Reference implementation of the stochastic-gradient-descent variant of the K minimum enclosing ball (K-MEB) outlier detector from:
D. Staps, T. Villmann, B. Paaßen, "K Minimum Enclosing Balls for Outlier Detection", WSOM+ 2024, Springer, 2024, pp. 174–184. DOI: 10.1007/978-3-031-67159-3_20
Fits K enclosing balls for outlier detection by stochastic gradient descent, inspired by
Generalized Learning Vector Quantization (GLVQ): points are visited one by one, balls are pushed
away from negatives and the closest ball is pulled toward positives, minimizing the soft
(hinge-relaxed) K-MEB objective (typically with Adam). Built on top of
prototorch. The companion alternating-optimization / QP
variant lives in [K-MEB_qp].
pip install -r requirements.txt
python3 examples/KMEB_irisAnomalie.py # Euclidean
python3 examples/KernelizedKMEB_irisAnomalie.py # kernelizedThe exact paper state is on branch published / tag v-paper-2024a.
| Path | Purpose |
|---|---|
kmeb_prototorch/oneclass.py |
K-MEB model (prototorch GLVQ-based) |
kmeb_prototorch/functions/ |
distances, competitions, losses, callbacks |
examples/ |
runnable examples (Euclidean + kernelized) |
prototorch / prototorch_models (PyPI) provide the LVQ building blocks. This repo does not
depend on the private prototorch_oneclass research package.
See CITATION.cff and cite the paper above.
BibTeX:
@inproceedings{staps2024kmeb,
title = {K Minimum Enclosing Balls for Outlier Detection},
author = {Staps, Daniel and Villmann, Thomas and Paa{\ss}en, Benjamin},
booktitle = {Advances in Self-Organizing Maps, Learning Vector Quantization, Interpretable Machine Learning, and Beyond (WSOM+ 2024)},
pages = {174--184}, year = {2024}, publisher = {Springer},
doi = {10.1007/978-3-031-67159-3_20}
}By Daniel Staps (0009-0002-4459-4544), Thomas Villmann and Benjamin Paaßen. Cleanup with assistance from Claude (Anthropic).
MIT — see LICENSE.