---
license: mit
pretty_name: PATH
tags:
- time-series
- anomaly-detection
- multivariate-time-series
- unsupervised-learning
size_categories:
- 1K<n<10K
configs:
- config_name: normal
  data_files:
  - split: data
    path: 1_postsim/normal.parquet
- config_name: anomalous
  data_files:
  - split: data
    path: 1_postsim/anomalous.parquet
- config_name: control
  data_files:
  - split: data
    path: 1_postsim/control.parquet
- config_name: train
  data_files:
  - split: data
    path: 1_postsim/train.parquet
- config_name: test
  data_files:
  - split: data
    path: 1_postsim/test.parquet
- config_name: train_clean
  data_files:
  - split: data
    path: 1_postsim/train_clean.parquet
- config_name: test_clean
  data_files:
  - split: data
    path: 1_postsim/test_clean.parquet
---

# PATH: A Discrete-sequence Dataset for Evaluating Online Unsupervised Anomaly Detection Approaches for Multivariate Time Series

PATH contains multivariate time series of 16 features, simulated for evaluating online unsupervised anomaly detection. This repository mirrors [Zenodo](https://doi.org/10.5281/zenodo.13255120), which remains the archival source. The raw `.mat` files in `0_simulation` are identical to the Zenodo files. The pickle files of `1_postsim` are provided here as Parquet with the same names and the same contents. To convert them to the required pickle format after download, see [Usage](#usage).

- Paper: [doi.org/10.1016/j.bdr.2025.100573](https://doi.org/10.1016/j.bdr.2025.100573)
- Source code: [github.com/lcs-crr/PATH](https://github.com/lcs-crr/PATH)
- Zenodo: [doi.org/10.5281/zenodo.13255120](https://doi.org/10.5281/zenodo.13255120)

## Repository structure

```
0_simulation/              3,891 raw .mat files
1_postsim/
    normal.parquet         all nominal series
    anomalous.parquet      all anomalous series
    control.parquet        the control counterpart of each anomalous series
    train.parquet          3-fold cross-validation, training sets
    test.parquet           3-fold cross-validation, test sets
    train_clean.parquet    train.parquet without anomalous series
    test_clean.parquet     test.parquet without anomalous series
scripts/                   pkl_to_parquet.py and parquet_to_pkl.py
```

## Usage

`0_simulation` and `1_postsim` can be downloaded separately. The `.mat` files of `0_simulation` are identical to the Zenodo files:

```bash
hf download lcs-crr/PATH --type dataset --local-dir PATH --include "0_simulation/*"
```

The pickle files of `1_postsim` are lists of 2D float32 NumPy arrays. The name of the corresponding `.mat` file, and by extension the label, is in the dtype metadata (`array.dtype.metadata["file_name"]`):

```bash
hf download lcs-crr/PATH --type dataset --local-dir PATH --include "1_postsim/*" --include "scripts/*"
python PATH/scripts/parquet_to_pkl.py --parquet-dir PATH/1_postsim
```

The script needs `numpy` and `pyarrow`.

The `2_preprocessed` folder is not included, as its contents are specific to the TensorFlow data pipeline. It is generated from `1_postsim` by [`src/1_data.py`](https://github.com/lcs-crr/PATH/blob/main/src/1_data.py) in the source code repository.

## Citation

If you use this dataset, please cite it:

```bibtex
@misc{correiaPATHDiscretesequenceDataset2025,
  title = {{{PATH}}: {{A Discrete-sequence Dataset}} for {{Evaluating Online Unsupervised Anomaly Detection Approaches}} for {{Multivariate Time Series}}},
  author = {Correia, Lucas and Goos, Jan-Christoph and B{\"a}ck, Thomas and Kononova, Anna V.},
  year = {2025},
  publisher = {Zenodo},
  doi = {10.5281/ZENODO.13255120},
  copyright = {MIT License}
}
```

## License

MIT