View publication
| Title | Spatial Transfer Learning for Urban Prediction using H3 Indexing and POI-based Representations |
| Authors | Pedro Olivares, Jonathan Frez, Nelson Baloian |
| Publication date | 2026 |
| Abstract | Machine learning is increasingly used to support urban planning, yet predictive models often fail in data-scarce or newly developed areas--precisely where accurate forecasts are most critical. This paper presents a general spatial transfer learning framework that addresses this gap by transferring knowledge across functionally similar urban zones. Urban structure is encoded through H3 hexagonal indexing and Point-of-Interest (POI)-based embeddings derived from open mapping data, enabling semantic similarity to guide model transfer. Models trained in data-rich zones are adapted to data-limited targets without task-specific architectures, supporting diverse applications such as real estate valuation, mobility prediction, and environmental assessment. Extensive city-scale case studies demonstrate that semantic proximity in POI-H3 space strongly predicts transfer success, consistently outperforming non-transfer baselines under sparse data conditions. We also identify operational regimes that guide neighborhood size, source selection, and transfer thresholds. The framework provides a scalable and principled approach to urban prediction, offering actionable insights for smart city analytics where data are limited. |
| Pages | 62-67 |
| Conference name | IEEE International Conference on Computer Supported Collaborative Work in Design |
| Publisher | IEEE Computer Society Press (Los Alamitos, CA, USA) |
| Reference URL |
|