_ _ _ _ _

fuzzy-rough-learn

fuzzy-rough-learn is a library of machine learning algorithms involving fuzzy rough sets, as well as data descriptors that can be used for one-class classification / novelty detection. It builds on scikit-learn, but uses a slightly different api, best illustrated with a concrete example:

from sklearn import datasets
from sklearn.metrics import accuracy_score, roc_auc_score
from sklearn.model_selection import train_test_split

from frlearn.base import probabilities_from_scores, select_class
from frlearn.classifiers import FRNN
from frlearn.feature_preprocessors import RangeNormaliser

# Import example data.
iris = datasets.load_iris()
X = iris.data
y = iris.target

# Split into train and test sets.
X_train, X_test, y_train, y_test = train_test_split(X, y, stratify=y, random_state=0)

# Create an instance of the FRNN classifier, construct the model, and query on the test set.
clf = FRNN(preprocessors=(RangeNormaliser(), ))
model = clf(X_train, y_train)
scores = model(X_test)

# Convert scores to probabilities and calculate the AUROC.
probabilities = probabilities_from_scores(scores)
auroc = roc_auc_score(y_test, probabilities, multi_class='ovo')
print('AUROC:', auroc)

# Select classes with the highest scores and calculate the accuracy.
classes = select_class(scores)
accuracy = accuracy_score(y_test, classes)
print('accuracy:', accuracy)

Both classifiers and feature preprocessors are functions that take training data and output a model. Models are functions that take data and output something else. Classifier models output class scores, preprocessor models output a transformation of the data. Preprocessors can be added as a keyword argument when initialising a classifier, which automatically creates a preprocessor model on the basis of the training data and applies it to the training and the test data.

Contents

At present, fuzzy-rough-learn contains the following algorithms:

Multiclass classifiers

• Fuzzy Rough Nearest Neighbours (FRNN; multiclass)
• Fuzzy Rough OVO COmbination (FROVOCO; muliclass, suitable for imbalanced data)
• Fuzzy ROugh NEighbourhood Consensus (FRONEC; multilabel)

Data descriptors

• Average Localised Proximity (ALP)
• Centre Distance (CD)
• Extended Isolation Forest (EIF)
• Isolation Forest (IF)
• Local Outlier Factor (LOF)
• Localised Nearest Neighbour Distance (LNND)
• Mahalanobis Distance (MD)
• Nearest Neighbour Distance (NND)
• Support Vector Machine (SVM)

Regressors

• Fuzzy Rough Nearest Neighbours (FRNN)

Feature preprocessors

• Fuzzy Rough Feature Selection (FRFS)

• Linear normaliser; in particular:

  • Interquartile range (IQR) normaliser
  • Maximum absolute value (MaxAbs) normaliser
  • Range normaliser
  • Standardiser

• Shrink Autoencoder (SAE; unsupervised)
• Vector size normaliser

Instance preprocessors

• Fuzzy Rough Prototype Selection (FRPS)

Other

• array functions
• dispersion measures
• location measures
• nearest neighbour search methods
• parametrisations
• t_norms
• transformations
• vector size measures
• weights

Documentation

The documentation is located here.

Dependencies

fuzzy-rough-learn requires python 3.7+ and the following packages:

• scipy >= 1.1.0
• numpy >=1.17.0
• scikit-learn >=0.24.0

In addition, some algorithms require optional dependencies:

• eif >= 2.0.0 (EIF)
• tensorflow >= 2.2.0 (SAE)

Citing fuzzy-rough-learn

If you use or refer to fuzzy-rough-learn in a scientific publication, please cite this paper:

Lenz OU, Cornelis C, Peralta D (2022).
fuzzy-rough-learn 0.2: a Python library for fuzzy rough set algorithms and one-class classification.
FUZZ-IEEE 2022: Proceedings of the IEEE International Conference on Fuzzy Systems.
doi: 10.1109/FUZZ-IEEE55066.2022.9882778

Bibtex entry:

@inproceedings{lenz22fuzzyroughlearn,
  title={{f}uzzy-rough-learn 0.2: a {P}ython library for fuzzy rough set algorithms and one-class classification},
  author={Lenz, Oliver Urs and Cornelis, Chris and Peralta, Daniel},
  booktitle={{FUZZ-IEEE} 2022: Proceedings of the IEEE International Conference on Fuzzy Systems},
  year={2022},
  publisher={IEEE},
}