Lite-STGNN

Official PyTorch implementation of Lite-STGNN: A Lightweight Spatial-Temporal Graph Neural Network for Long-Term Time Series Forecasting

Published Paper | arXiv

Get Started

1. Install Dependencies

pip install -r requirements.txt

2. Download Data

All datasets can be obtained from Time-Series-Library.

Place datasets in ./data/:

data/
├── electricity.csv
├── traffic.csv
├── exchange_rate.csv
└── weather.csv

3. Run Experiments

Reproduce Paper Results:

bash scripts/reproduce_paper_results.sh

Or run individual experiments:

python src/lite_stgnn_modular.py \
  --dataset electricity \
  --data-root ./data \
  --seq-len 96 --pred-len 720 --epochs 50 \
  --learning-rate 5e-4 \
  --adj-rank 16 --adj-topk 10 --adj-tau 1.2 \
  --use-input-residual \
  --use-temporal-head --temporal-head-ratio 0.5 \
  --residual-dropout 0.1 \
  --gate-mode band --prop-orders 2

Architecture

Lite-STGNN combines temporal modeling with learnable spatial dependencies for efficient long-term time-series forecasting:

Key Components:

• Temporal Modeling (Dlinear): Trend-seasonal decomposition with separate linear projections
• Learnable Adjacency Matrix: Low-rank factorization with TopK sparsification for discovering spatial dependencies
• Graph Propagation: Multi-hop spatial aggregation with learnable graph structure
• Residual Gating: Adaptive horizon-wise weighting mechanism
• Optional Enhancements: Temporal refinement head, input skip connections, feature normalization

Citation

If you find this work useful, please cite our paper:

@inproceedings{moges2026litestgnn,
  title={A Lightweight Spatial-Temporal Graph Neural Network for Long-Term Time Series Forecasting},
  author={Moges, H.T. and Moodley, D.},
  booktitle={Proceedings of the 18th International Conference on Agents and Artificial Intelligence (ICAART)},
  volume={3},
  pages={2743--2750},
  year={2026},
  publisher={SciTePress},
  doi={10.5220/0014322300004052}
}

Published paper: https://www.scitepress.org/Link.aspx?doi=10.5220/0014322300004052
arXiv preprint: https://arxiv.org/abs/2512.17453

Acknowledgement

We appreciate the following repositories for their valuable code and datasets:

• Time-Series-Library - Benchmark datasets
• LTSF-Linear - DLinear implementation
• Graph-WaveNet - Learnable adjacency matrix inspiration

License

MIT License