Python SKLearn: How to Get Feature Names After One

问题:

I would like to get the feature names of a data set after it has been transformed by SKLearn OneHotEncoder.

In active_features_ attribute in OneHotEncoder one can see a very good explanation how the attributes n_values_, feature_indices_ and active_features_ get filled after transform() was executed.

My question is:

For e.g. DataFrame based input data:

data = pd.DataFrame({"a": [0, 1, 2,0], "b": [0,1,4, 5], "c":[0,1,4, 5]}).as_matrix()

How does the code look like to get from the original feature names a, b and c to a list of the transformed feature names (like e.g:

a-0,a-1, a-2, b-0, b-1, b-2, b-3, c-0, c-1, c-2, c-3

or

a-0,a-1, a-2, b-0, b-1, b-2, b-3, b-4, b-5, b-6, b-7, b-8

or anything that helps to see the assignment of encoded columns to the original columns).

Background: I would like to see the feature importances of some of the algorithms to get a feeling for which feature have the most effect on the algorithm used.

回答1:

You can use pd.get_dummies():

pd.get_dummies(data["a"],prefix="a")

will give you:

    a_0 a_1 a_2
0   1   0   0
1   0   1   0
2   0   0   1
3   1   0   0

which can automatically generates the column names. You can apply this to all your columns and then get the columns names. No need to convert them to a numpy matrix.

So with:

df = pd.DataFrame({"a": [0, 1, 2,0], "b": [0,1,4, 5], "c":[0,1,4, 5]})
data = df.as_matrix()

the solution looks like:

columns = df.columns
my_result = pd.DataFrame()
temp = pd.DataFrame()
for runner in columns:
    temp = pd.get_dummies(df[runner], prefix=runner)
    my_result[temp.columns] = temp
print(my_result.columns)

>>Index(['a_0', 'a_1', 'a_2', 'b_0', 'b_1', 'b_4', 'b_5', 'c_0', 'c_1', 'c_4',
       'c_5'],
      dtype='object')

回答2:

If I understand correctly you can use feature_indices_ to identify which columns correspond to which feature.

e.g.

import pandas as pd
from sklearn.preprocessing import OneHotEncoder
data = pd.DataFrame({"a": [0, 1, 2,0], "b": [0,1,4, 5], "c":[0,1,4, 5]}).as_matrix()
ohe = OneHotEncoder(sparse=False)
ohe_fitted = ohe.fit_transform(data)
print(ohe_fitted)
print(ohe.feature_indices_) # [ 0  3  9 15]

From the above feature_indices_ we know if we spliced the OneHotEncoded data from 0:3 we would get the features corresponding to the first column in data like so:

print(ohe_fitted[:,0:3])

Each column in the spliced data represents a value in the first feature. The first column is 0, the second 1 and the third column is 2. To illustrate this on the spliced data, the column labels would look like:

  a_0 a_1 a_2
[[ 1.  0.  0.]
 [ 0.  1.  0.]
 [ 0.  0.  1.]
 [ 1.  0.  0.]]

Note that features are sorted first before they are encoded.

回答3:

You can do that with the open source package feature-engine:

import pandas as pd
from sklearn.model_selection import train_test_split
from feature_engine.categorical_encoders import OneHotCategoricalEncoder

# load titanic data from openML
pd.read_csv('https://www.openml.org/data/get_csv/16826755/phpMYEkMl')

# divide into train and test
X_train, X_test, y_train, y_test = train_test_split(
    data[['sex', 'embarked']],  # predictors for this example
    data['survived'],  # target
    test_size=0.3,  # percentage of obs in test set
    random_state=0)  # seed to ensure reproducibility

ohe_enc = OneHotCategoricalEncoder(
    top_categories=None,
    variables=['sex', 'embarked'],
    drop_last=True)

ohe_enc.fit(X_train)

X_train = ohe_enc.transform(X_train)
X_test = ohe_enc.transform(X_test)

X_train.head()

You should see this output returned:

   sex_female  embarked_S  embarked_C  embarked_Q
501            1           1           0           0
588            1           1           0           0
402            1           0           1           0
1193           0           0           0           1
686            1           0           0           1

More details about feature engine here:

https://www.trainindata.com/feature-engine

https://github.com/solegalli/feature_engine

https://feature-engine.readthedocs.io/en/latest/

回答4:

There is a OneHotEncoder that does all the work for you.

Package sksurv has a OneHotEncoder that will return a pandas Dataframe with all the column names set-up for you. Check it out. Make sure you set-up an environment to play with the encoder to ensure it doesn't break your current environment. This encoder saved me a lot of time and effort.

scikit-suvival GitHub

OneHotEncoder Documentation

回答5:

OneHotEncoder now has a method get_feature_names. You can use input_features=data.columns to match to the training data.