Pretraining-improved Spatiotemporal graph network for the generalization performance enhancement of traffic...

Area of Science:

• * Artificial Intelligence
• * Smart City Development
• * Traffic Engineering

Background:

• * Existing traffic prediction models, often based on Graph Convolutional Networks (GCNs), struggle with long-term spatiotemporal dependencies and high computational costs.
• * Retraining these models for new datasets reduces accuracy and increases time investment.
• * Sophisticated modules improve performance but escalate computational demands.

Purpose of the Study:

• * To propose an enhanced pre-training method, Improved Spatiotemporal Diffusion Graph (ImPreSTDG), for traffic forecasting.
• * To address limitations in capturing long-term spatiotemporal dependencies and high computational costs in current models.
• * To improve model generalization ability and reduce retraining overhead.

Main Methods:

• * Integration of a Denoised Diffusion Probability Model (DDPM) into the pre-training process to enhance learning from long-term data and reduce computational load.
• * Implementation of a data masking and recovery strategy during pre-training, with DDPM reconstructing masked segments.
• * Inclusion of the Mamba module, utilizing Selective State Space Models (SSM), for efficient processing of long sequences and capturing multivariate spatiotemporal correlations.

Main Results:

• * The proposed ImPreSTDG method significantly improves the model's capability to handle long-term spatiotemporal dependencies.
• * The approach effectively addresses challenges related to missing data and high computational expenses.
• * Experiments on three real-world traffic datasets validate the enhanced efficiency and accuracy of the pre-training method.

Conclusions:

• * The ImPreSTDG pre-training method offers a more efficient and accurate solution for traffic forecasting, particularly for long-term dependencies.
• * The study demonstrates a significant reduction in computational costs without compromising prediction accuracy.
• * This enhanced approach provides a robust framework for smart city traffic management systems.