sklearn.preprocessing.OneHotEncoder: using drop and handle_unknown='ignore'

Question

I have some pandas.Series – s, below – that I want to one-hot-encode. I've found through research that the 'b' level is not important for my predictive modeling task. I can exclude it from my analysis like so:

import pandas as pd
from sklearn.preprocessing import OneHotEncoder

s = pd.Series(['a', 'b', 'c']).values.reshape(-1, 1)

enc = OneHotEncoder(drop=['b'], sparse=False, handle_unknown='error')
enc.fit_transform(s)
# array([[1., 0.],
#        [0., 0.],
#        [0., 1.]])
enc.get_feature_names()
# array(['x0_a', 'x0_c'], dtype=object)

But when I go to transform a new series, one containing both 'b' and a new level, 'd', I get an error:

new_s = pd.Series(['a', 'b', 'c', 'd']).values.reshape(-1, 1)
enc.transform(new_s)

Traceback (most recent call last): File "", line 1, in File "/Users/user/Documents/assets/envs/data-science/venv/lib/python3.7/site-packages/sklearn/preprocessing/_encoders.py", line 390, in transform X_int, X_mask = self._transform(X, handle_unknown=self.handle_unknown) File "/Users/user/Documents/assets/envs/data-science/venv/lib/python3.7/site-packages/sklearn/preprocessing/_encoders.py", line 124, in _transform raise ValueError(msg) ValueError: Found unknown categories ['d'] in column 0 during transform

This is to be expected since I set handle_unknown='error' above. However, I'd like to completely ignore all classes except for ['a', 'c'] in both the fitting and subsequent transforming steps. I tried this:

enc = OneHotEncoder(drop=['b'], sparse=False, handle_unknown='ignore')
enc.fit_transform(s)
enc.transform(new_s)

Traceback (most recent call last): File "", line 1, in File "/Users/user/Documents/assets/envs/data-science/venv/lib/python3.7/site-packages/sklearn/preprocessing/_encoders.py", line 371, in fit_transform self._validate_keywords() File "/Users/user/Documents/assets/envs/data-science/venv/lib/python3.7/site-packages/sklearn/preprocessing/_encoders.py", line 289, in _validate_keywords "handle_unknown must be 'error' when the drop parameter is " ValueError: handle_unknown must be 'error' when the drop parameter is specified, as both would create categories that are all zero.

It seems this pattern is not supported in scikit-learn. Does anyone know a scikit-learn-compatible pattern to accomplish this task?

Answer 1

You could also approach this using the following:

class IgnorantOneHotEncoder(OneHotEncoder):
    def transform(self, X, y=None):
        try:
            return super().transform(X)
        except ValueError as e:
            if 'Found unknown categories' in str(e):
                X = np.copy(X)
                # Keep track of indices corresponding to unknown categories
                unknown_categories_mask = ~np.isin(X, self.categories_[0]).ravel()
                # Overwrite the unknown categories in the input matrix, X, with the first known category
                X[unknown_categories_mask] = self.categories_[0][0]
                # Transform X, whose categories are all known now
                X = super().transform(X)
                # Overwrite originally unknown-category records with 0 to indicate
                # absence of any value for any category for that feature
                X[unknown_categories_mask, 0] = 0
                return X
            else:
                raise

Try it out:

>>> ienc = IgnorantOneHotEncoder(sparse=False)
>>> ienc.fit(s)
IgnorantOneHotEncoder(sparse=False)
>>> ienc.transform(s)
array([[1., 0., 0.],
       [0., 1., 0.],
       [0., 0., 1.]])
>>> ienc.transform(new_s)
array([[1., 0., 0.],
       [0., 1., 0.],
       [0., 0., 1.],
       [0., 0., 0.]])