GSF-LLM: Graph-Enhanced Spatio-Temporal Fusion-Based Large Language Model for Traffic Prediction

Area of Science:

• Artificial Intelligence
• Transportation Engineering
• Data Science

Background:

• Accurate traffic prediction is crucial for intelligent transportation systems and urban mobility.
• Existing deep learning models face challenges in capturing complex spatial-temporal dynamics and preventing overfitting in large-scale networks.

Purpose of the Study:

• To propose a novel framework, GSF-LLM (graph-enhanced spatio-temporal fusion-based large language model), that integrates LLMs with graph-based spatio-temporal learning.
• To address the limitations of current deep learning approaches in traffic prediction, specifically regarding spatial dependencies, temporal dynamics, and overfitting.

Main Methods:

• GSF-LLM utilizes a spatio-temporal fusion module for joint encoding of spatial and temporal correlations.
• A partially frozen graph attention (PFGA) mechanism models topological dependencies while mitigating overfitting.
• Low-rank adaptation (LoRA) fine-tunes a subset of LLM parameters for improved training efficiency and generalization.

Main Results:

• GSF-LLM demonstrates superior performance compared to state-of-the-art baselines on multiple real-world traffic datasets.
• The framework effectively captures complex spatial and temporal traffic patterns.
• The proposed methods successfully mitigate overfitting issues common in large-scale traffic network modeling.

Conclusions:

• GSF-LLM offers a powerful new approach for accurate traffic prediction by integrating LLMs and graph-based spatio-temporal learning.
• The framework shows significant improvements in handling spatial dependencies, temporal dynamics, and overfitting.
• GSF-LLM has potential applications in related intelligent transportation tasks like data imputation, trajectory generation, and anomaly detection.