Spatial Augmentation Analysis for Long-Term MTS Forecasting

Official implementation of "How Much Does the Spatial Module Really Help? A Systematic Analysis for Long-Term MTS Forecasting"

Paper: H. T. Moges and D. Moodley — SAICSIT 2026, CCIS, Springer Contact: ht.moges@gmail.com
Homepage: https://htmoges.github.io

Get Started

1. Install

pip install -r requirements.txt

2. Download Data

Datasets from Time-Series-Library. Place in ./data/:

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

3. Reproduce Results

bash scripts/run_linear.sh       # Table 4 — DLinear / NLinear / RLinear
bash scripts/run_patchtst.sh     # Table 5 — PatchTST
bash scripts/run_horizon.sh      # Table 6 — horizon sweep

Or run a single experiment:

python -m gdf.train --config configs/linear/dlinear_electricity.yaml --seeds 42 1 2 3

Evaluation Configurations (TTS Design)

The study uses four configurations formed by selectively enabling or disabling the prediction head (PH), input skip connection (R), and spatial module (S) within a shared TTS architecture. This isolates each component's contribution independently of the others:

Config	Pred. Head	Input Residual	Spatial
L	✗	✗	✗
L+PH	✓	✗	✗
L+PH+R	✓	✓	✗
L+PH+R+S	✓	✓	✓

Same protocol for PatchTST: Base / +R / +R+S / +S, evaluating spatial behaviour as a function of the residual connection.

Citation

@inproceedings{moges2026spatial,
  title={How Much Does the Spatial Module Really Help? A Systematic Analysis for Long-Term MTS Forecasting},
  author={Moges, H.T. and Moodley, D.},
  booktitle={Proceedings of SAICSIT 2026, CCIS, Springer},
  year={2026}
}

Acknowledgements

• Time-Series-Library — benchmark datasets and baselines
• Lite-STGNN — STGNN modeling framework

License

MIT License