/PU-OC

This repository accompanies the paper "Improving State-of-the-Art in One-Class Classification by Leveraging Unlabeled Data".

Primary LanguagePythonMIT LicenseMIT

PU-OC

This repository accompanies the paper "Improving State-of-the-Art in One-Class Classification by Leveraging Unlabeled Data".

Structure

  • OC models with their PU modifications can be found in models

  • Pretrained autoencoder weights can be found in AE weights

  • OC and PU models tested in experimentŃ‹ 5.3, A.6, and A.5 can be found in models/reference

  • Testing functions for all settings can be found in test

    • Code for the experiment One-vs-all can be found in test/drocc/one-vs-all.py for DROCC-based models and in test/one-vs-all.py for other models
    • Code for the experiment Number of positive modes can be found in test/drocc/pos modality.py for DROCC-based models and in test/pos modality.py for other models
    • Code for the experiment Shift of the negative distribution can be found in test/drocc/neg shift.py for DROCC-based models and in test/neg shift.py for other models
    • Code for the experiment Size and contamination of unlabeled data can be found in test/unlabeled.py
    • Code for Abnormal1001 is the same as for One-vs-all
    • Code for the lstm-based models can be found in lstm.ipynb