[ICML 2025] Causality-Aware Contrastive Learning for Robust Multivariate Time-Series Anomaly Detection

August 14, 2025 · View on GitHub

Overview

The figures below illustrate the overview pipeline and components of the CAROTS framework. For more detailed information, please refer to our paper.

Overview

Component

Dataset Preparation

For the SMD and SMAP datasets, preprocessing was performed based on the script available at OmniAnomaly's data_preprocess.py. After preprocessing, the datasets were placed in the following directories:

data/
├── ServerMachineDataset/  # Preprocessed SMD dataset
└── SMAP_MSL/              # Preprocessed SMAP dataset

Synthetic Dataset Generation

To generate synthetic datasets (Lorenz96 and VAR), use the following commands:

cd data/Lorenz96
python generate.py

cd data/VAR
python generate.py

Dataset Organization

Place the downloaded datasets in the data/ directory, ensuring each dataset resides in its respective subdirectory:

data/
├── Lorenz96/
├── VAR/
├── SWaT/
├── WADI/
├── PSM/
├── ServerMachineDataset/
└── SMAP_MSL/

Example Execution

Scripts for running the model are located in the scripts/ directory. Use the following command to execute the model:

bash scripts/{dataset}.sh

Example: To run the model on the SWaT dataset:

bash scripts/SWaT.sh

The results of the model execution are saved in the results/{dataset} directory.

Citation

If you find our work useful, please cite our paper:

@article{kim2025causality,
  title={Causality-Aware Contrastive Learning for Robust Multivariate Time-Series Anomaly Detection},
  author={Kim, HyunGi and Mok, Jisoo and Lee, Dongjun and Lew, Jaihyun and Kim, Sungjae and Yoon, Sungroh},
  journal={arXiv preprint arXiv:2506.03964},
  year={2025}
}

License

This project is licensed under the MIT License. For commercial use, permission is required.

Acknowledgements

Please provide proper attribution if you use our codebase.
If you use our work, kindly cite our paper as mentioned in the Citation section.