Undersampling

In the case where method = 'majority', the random undersampling consists in randomly removing points which have the biggest class as their majority class. In other words, let \(M\) be the biggest class of the dataset, i.e. \(M = \mbox{argmax}_j (V_j)\), where \(V\) is the vector of the sum of the classes (as defined in Introduction). Then, all the points where the softmax of their label is \(M\) are considered as points that can be removed. Some points are randomly removed one by one until \(V_M \leq V_{M_2}\), where \(M_2\) is the second biggest class of the dataset.

In the case where method is an integer b, then the b majority classes are similarly undersampled until the sum of each of them is \(\leq V_{M_{b+1}}\).

As the function smote_cd.random_undersampling only returns the indexes of the points to be removed, you will have to remove these indexes from your original dataset. For instance, you can do the following:

indexes_to_remove = smote_cd.random_undersampling(y)
y_undersampled=np.delete(y,indexes_to_remove,axis=0)
X_undersampled=np.delete(X,indexes_to_remove,axis=0)

Practical examples are available on the page smote_cd.random_undersampling().