Installation

Preferrably, install from GitHub source. The use of a virtual environment is strongly advised.

git clone https://github.com/jafetgado/resreg.git
cd resreg
pip install -r requirements.txt
python setup.py install

Or, install with pip (less preferred)

pip install resreg

Prerequisites

1. Python 3

2. Numpy

3. Scipy

4. Pandas

5. Scikit-learn

Usage

A regression dataset (X, y) can be resampled to mitigate the imbalance in the distribution with any of six strategies: random oversampling, random undersampling, SMOTER, Gaussian noise, WERCS, or Rebagg.

1. Random oversampling (RO): randomly oversample rare values selected by the user via a relevance function.

2. Random undersampling (RU): randomly undersample abundant values.

3. SMOTER: randomly undersample abundant values; oversample rare values by interpolation between nearest neighbors.

4. Gaussian noise (GN): randomly undersample abundant values; oversample rare values by adding Gaussian noise.

5. WERCS: resample the dataset by selecting instances using user-specified relevance values as weights.

6. REBAGG: Train an ensemble of Scikit-learn base learners on independently resampled bootstrap subsets of the dataset.

See the tutorial for more details.

Examples

import resreg
from sklearn.metrics import train_test_split
from sklearn.metrics import RandomForestRegressor

# Split dataset to training and testing sets
X_train, X_test, y_train, y_test = resreg.train_test_split(X, y, test_size=0.25)

# Resample training set with random oversampling
relevance = resreg.sigmoid_relevance(y, cl=None, ch=np.percentile(y, 90))
X_train, y_train = resreg.random_oversampling(X_train, y_train, relevance, relevance_threshold=0.5,
                                              over='balance')

# Fit regressor to resampled training set
reg = RandomForestRegressor()
reg.fit(X_train, y_train)
y_pred = reg.predict(X_train, y_train)