Related Experiment Videos
DG-LLM: Decomposition-based dynamic graph adaptation of large language models for spatiotemporal traffic forecasting.
Sadia Tabassum1, Naushin Nower1
1Institute of Information Technology, University of Dhaka, Dhaka, Bangladesh.
Plos One
|May 19, 2026
Summary
DG-LLM enhances traffic forecasting by decomposing signals and using adaptive graphs with Large Language Models (LLMs). This novel framework improves accuracy in complex spatiotemporal modeling, outperforming existing methods for both short and long-term predictions.
Area of Science:
- Artificial Intelligence
- Urban Planning
- Data Science
Background:
- Accurate traffic forecasting is vital for urban planning.
- Existing methods face challenges in modeling complex spatiotemporal dependencies and long-term patterns due to multiscale traffic data.
- Limitations include difficulties in capturing intricate spatial relationships and temporal dynamics inherent in traffic flow.
Purpose of the Study:
- To introduce DG-LLM, a novel framework for enhanced traffic forecasting.
- To address limitations in modeling complex spatiotemporal dependencies and long-term patterns.
- To improve the accuracy and robustness of traffic prediction models.
Main Methods:
- Decomposition of traffic signals into intrinsic modes.
- Learning dynamic graphs for each mode to represent spatial dependencies.
- Integration of these representations with pre-trained Large Language Models (LLMs) for temporal dependency modeling.
Main Results:
- DG-LLM demonstrated significant improvements over state-of-the-art spatiotemporal forecasting models.
- Achieved superior performance in both short- and long-term forecasting across six real-world datasets.
- Showcased substantial gains in Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) compared to baseline approaches.
Conclusions:
- The proposed DG-LLM framework effectively models complex spatiotemporal dependencies in traffic data.
- The model exhibits strong generalization capabilities, validated through ablation studies and cross-dataset evaluations.
- DG-LLM offers a robust solution for accurate traffic forecasting, even with temporal instability and missing data.
