CrossLinear

Official implementation of "CrossLinear: Plug-and-Play Cross-Correlation Embedding for Time Series Forecasting with Exogenous Variables" (KDD 2025).

Slide is [here], and Poster is [here].

Usage

1. Install Python 3.8.20 first. For convenience, then execute the following command.

pip install -r requirements.txt

2. Prepare Data. You can obtain the datasets from [Google Drive]. Then unzip the downloaded data.

unzip dataset.zip

3. Train and evaluate model. We provide the experiment scripts under the folder ./scripts/. You can reproduce the experiment results as the following examples:

# exogenous forecast
sh ./scripts/exogenous_forecast/ECL/CrossLinear.sh
# multivariate forecast
sh ./scripts/multivariate_forecast/ECL/CrossLinear.sh

Citation

If you find this repo useful, please cite our paper.

@inproceedings{CrossLinear,
  title = {CrossLinear: Plug-and-Play Cross-Correlation Embedding for Time Series Forecasting with Exogenous Variables},
  author = {Zhou, Pengfei and Liu, Yunlong and Liang, Junli and Song, Qi and Li, Xiangyang},
  booktitle = {Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2},
  year = {2025},
  doi = {10.1145/3711896.3736899},
}

Contact

If you have any questions or suggestions, feel free to contact Pengfei Zhou (pengfeizhou@mail.ustc.edu.cn).

Acknowledgement

This work is constructed based on the following repos:

• Autoformer: https://github.com/thuml/Autoformer;

• Informer: https://github.com/zhouhaoyi/Informer2020;

• Time Series Library: https://github.com/thuml/Time-Series-Library.