAOE disrupts this paradigm by leveraging hypernetworks that generate model parameters directly from publicly available spatial meta‑data—such as points of interest, road networks, and population density—thereby sidestepping the need for any ground‑truth trajectories in the target city.

The core of HyperAOE’s flexibility lies in its modular three‑stage design. First, spatial meta‑data are ingested and encoded into rich location embeddings via a city encoder, which can be any architecture ranging from graph neural networks to simple multilayer perceptrons. Second, these embeddings condition a hypernetwork that produces the weights for a downstream sequence‑to‑sequence backbone, responsible for capturing the temporal dynamics of human movement. This separation permits researchers to experiment with novel encoders or backbone models without redesigning the entire system, fostering rapid iteration and cross‑disciplinary collaboration. Empirical results across diverse metropolitan areas demonstrate that HyperAOE not only matches but often exceeds the performance of models that rely on extensive fine‑tuning, highlighting its strong zero‑shot transfer capability.

For industry stakeholders, HyperAOE opens a pathway to deploy predictive mobility services—such as demand‑responsive transit, ride‑hailing fleet optimization, and urban infrastructure planning—in markets previously deemed too data‑poor to justify investment. The open‑source release accelerates adoption, allowing city planners and mobility providers to integrate the framework into existing pipelines and tailor it with locally relevant meta‑data. As smart‑city ecosystems evolve, tools like HyperAOE that democratize high‑quality forecasting will become pivotal in shaping equitable, efficient, and data‑light transportation solutions.