DualSG

[中文解读1] [中文解读2] [中文解读3]

This is the official implementation of our paper: DualSG: A Dual-Stream Explicit Semantic-Guided Multivariate Time Series Forecasting Framework

Multivariate Time Series Forecasting (MTSF) is fundamental to many real-world applications. Recent studies leverage Large Language Models (LLMs) for MTSF, but treating them as end-to-end forecasters often sacrifices numerical precision, while latent-space alignment between text and time series remains challenging. We propose DualSG, a dual-stream framework that uses LLMs as semantic guidance modules rather than standalone predictors. DualSG introduces Time Series Caption, an explicit natural-language summary of temporal patterns, providing interpretable context for LLMs. A caption-guided fusion module further captures inter-variable dependencies with reduced noise and computation. Experiments across diverse datasets show consistent improvements over 15 state-of-the-art methods.

Distinctive Aspects of Our Approach

Our work departs from existing LLM-based time series forecasting methods by redefining the role of LLMs. Instead of using them as end-to-end predictors or relying on implicit latent-space alignment, we position LLMs as explicit semantic guides that refine numerical forecasts. By introducing Time Series Caption, we convert temporal patterns into interpretable natural language, enabling effective semantic guidance without sacrificing numerical precision. This design preserves the strengths of traditional forecasting models while incorporating high-level reasoning in a controlled and transparent manner.

Project Structure

├── models/ 
├── exp/ # Experiment runners
├── data_provider/ # Dataset loaders
├── layers/ # Transformer components
├── utils/ # Tools, losses, metrics
├── TS_Caption_GPT/ # Time-aware GPT2 decoder and checkpoints
├── scripts/ # Shell scripts for reproducibility
├── run.py # Entry point for numerical forecasting
├── requirements.txt # Python dependencies
└── README.md

Setup

conda create -n dualsg python=3.9
conda activate dualsg
pip install -r requirements.txt

Place your datasets under ./dataset/. See data_provider/ for supported formats.

process data can be download in : https://drive.google.com/file/d/15hsRcUyx79uJ5aiXFK1LA-zxbMY3aw8W/view?usp=sharing

Usage

Long-term Forecasting

bash DualSG/scripts/DualSG/DualSG_ETTh1.sh

Citation

If you find our work helpful, please consider citing us:

@inproceedings{ding2025dualsg,
  title={DualSG: A Dual-Stream Explicit Semantic-Guided Multivariate Time Series Forecasting Framework},
  author={Ding, Kuiye and Fan, Fanda and Wang, Yao and Jian, Ruijie and Wang, Xiaorui and Gong, Luqi and Jiang, Yishan and Luo, Chunjie and Zhan, Jianfeng},
  booktitle={Proceedings of the 33rd ACM International Conference on Multimedia},
  pages={508--517},
  year={2025}
}

Acknowledgements

All the experiment datasets are public, and we obtain them from the following links:

• Time-Series-Library: https://github.com/thuml/Time-Series-Library/.

• TFB: https://github.com/decisionintelligence/TFB.

• TimeMixer: https://github.com/kwuking/TimeMixer.