Considerations about OneHotEncoder


When I was answering questions on StackOverflow, I found this question interesting. I also found that there are lots of questions about sklearn.preprocessing.OneHotEncoder, which seemed to be pretty confusing, so I reprinted my answer here.

These following informations might be helpful:

  1. The type of some of the objects:
    • data[feature]: pandas.Series
    • data[feature].values: numpy.ndarray
  2. You can reshape a numpy.ndarray but not a pandas.Series, so you need to use .values to get a numpy.ndarray
  3. When you assign a numpy.ndarray to data[feature], automatic type conversion occurs, so data[feature] = data[feature].values.reshape(-1, 1) doesn’t seem to do anything.
  4. fit_transform takes an array-like(Need to be a 2D array, e.g. pandas.DataFrame or numpy.ndarray) object as argument because sklearn.preprocessing.OneHotEncoder is designed to fit/transform multiple features at the same time, input pandas.Series(1D array) will cause error.
  5. fit_transform will return sparse matrix(or 2-d array), assign it to a pandas.Series may cause a disaster.

(Not Recommended) If you insist on processing one feature after another:

for f in categorical_feats:
    encoder = OneHotEncoder()
    tmp_ohe_data = pd.DataFrame(
        encoder.fit_transform(data[f].values.reshape(-1, 1)).toarray(),
    data = pd.concat([ohe_data, data], axis=1).drop([feature], axis=1)

I Recommended do encoding like this:

encoder = OneHotEncoder()

ohe_data = pd.DataFrame(
data = pd.concat([ohe_data, data], axis=1).drop(categorical_feats, axis=1)

pandas.get_dummies is also a good choice, but the downside is that, you can’t pickle an encoder for later use.

for f in categorical_feats:
    dummies = pd.get_dummies(data[f], prefix=f)
    data = data.join(dummies)
本文总字数: 1611