Discovering Key Sub-Trajectories to Explain Traffic Prediction.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Traffic flow prediction is crucial but challenging to achieve with interpretable and efficient models.
• Existing Shapley methods offer interpretability but face issues with traffic data complexity and computational cost.
• Direct application of Shapley values in traffic prediction is hindered by understanding correlations and NP-hard problem complexities.

Purpose of the Study:

• To develop an interpretable and efficient unified model for traffic flow prediction.
• To overcome the limitations of traditional Shapley methods in handling complex trajectory data.
• To provide a method for understanding the underlying mechanisms of deep learning models in traffic prediction.

Main Methods:

• Proposed Trajectory Shapley, an approximate Shapley approach for flow tensor decomposition.
• Decomposed input into trajectories and calculated Shapley values for specific regions.
• Developed a feature-based submodular algorithm to summarize representative Shapley patterns for interpretability.

Main Results:

• The algorithm successfully identified multiple traffic trends from arterial roads using Shapley distributions.
• Demonstrated improved interpretability and efficiency compared to existing explainable baseline models.
• Validated on real-world taxi trajectory datasets, showing robust performance.

Conclusions:

• Trajectory Shapley offers a novel and effective approach to interpretable traffic flow prediction.
• The feature-based summarization enhances the understanding of deep model mechanisms in traffic analysis.
• This method provides valuable insights for urban traffic management and planning.