sklearn LabelBinarizer choose which label (string) is the positive

From my tests using sklearn 0.24.1 it seems that LabelBinarizer orders the strings in the data to be binarized and chooses the last value as "positive". For instance:

from sklearn import preprocessing
lb = preprocessing.LabelBinarizer()
print(lb.fit_transform(['a', 'd', 'd', 'a']))
lb = preprocessing.LabelBinarizer()
print(lb.fit_transform(['e', 'd', 'd', 'e']))

In the first case 'd' is chosen to be '1' and in the second one 'e' is chosen to be '1'. I haven't been able to find a parameter or a way to specify with string (e.g. 'a' in the first case or 'd' in the second) should be the '1'. I have even tried ChatGPT, which showed up weird things.

In any case, the following did the trick:

print(-1 * (lb.fit_transform(['e', 'd', 'd', 'e']) -1))

If you're dealing with fixed classes, you can use label_binarize() instead:

print(preprocessing.label_binarize(['e', 'd', 'd', 'e'], classes = ['e', 'd']))

Worth noting this also allows for custom order label encoding if you apply argmax() to it.

The order of class labels is stored in LabelBinarizer.classes_ attribute.

After fit(X), it should be totally safe to re-order its elements if it doesn't agree with your requirements.

from sklearn import preprocessing
import numpy
lb = preprocessing.LabelBinarizer()
print(lb.fit_transform(['a', 'd', 'd', 'a']))
# Take a look at the learned order
print(lb.classes_)
# Invert it
lb.classes_ = numpy.asarray(["d", "a"])
print(lb.transform(['a', 'd', 'd', 'a']))