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

Area of Science:

• Intelligent transportation systems
• Graph representation learning
• Geospatial data analysis

Background:

• Effective road network characterization is crucial for intelligent transportation systems.
• Existing methods lack comprehensive feature integration for road intersection representation.

Purpose of the Study:

• To propose IRN2Vec, an intersection-oriented representation learning model.
• To generate discriminative road intersection embeddings by integrating diverse features.
• To enhance performance in traffic analysis and road network optimization tasks.

Main Methods:

• Developed the LEIRN framework for integrating geospatial, semantic, and mobility features.
• Employed a shortest-path sampling strategy for training data construction.
• Utilized a multi-task learning approach to optimize geographical proximity, label consistency, and categorical similarity.

Main Results:

• IRN2Vec demonstrated significant improvements in F1-Score for traffic signal and pedestrian crossing classification tasks compared to UID, GCN, and GAT models.
• Achieved substantial reductions in Mean Absolute Error (MAE) for travel time estimation.
• Showcased superior performance across real-world datasets from San Francisco, Porto, and Tokyo.

Conclusions:

• IRN2Vec provides effective feature support for advanced traffic state perception.
• The model contributes to optimizing road network functionalities within intelligent transportation systems.
• The proposed approach offers a robust method for learning rich intersection representations.