Reference / maintained implementation of the SVQ-OCC model from:

D. Staps, R. Schubert, M. Kaden, A. Lampe, W. Hermann, T. Villmann, "Prototype-based one-class-classification learning using local representations", 2022 International Joint Conference on Neural Networks (IJCNN), IEEE, 2022, pp. 1–8. DOI: 10.1109/IJCNN55064.2022.9892912

SVQ-OCC is a prototype-based one-class classifier for outlier / novelty detection built in the spirit of Generalized Learning Vector Quantization (GLVQ). Each prototype acts as the center of a flexible enclosing ball; the centers are learned from the data density and the per-ball radius is adapted via the local underlying distribution, optimized end-to-end by stochastic gradient descent. The "local representations" give an interpretable decision boundary (a point is an outlier if it falls outside every ball).

Implemented as the Python package prototorch_oneclass, built on prototorch.

• Models: OneClassGLVQ, OneClassGMLVQ, SVQ_OCC (Euclidean and kernelized variants).
• Classification costs: brier, pcs, ce. Representation cost: neural-gas (NG).
• Key parameter: K (number of prototypes/balls).

pip install -e .        # or: pip install -r requirements.txt
python3 examples/...    # see examples/ for runnable scripts

This repository is the canonical, self-contained home of prototorch_oneclass; it no longer depends on private si-cim research repositories. It is also the package used as the SVQ-OCC baseline by the K-MEB work (companion experiments repo, kept private).

See CITATION.cff and cite the paper above.

BibTeX:

@inproceedings{staps2022svqocc,
  title        = {Prototype-based one-class-classification learning using local representations},
  author       = {Staps, Daniel and Schubert, Ronny and Kaden, Marika and Lampe, Alexander and Hermann, Wieland and Villmann, Thomas},
  booktitle    = {2022 International Joint Conference on Neural Networks (IJCNN)},
  pages        = {1--8}, year = {2022}, organization = {IEEE},
  doi          = {10.1109/IJCNN55064.2022.9892912}
}

By Daniel Staps (0009-0002-4459-4544) with R. Schubert, M. Kaden, A. Lampe, W. Hermann and Thomas Villmann (0000-0001-6725-0141). Repository cleanup with assistance from Claude (Anthropic).

MIT — see LICENSE.