One Size Fits All: A Unified Traffic Predictor for Capturing the Essential Spatial-Temporal Dependency.

Area of Science:

• Urban planning and intelligent transportation systems.
• Data science and machine learning applications in infrastructure.

Background:

• Accurate traffic prediction is crucial for smart city development, impacting traffic management, public safety, and environmental policy.
• Existing traffic predictors are often city-specific, limiting their scalability and applicability.

Purpose of the Study:

• To develop a unified network model for universal traffic prediction across diverse cities.
• To capture essential spatial-temporal dependencies between traffic flow and fine-grained road networks.

Main Methods:

• A joint knowledge- and data-driven mechanism categorizing dependencies into road segment, intra-intersection, and inter-intersection correlations.
• Utilizing meta road segment learners, connection-aware spatial-temporal graph convolutional networks (GCN), and multiscale residual networks.
• Modeling the road network as a graph using fine-grained road topology from real-world trajectory data.

Main Results:

• The proposed model achieves superior prediction accuracy compared to existing approaches.
• Demonstrated transferability: a model trained in one city performs well in another without fine-tuning, highlighting its ability to capture universal traffic-road network relationships.

Conclusions:

• The unified model effectively learns and leverages spatial-temporal dependencies for robust traffic prediction.
• The approach offers a scalable and adaptable solution for smart city traffic management, reducing the need for city-specific model development.