EMMV

Implementation of EM/MV metrics based on N. Goix et al.

This is a means of evaluating anomaly detection models without anomaly labels.

Installation

pip install emmv

Example Use

from emmv import emmv_scores
excess_mass_score, mass_volume_score = emmv_scores(model, features)

• Where 'model' is your trained scikit-learn/PyOD/PyCaret/etc. model
• Where 'features' is a 2D DataFrame of features (the X matrix)

If you are using models without a built-in decision_function (e.g. Keras or ADTK models), then you need to specify an anomaly scoring function. Please see examples in ./src/emmv/examples folder.

Running Examples

Examples exist for ADTK, Alibi-Detect, PyCaret, PyOD, scikit-learn, and TensorFlow (Keras).

pip install -r requirements.txt
pip install -e .
python src/emmv/examples/adtk_example.py          # For an ADTK example
python src/emmv/examples/alibi_detect_example.py  # For a Alibi Detect example
python src/emmv/examples/keras_example.py         # For a Keras example
python src/emmv/examples/pycaret_example.py       # For a PyCaret example
python src/emmv/examples/pyod_example.py          # For a PyOD example
python src/emmv/examples/sklearn_example.py       # For a scikit-learn example

Interpreting scores

• The best model should have the highest Excess Mass score
• The best model should have the lowest Mass Volume score
• Probably easiest to just use one of the metrics
• Extreme values are possible

Contact

Please feel free to get in touch at [email protected]

Citation

Christian O'Leary (2024) EMMV library

@Misc{emmv,
    author = {Christian O'Leary},
    title = {EMMV library},
    howpublished = {\url{https://pypi.org/project/emmv/}},
    year = {2021}
}

Development

conda create -n emmv python=3.10 -y
conda activate emmv
pip install -r ./tests/requirements.txt
pip install -e .
conda install pre-commit
pre-commit install
pytest
pylint src