M2FMoE: Multi-Resolution Multi-View Frequency Mixture-of-Experts for Extreme-Adaptive Time Series Forecasting

May 7, 2026 · View on GitHub

Motivation

Fig. 1: Comparison of frequency spectra between regular and extreme events.

Frequency-domain representations provide an effective way to decompose temporal dynamics into spectral components, separating high-frequency fluctuations from low-frequency trends.

As illustrated in Fig. 1(a)–(d), the differenced sequences $\Delta\mathbf{X}$ show clear contrasts between extreme and regular events. These differences are amplified in the wavelet domain (Fig. 1(e), (g)), where extreme events yield sharp, localized energy at fine resolutions; as the resolution becomes coarser, energy shifts toward lower frequencies with reduced intensity while the event structure remains consistent. In contrast, regular events exhibit smooth low-frequency dynamics, producing diffuse and uniform energy across resolutions. Similar patterns appear in the Fourier domain (Fig. 1(f), (h)): extreme sequences show broad-spectrum, multi-peaked energy with slow spectral decay, whereas regular sequences concentrate energy within narrow low-frequency bands.

Model Architecture

Fig. 2: The proposed M2FMoE

Acknowledgements

We thank the authors of the following repositories for their open-source code or datasets used in our experiments:

MC-ANN: https://github.com/davidanastasiu/mcann
FreqMoE: https://github.com/sunbus100/FreqMoE-main
Time-Series-Library: https://github.com/thuml/Time-Series-Library
MoLE: https://github.com/RogerNi/MoLE

Datasets

Datasets can be found in the MC-ANN repository:
https://github.com/davidanastasiu/mcann

Citation

If you find this work useful, please cite:

@inproceedings{huang2026m2fmoe_aaai,
  title = {{M2FMoE}: Multi-Resolution Multi-View Frequency Mixture-of-Experts for Extreme-Adaptive Time Series Forecasting},
  author = {Yaohui Huang and Runmin Zou and Yun Wang and Laeeq Aslam and Ruipeng Dong},
  booktitle = {Proceedings of the 40th AAAI Conference on Artificial Intelligence},
  year = {2026},
  volume = {40},
  number = {26},
  pages = {22075-22083},
  doi = {https://doi.org/10.1609/aaai.v40i26.39362},
  url = {https://ojs.aaai.org/index.php/AAAI/article/view/39362},
}

@misc{huang2026m2fmoe_arxiv,
  title         = {{M$^2$FMoE}: Multi-Resolution Multi-View Frequency Mixture-of-Experts for Extreme-Adaptive Time Series Forecasting},
  author        = {Yaohui Huang and Runmin Zou and Yun Wang and Laeeq Aslam and Ruipeng Dong},
  year          = {2026},
  eprint        = {2601.08631},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG},
  url           = {https://arxiv.org/abs/2601.08631}
}