STM3 is designed for long-term spatio-temporal forecasting, where multiscale temporal patterns and heterogeneous spatial dependencies are deeply entangled. It addresses this challenge with three core components:

• Multiscale Mamba captures long-range temporal dynamics across multiple scales within an efficient Mamba block.
• AGCCN performs causal scale-wise spatio-temporal fusion, preserving scale distinguishability while modeling adaptive inter-node correlations.
• DMoE routes spatial nodes with static embeddings and contrastive regularization, disentangling heterogeneous node-level patterns across experts.

News

• [2026.05]: 🎉 STM3 has been accepted by KDD 2026!

Installation

We recommend using a Conda environment.

conda create -n stpredict python=3.10
conda activate stpredict
pip install -r requirements.txt

Some dependencies, such as mamba-ssm and causal-conv1d, are sensitive to CUDA, PyTorch, and compiler versions. If installation fails, install these packages following the version requirements of your local CUDA environment.

Data

Datasets are available here.

After downloading, place each dataset in the corresponding subdirectory under data/.

Quick Start

python run.py --dataset METR_LA --model STM3 --lag 96 --horizon 96 --cuda_devices 0

Citation

If you find this repository useful, please cite:

@article{chen2025stm3,
  title={STM3: Mixture of Multiscale Mamba for Long-Term Spatio-Temporal Time-Series Prediction},
  author={Chen, Haolong and Zhang, Liang and Xin, Zhengyuan and Zhu, Guangxu},
  journal={arXiv preprint arXiv:2508.12247},
  year={2025}
}

Further Reading

1. An overview of domain-specific foundation model: key technologies, applications and challenges in SCIS 2026: Broader discussion of key technologies, applications, and challenges in domain-specific foundation models.

@article{chen2026overview,
  title={An overview of domain-specific foundation model: key technologies, applications and challenges},
  author={Chen, Haolong and Chen, Hanzhi and Zhao, Zijian and Han, Kaifeng and Zhu, Guangxu and Zhao, Yichen and Du, Ying and Xu, Wei and Shi, Qingjiang},
  journal={Science China Information Sciences},
  volume={69},
  number={1},
  pages={111301},
  year={2026},
  publisher={Springer}
}

2. DK-Root: A Joint Data-and-Knowledge-Driven Framework for Root Cause Analysis of QoE Degradations in Mobile Networks, arXiv 2025: Root cause classification of QoE degradations using multidimensional time-series KPIs in mobile networks.

@article{li2025dk,
  title={DK-Root: A Joint Data-and-Knowledge-Driven Framework for Root Cause Analysis of QoE Degradations in Mobile Networks},
  author={Li, Qizhe and Chen, Haolong and Li, Jiansheng and Chai, Shuqi and Li, Xuan and Hou, Yuzhou and Shao, Xinhua and Li, Fangfang and Han, Kaifeng and Zhu, Guangxu},
  journal={arXiv preprint arXiv:2511.11737},
  year={2025}
}

3. FedRMamba: Federated Residual Mamba for Multivariate Time-Series Forecasting in WWW 2026: A federated foundation model for multivariate time-series forecasting with residual Mamba architectures.

@inproceedings{hu2026fedrmamba,
  title={FedRMamba: Federated Residual Mamba for Multivariate Time-Series Forecasting},
  author={Hu, Zhiwei and Zhang, Liang and Zhu, Guangxu},
  booktitle={Proceedings of the ACM Web Conference 2026},
  pages={7610--7620},
  year={2026}
}

4. CSI-BERT2: A BERT-inspired Framework for Efficient CSI Prediction and Classification in Wireless Communication and Sensing in TMC 2025: Efficient prediction and classification of CSI time-series data in wireless communication and sensing.

@article{zhao2025csi,
  title={CSI-BERT2: A BERT-inspired Framework for Efficient CSI Prediction and Classification in Wireless Communication and Sensing},
  author={Zhao, Zijian and Meng, Fanyi and Lyu, Zhonghao and Li, Hang and Li, Xiaoyang and Zhu, Guangxu},
  journal={IEEE Transactions on Mobile Computing},
  year={2025},
  publisher={IEEE}
}