SpikF-GO: Spiking Fourier Graph Operators for Multivariate Time Series Forecasting

📄 Paper (arXiv): https://arxiv.org/abs/2606.13901
💻 GitHub: https://github.com/jafarbakhshaliyev/SpikF-GO

Official implementation of SpikF-GO: Spiking Fourier Graph Operators for Multivariate Time Series Forecasting, accepted to the ECML PKDD 2026 Research Track.

---

Abstract

SpikF-GO is a spiking neural architecture for multivariate time series forecasting. It combines the hypervariate graph formulation of FourierGNN with spike-driven Fourier-domain graph processing, enabling joint modeling of intra-series temporal dependencies, inter-series dependencies, and time-varying cross-variable interactions. The model introduces sparse frequency selection and Complex LIF-based spectral gating to preserve event-driven computation in the Fourier domain. We also provide SpikF-GO w/ CPG, which incorporates Central Pattern Generator-based positional signals for improved long-range temporal modeling.

---

Key Contributions

• Graph-based SNN forecasting: SpikF-GO brings hypervariate graph modeling into SNN-based multivariate time series forecasting.
• Spike-driven Fourier graph operators: The model combines sparse frequency gating with Complex LIF-based spectral processing to preserve event-driven computation in the Fourier domain.
• Unified SNN benchmark: We evaluate SpikF-GO against major SNN forecasting families under a common experimental protocol across eight benchmark datasets.
• Energy-aware forecasting: SpikF-GO achieves competitive-to-superior forecasting performance while reducing theoretical energy consumption relative to FourierGNN.

---

Related Library: SpikingTSF

We also maintain SpikingTSF, a broader open-source benchmark library for spiking neural network-based time series forecasting. SpikingTSF unifies SNN forecasting architectures and ANN baselines under a common training and evaluation protocol across datasets, horizons, metrics, and random seeds.

Note: SpikingTSF is a benchmarking library and may not reproduce all experiments from this repository directly.

---

Repository Structure

SpikF-GO/
├── README.md
├── LICENSE
├── CITATION.cff
├── requirements.txt
├── train.py                  # main training & evaluation entry point
├── model/                    # SpikF-GO + all baseline implementations
├── utils/                    # shared utilities (metrics, helpers)
├── data/
│   └── data_loader.py        # dataset loading (raw files placed here at runtime)
├── scripts/
│   ├── ecg.sh
│   ├── covid.sh
│   ├── solar.sh
│   ├── ecl.sh
│   ├── traffic.sh
│   ├── metr_la.sh
│   ├── pems_bay.sh
│   └── wiki.sh
└── assets/
    ├── spikf-go-architecture.png
    └── supplementary.pdf

---

Environment Setup

Create and activate a virtual environment:

Linux / macOS

python3 -m venv venv
source venv/bin/activate

Windows

python -m venv venv
venv\Scripts\activate

Install dependencies:

pip install -r requirements.txt

Experiments were run with PyTorch 2.5.1 on a single NVIDIA RTX 4090.

---

Dataset

Download the processed datasets from Figshare:

https://figshare.com/s/7617530bce306584fe95?file=62576929

Place all dataset files directly inside the data/ folder (do not create subfolders):

SpikF-GO/
├── data/
│   ├── dataset_file_1
│   ├── dataset_file_2
│   └── ...
├── model/
├── scripts/
└── train.py

---

Run Experiments

Scripts are in scripts/, one per dataset:

bash scripts/ecg.sh
bash scripts/covid.sh
bash scripts/solar.sh
bash scripts/ecl.sh
bash scripts/traffic.sh
bash scripts/metr_la.sh
bash scripts/pems_bay.sh
bash scripts/wiki.sh

Each script sets the exact hyperparameters used to produce the results reported in the paper.

---

Supplementary Material

Available at assets/supplementary.pdf.

---

Citation

If you use this code or build on SpikF-GO, please cite our paper:

arXiv preprint:

@misc{bakhshaliyev2026spikfgo,
  title        = {SpikF-GO: Spiking Fourier Graph Operators for Multivariate Time Series Forecasting},
  author       = {Bakhshaliyev, Jafar and Landwehr, Niels},
  year         = {2026},
  eprint       = {2606.13901},
  archivePrefix= {arXiv},
  primaryClass = {cs.LG},
  url          = {https://arxiv.org/abs/2606.13901}
}

ECML PKDD 2026 proceedings:

@inproceedings{bakhshaliyev2026spikfgo,
  title     = {SpikF-GO: Spiking Fourier Graph Operators for Multivariate Time Series Forecasting},
  author    = {Bakhshaliyev, Jafar and Landwehr, Niels},
  booktitle = {ECML PKDD},
  year      = {2026}
}

See CITATION.cff for full citation metadata.

---

Acknowledgements

The baselines in model/ build on prior work. We thank the authors for releasing their code; original licenses are respected.

• SpikF.py — adapted from SpikF (Wu, Huo & Chen, "SpikF: Spiking Fourier Network for Efficient Long-term Prediction", ICML 2025 / PMLR v267).
• TS_Former.py, TS_GRU.py, TS_TCN.py — adapted from TS-LIF (Feng et al., "TS-LIF: A Temporal Segment Spiking Neuron Network for Time Series Forecasting", arXiv:2503.05108).
• iSpikformer.py, SpikeGRU.py — adapted from SeqSNN (Lv et al., "Efficient and Effective Time-Series Forecasting with Spiking Neural Networks", arXiv:2402.01533), microsoft/SeqSNN.
• SpikeRNN_CPG.py, SpikeTCN_CPG.py, Spikformer_CPG.py — CPG variants build on arXiv:2405.14362 / microsoft/SeqSNN.
• FourierGNN.py — adapted from FourierGNN, arXiv:2311.06190 / aikunyi/FourierGNN.

---

License

This project is released under the MIT License.