How to Break It Down for Building It Up? Theory-Guided Graph Decomposition Learning for Spatiotemporal Traffic...

Area of Science:

• Artificial Intelligence
• Data Science
• Transportation Engineering

Background:

• Current traffic state prediction models often oversimplify spatial-temporal correlations, leading to suboptimal performance.
• The "decompose, then predict" paradigm shows promise but lacks theoretical grounding for effective data decomposition.
• Human mobility patterns exhibit complex heterogeneity not captured by uniform correlation assumptions.

Purpose of the Study:

• To theoretically analyze data decomposition for traffic prediction, establishing conditions for reduced prediction errors.
• To introduce a novel framework, Theory-guided Graph Decomposition Learning (TGDL), for improved traffic state prediction.
• To enhance the portability and performance of existing graph-based traffic prediction models.

Main Methods:

• Information theory-based analysis to derive the Component Independence Principle for data decomposition.
• Development of the TGDL framework to decompose graph-based multivariate time series data into independent subgraph components.
• Integration and evaluation of TGDL with diverse graph-based traffic prediction models.

Main Results:

• TGDL effectively decomposes traffic data into approximately independent subgraph components.
• The framework significantly improves the predictive performance of various graph-based traffic prediction models.
• Experiments on four public datasets show an average performance enhancement of 19.37%.

Conclusions:

• The Component Independence Principle provides a theoretical basis for effective data decomposition in traffic prediction.
• TGDL offers a robust and portable solution for enhancing traffic state prediction accuracy.
• The proposed method addresses limitations of uniform correlation assumptions in spatiotemporal data analysis.