DualSG: A Dual-Stream Explicit Semantic-Guided Multivariate Time Series Forecasting Framework

🚀 Highlights

• 💡 Dual-stream forecasting decouples numerical and semantic modeling to resolve LLM precision bottlenecks.
• 🔁 Time series captioning generates intermediate semantic representations to guide numerical forecasting.
• 🤖 TimeAwareGPT2 decoder enhances LLM alignment with temporal patterns using temporal position control.

---

🗂️ 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.

Please download the weights from the link below and place them in the path TS_Caption_GPT/checkpoints/.

https://drive.google.com/file/d/1h5IaCC41lM-sHJbf9-2w192LUlMv9BCS/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 Wang, Xiaorui and Gong, Luqi and Jiang, Yishan and Luo, Chunjie and Zhan, Jianfeng and others},
  booktitle={arXiv preprint arXiv:2507.21830},
  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.