Activity-aware urban area embedding with contrastive learning for intelligent transportation systems applications

Embedding is a machine learning technique that represents data entities as continuous vector representations, capturing the underlying semantic relationships between them. Urban area embedding applies this concept to urban regions, representing each area as a vector that encapsulates its key characteristics. These embeddings enable models to better understand the relationships between different urban areas, facilitating applications such as traffic management, urban planning, and resource allocation. In this paper, we propose a comprehensive framework called AUAEC (Activity-aware Urban Area Embedding with Contrastive Learning) that integrates diverse open datasets including Location-Based Social Network (LBSN) check-ins, taxi flow data, and Points of Interest (POI) to produce enriched and context-aware region embeddings. To capture both mobility patterns and activity-aware semantics of LBSN users, we apply spatial interpolation based on road network, coupled with activity vector construction to represent user daily activity and movement patterns. To refine these embeddings into comprehensive urban regional representations, the AUAEC incorporates two complementary contrastive learning strategies: View-wise Contrastive Learning, which aligns representations across multiple data views, and Activity-aware Contrastive Learning, which captures inter-region relationships based on activity-aware semantics. The resulting embeddings are evaluated across four critical ITS tasks including land use distribution classification, traffic incident prediction, public transport delay prediction and traffic volume prediction using real-world data. Our approach demonstrates promising results, outperforming state-of-the-art solutions and highlighting the superiority of AUAEC in providing robust, contextual representations of urban areas for ITS and urban planning applications. • H3-cell level interpolation and aggregated LBSN data address spatial and temporal sparsity. • Contrastive learning aligns heterogeneous urban data views, capturing complementary insights. • Activity-aware pattern semantics through triplet learning produce context-aware embeddings. • Utility and robustness demonstrated across land use, transport delay, incident, and volume prediction tasks.