bae-sohee / oneclassclassifier

One-class classifiers for anomaly detection (outlier detection)

OneClassClassifier

Outlier detection using One-class classifier.

Install

python setup.py install

Usage

from occ.occ import *

## Load data
ocsvm = occ()
ocsvm.load_data_mat(file)
"""
Or
ocsvm.load_data_csv(file)
Or
ocsvm.X = some_np_array ##(n_samples, n_features)
"""

## Train the model and get results
ocsvm.train(model='ocsvm', kernel='rbf', norm=True) # If norm=True, data will be normalized(L2)
Y_scores = ocsvm.get_score(norm=True)
Y_hat = ocsvm.predict(norm=True)

## Visualization of the results
occ.show_projection(ocsvm.X, Y_scores, title="Score ocsvm with rbf kernel", markersize=100)
occ.show_projection(ocsvm.X, Y_hat, title="Prediction ocsvm with rbf kernel", markersize=100)

## Export the result score and outliers
ocsvm.export_outliers("outliers.csv", Y_hat) ## Raw data of outliers
ocsvm.export_csv("scores.csv", Y_scores) ## Raw data with scores
ocsvm.export_csv("predictions.csv", Y_hat) ## Raw data with predictions

Adjusting "nu" as the prior knowledge of the proportion of outliers

few_outliers = ocsvm.train(model='ocsvm', nu=0.01) ## 1% outliers assumed
many_outliers = ocsvm.train(model='ocsvm', nu=0.2) ## 20% outliers assumed

Using small subset for training

ocsvm = occ()
ocsvm.X = some_np_array
ocsvm.train(model='isolationForest', sampling=0.1) ## Only 10% of data will be used for training

Implemented Models

occ = occ()
occ.X = some_np_array
occ.train(model='SOMEMODEL')

• deepsvdd : Deep-SVDD; Deep One-Class Classification [ICML2018; Ruff, Lukas et al.,] ; Tensorflow2 implementation
• autoEncoder : AutoEncoder and reconstruction loss; pyod library
• vae : Variational AutoEncoder and reconstruction loss; pyod library
• isoForest : Isolation Forest; Isolation Forest[IEEE data mining conf 2008; Liu, Fei Tony, Kai Ming Ting, and Zhi-Hua Zhou.]; scikit-learn library
• ocnn : One-Class Neural Networks; Anomaly Detection Using One-Class Neural Networks [Chalapathy, R., Menon, A. K., & Chawla, S.]; Tensorflow2 implementation
• ocsvm : One-Class SVM; Support Vector Method for Novelty Detection [NIPS2000; Schölkopf, Bernhard, et al.]; scikit-learn library

Notice

• This module needs the Tensorflow 2.X.X version.