resampling

• There is a dedicated Python package for that called imbalanced-learn

• in practice resampling only on the training data set - This is in order to avoid data leakage.

undersampling

# mX is the feature matrix
# vY is the target vector
len_feat_1 = mX0.shape[0]       ## assume large ; for example 1000
len_feat_0 = mX1.shape[0]       ## assume small ; for example 50

## create a random index from 1-1000 with 50 samples
vIdx0UnderSample = np.random.choice(len_feat_1, len_feat_0, replace = False)

## filter by index to get the undersampled feature matrix in size of 50
mX0UnderSample   = mX0[vIdx0UnderSample]

## create the new feature matrix with both classes in len 50 ; use the small one as is
mXS = np.vstack((mX0UnderSample, mX1))

## build the label vector accordingly
vYS = np.concatenate((np.zeros(mX0UnderSample.shape[0], dtype = vY.dtype), np.ones(mX1.shape[0], dtype = vY.dtype)))

oversampling