This is an official implementation of a KDD2024 paper, "Fredformer: Frequency Debiased Transformer for Time Series Forecasting." https://arxiv.org/abs/2406.09009

Fredformer focuses on addressing the issue of frequency bias in Transformer models for time series forecasting. This bias can cause the model to fail to capture mid-to-high-frequency information in the data. We have conducted empirical analyses on this issue and proposed a solution in this work. Extensive experimental results on eight datasets show the effectiveness of Fredformer.

For more details, please refer to our original paper.

Fredformer is built based on PyTorch. You can install PyTorch following the instructions in PyTorch. For example:

pip install torch==1.10.0+cu111 torchvision==0.11.0+cu111 torchaudio==0.10.0 -f https://download.pytorch.org/whl/torch_stable.html

After ensuring that PyTorch is installed correctly, you can install other dependencies via:

pip install -r requirements.txt

We have prepared a dataset for this example: Weather. The CSV file is in the dataset folder.

If you are using your data, please format it as a CSV file, with each row representing a sample and each column representing a channel. When selecting the data type "data_name," please choose "Custom" and ensure the CSV file is stored in the dataset folder.

• --patch_len (int, default=16): Frequency patching length.
• --cf_dim (int, default=48): Feature dimension.
• --cf_drop (float, default=0.2): Dropout rate.
• --cf_depth (int, default=2): Number of Transformer layers.
• --cf_heads (int, default=6): Number of multi-heads.
• --cf_mlp (int, default=128): Feed-forward network dimension.
• --cf_head_dim (int, default=32): Dimension for each single head.
• --use_nys (int, default=0): Use Nyström method (0 = No, 1 = Yes).
• --mlp_drop (float, default=0.3): Dropout rate for MLP.

For Fredfromer: The scripts for our are in the directory ./scripts/Fredformer. You can run the following command, then open ./result.txt to see the results once the training is done:

sh ./scripts/Fredformer/weather.sh

Log files will be generated and updated in ./logs/ during training.

All dataset scripts can currently be found in the scripts/Fedformer/ directory.

Please note that the Electricity and Traffic datasets use the Nystrom variant of Fedformer by default. If you need to change this, set the value of --use_nys to 0 in the corresponding .sh file.

# Example: Modify the use of the Nystrom variant in the shell script for the Electricity dataset
# Open the corresponding script file (e.g., train_electricity.sh) and find the line with --use_nys

# Original line in train_electricity.sh
--use_nys 1

# Modify to disable the Nystrom variant
--use_nys 0

Here are the main results of our experiment:

If you find our work useful in your research, please consider citing:

@inproceedings{Piao2024fredformer,
  title={Fredformer: Frequency Debiased Transformer for Time Series Forecasting},
  author={Xihao Piao and Zheng Chen and Taichi Murayama and Yasuko Matsubara and Yasushi Sakurai},
  booktitle={Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
  series = {KDD '24}
  year={2024}
}