FreqMoE: Enhancing Time Series Forecasting through Frequency Decomposition Mixture of Experts (AISTATS 2025 poster)

March 17, 2025 ยท View on GitHub

The official implementation of "FreqMoE: Enhancing Time Series Forecasting through Frequency Decomposition Mixture of Experts".

[Arxiv]

TLDR

In this project we propose FreqMoE, a frequency-based Mixture of Experts model for long-term time series forecasting. Unlike existing methods, FreqMoE dynamically decomposes time series into frequency bands, with specialized experts processing each band. A gating mechanism adjusts expert contributions, and a prediction module refines forecasts via residual connections. Experiments show FreqMoE achieves SOTA performance across eight datasets while keeping parameters under 50k, ensuring high efficiency.

Overall Architecture

structure

Datasets Preparation

You can access all nine benchmark datasets from the Google Drive link provided in Autoformer. These datasets are well pre-processed and readily usable. Please download the datasets and put them in the ./dataset folder. Each dataset is an .csv file.

Environment requirements

pip install -r requirements.txt

Please refer to the requirements.txt file for the required packages.

Acknowledgement

We thank the authors of the following repositories for their open-source code or dataset, which we used in our experiments:

Citation

If you find our work useful, please consider citing our paper using the following BibTeX:

@inproceedings{
liu2025freqmoe,
title={FreqMoE: Enhancing Time Series Forecasting through Frequency Decomposition Mixture of Experts},
author={Ziqi Liu},
booktitle={The 28th International Conference on Artificial Intelligence and Statistics},
year={2025},
url={https://openreview.net/forum?id=LGkLTYLIBq}
}