MK-TSOD (Multi-Kernel Time Series Outlier Detection)

This repository hosts the supplementary material for the paper:

• Multi-Kernel Time Series Outlier Detection. (Under review).

The repository contains our code, and instructions to reproduce our experiments.

We use publicly available data from the UCR Time Series Classification Archive.

We release our code with an AGPLv3 license. If you are using code from this repository, please cite our paper.

Instructions for reproducing the experiments

Follow the steps below to reproduce the results of MK-TSOD. We give the random seeds, to ensure reproducibility. We do not share the source code of the competitors, which may be obtained from the respective sources, or by contacting the authors. We include our result file, for ease of viewing with notebooks/BA.ipynb.

1. Clone this repository.

2. Obtain the data from UCR per instructions on their webpage.

    cp UCRArchive_2018.zip tsod-svdd/data/raw/
    cd data/raw/
    unzip UCRArchive_2018.zip (requires password)
   

3. Install dependencies.

    conda create -n tsod python=3.7 pandas=1.3.3 scikit-learn=1.0 numexpr=2.7.3 jupyter numpy=1.21.5
    conda activate tsod
    pip install tsfresh dtaidistance
    pip install git+https://github.com/anon-repository/tsvdd.git@main
   

4. Create train data / CV folds for outlier detection.

    cd src/
    python preprocess.py
   

5. Create data for runtime experiments.

    cd src/
    python preprocess_runtime.py
   

6. Run the algorithm.

    cd src/
    python run.py
    # to run runtime experiments, uncomment the code in `run.py`.
   

7. View the results.

    cd <repository root>
    jupyter notebook
    # open notebooks/visualize.ipynb
   

Acknowledgements

• We base our SimpleMKL implementation on Wenyi Qin's, available here.