Traffic data imputation via knowledge graph-enhanced generative adversarial network.

Area of Science:

• Intelligent Transportation Systems (ITS)
• Data Science
• Artificial Intelligence

Background:

• Traffic data imputation is vital for ITS, but current deep learning methods falter with high missing data rates and limited external factor integration.
• Challenges include maintaining system reliability and efficiency under incomplete traffic information.

Purpose of the Study:

• To develop a novel deep learning model for accurate traffic data imputation, addressing limitations of existing methods.
• To enhance imputation performance by incorporating external traffic-related knowledge and improving spatiotemporal dependency modeling.

Main Methods:

• Propose a knowledge graph-enhanced generative adversarial network (KG-GAN) for traffic data imputation.
• Construct a fine-grained knowledge graph (KG) to represent external factors like POI and weather.
• Introduce a knowledge-aware embedding cell (EM-cell) to integrate external knowledge with traffic data for spatiotemporal GAN input.

Main Results:

• KG-GAN significantly outperforms state-of-the-art methods in traffic data imputation across various missing data scenarios.
• Ablation studies confirm the substantial performance gains from incorporating external knowledge.
• The model demonstrates improved accuracy and robustness in handling incomplete traffic datasets.

Conclusions:

• The proposed KG-GAN effectively addresses the challenges of traffic data imputation, particularly under high missing data rates.
• Integrating external knowledge through a fine-grained KG significantly enhances imputation quality for ITS applications.
• This approach offers a more robust solution for reliable traffic prediction and management systems.