OneHotEncoder should handle unknown values more gracefully #2169
Description
As reported over at SO, it would be nice if OneHotEncoder could handle unknown values for categorical features at transform time. Currently, the following throws an exception:
>>> from sklearn.preprocessing import OneHotEncoder
>>> oh = OneHotEncoder().fit([[0]])
>>> oh.transform([[1]])
Traceback (most recent call last):
File "<ipython-input-17-54f21ed7c610>", line 1, in <module>
oh.transform([[1]])
File "/home/lars/src/scikit-learn/sklearn/preprocessing.py", line 878, in transform
self.categorical_features, copy=True)
File "/home/lars/src/scikit-learn/sklearn/preprocessing.py", line 662, in _transform_selected
return transform(X)
File "/home/lars/src/scikit-learn/sklearn/preprocessing.py", line 851, in _transform
raise ValueError("Feature out of bounds. Try setting n_values.")
ValueError: Feature out of bounds. Try setting n_values.
I personally find this a bit strict, and would expect at most a warning and an appropriate number of zero columns for an unknown value.
This would be consistent with DictVectorizer and CountVectorizer, which ignore whatever features were not in their training set.