---

osman: A pip package which oversamples class imbalanced binary data by Deep Generative Models.

This package offers two APIs.

1. Variational Auto Encoders

2. WGAN-GP

Usage Guide

Installation:

pip install osman

The APIs are very simple to use. This example shows the usage of the VAE

# Applying variational auto encoder on class imbalanced data 
from osman.oversampler import VAEify 

X_vae,y_vae=VAEify(X_train,y_train)

logreg.fit(X_vae,y_vae)

y_pred_vae=logreg.predict(X_test)

score_vae=roc_auc_score(y_pred_vae,y_test)

This example shows the usage of the WGAN-GP

from osman.oversampler import WGANify

X_wgan,y_wgan=WGANify(X_train,y_train)

logreg.fit(X_wgan,y_wgan)

y_pred_wgan=logreg.predict(X_test)

score_wgan=roc_auc_score(y_pred_wgan,y_test)

Furthermore, consult the tutorial ipython notebook to see the usage on real world credit default dataset.

Results

Results on UCI Credit Default dataset with Logistic Regression Classifier.

Data	ROC AUC Score
Original Imbalanced data	~38%
Oversampled by osman's VAE API	~55%
Oversampled by osman's WGAN-GP API	~50%

Acknowledgement:

1. https://arxiv.org/abs/1312.6114
2. https://arxiv.org/abs/2008.09202
3. https://github.com/justinengelmann/GANbasedOversampling