scikit-learn-contrib/category_encoders
Error handling in inverse_transform is broken
Closed
#190 opened on May 7, 2019
bughelp wanted
Repository metrics
- Stars
- (2,322 stars)
- PR merge metrics
- (Avg merge 4d 10h) (2 merged PRs in 30d)
Description
Inverse_transform should ideally handle absence of the columns dropped because of drop_invariant=True. But if it is not possible, inverse_transform() should at least return the correct error message instead of just crashing.
Example in form of a parameterized unit test:
def test_inverse_wrong_feature_count_wit_drop_invariant(self):
x = [['A', 'B', 'C'], ['D', 'E', 'C'], ['F', 'G', 'C']] # the last column is constant
for encoder_name in {'BaseNEncoder', 'BinaryEncoder', 'OrdinalEncoder', 'OneHotEncoder'}:
with self.subTest(encoder_name=encoder_name):
enc = getattr(encoders, encoder_name)(drop_invariant=True)
transformed = enc.fit_transform(x)
# run inverse_transform() and check the raised exception text
with self.assertRaises(ValueError) as cm:
enc.inverse_transform(transformed)
self.assertTrue(str(cm.exception).startswith('Unexpected input dimension'))