IRN2Vec: A representation learning model for road network intersections by integrating geospatial attributes and travel behaviors

Abstract

The structural characterization of road networks serves as a critical foundation for enabling high performance in intelligent transportation systems. This paper proposes IRN2Vec, an intersection-oriented representation learning model that generates discriminative road intersection embeddings by integrating geospatial attributes, semantic homogeneity, and mobility behavior features through the LEIRN framework. The model employs a shortest-path sampling strategy to construct training data and adopts a multi-task learning approach to jointly optimize three types of relationships: geographical proximity, label consistency, and categorical similarity. Experiments conducted on real-world road network data from San Francisco, Porto, and Tokyo demonstrate that IRN2Vec achieves average improvements in F1-Score of 31.6%/25.1%, 16.2%/8.6%, and 27.8%/20.2% over UID, GCN, and GAT models, respectively, in traffic signal classification and pedestrian crossing classification tasks. In travel time estimation, it reduces the mean absolute error (MAE) by 12.2%–24.6%. The findings provide effective feature support for traffic state perception and road network optimization.