Non-end-to-end adaptive graph learning for multi-scale temporal traffic flow prediction.

Area of Science:

• Transportation Science
• Artificial Intelligence
• Network Analysis

Background:

• Accurate traffic flow prediction is crucial for intelligent transportation systems (ITS).
• Existing prediction models struggle with multi-scale temporal interactions and fixed spatial graph structures.
• End-to-end training in current methods can lead to suboptimal parameter optimization and limited performance gains.

Purpose of the Study:

• To develop an advanced non-end-to-end adaptive graph learning algorithm for traffic flow prediction.
• To overcome limitations in capturing complex temporal dependencies and spatial relationships in road networks.
• To improve the predictive accuracy and overall performance of traffic flow forecasting models.

Main Methods:

• Proposed a non-end-to-end adaptive graph learning algorithm incorporating multi-scale temporal attention and convolution modules.
• Introduced a novel graph learning module for adaptive node correlation discovery during training.
• Implemented alternate optimization of prediction and graph learning modules with dynamic graph structure updates.

Main Results:

• The proposed method significantly enhanced prediction accuracy on the PeMSD4 and PeMSD8 datasets.
• Ablation studies confirmed the effectiveness of individual modules (multi-scale temporal extraction, adaptive graph learning).
• Visualizations validated the rationality of the dynamically generated graph structures.

Conclusions:

• The adaptive graph learning algorithm effectively captures complex dependencies for improved traffic flow prediction.
• The non-end-to-end approach with alternate optimization enhances model performance.
• The dynamic graph structure learning is crucial for modeling intricate spatial relationships in road networks.