CityRep Benchmark for Urban Representation Learning
A new benchmark called CityRep has been introduced to evaluate urban representations across multiple cities, data modalities, and tasks. It addresses limitations in current evaluations that rely on random splits causing spatial leakage and inflated performance. CityRep uses spatially structured splits and includes a spatial unit-agnostic evaluation framework, a unified evaluation protocol with block-based spatial splits, and an extensible multi-city, multi-task design.
Key facts
- CityRep is a unified benchmark for urban representations.
- It evaluates across data modalities, cities, and tasks.
- Uses spatially structured splits to mitigate spatial leakage.
- Includes a spatial unit-agnostic evaluation framework.
- Has a unified evaluation protocol with block-based spatial splits.
- Supports heterogeneous urban representations through standardized alignment.
- Addresses limitations of current evaluations with random splits.
- Designed for cross-location generalization and fair comparison.
Entities
—