Impute categorical missing values in scikit-learn

I've got pandas data with some columns of text type. There are some NaN values along with these text columns. What I'm trying to do is to impute those NaN's by sklearn.preprocessing.Imputer (replacing NaN by the most frequent value). The problem is in implementation. Suppose there is a Pandas dataframe df with 30 columns, 10 of which are of categorical nature. Once I run:

from sklearn.preprocessing import Imputer
imp = Imputer(missing_values='NaN', strategy='most_frequent', axis=0)
imp.fit(df) 

Python generates an error: 'could not convert string to float: 'run1'', where 'run1' is an ordinary (non-missing) value from the first column with categorical data.

Any help would be very welcome

Solution 1:

To use mean values for numeric columns and the most frequent value for non-numeric columns you could do something like this. You could further distinguish between integers and floats. I guess it might make sense to use the median for integer columns instead.

import pandas as pd
import numpy as np

from sklearn.base import TransformerMixin

class DataFrameImputer(TransformerMixin):

    def __init__(self):
        """Impute missing values.

        Columns of dtype object are imputed with the most frequent value 
        in column.

        Columns of other types are imputed with mean of column.

        """
    def fit(self, X, y=None):

        self.fill = pd.Series([X[c].value_counts().index[0]
            if X[c].dtype == np.dtype('O') else X[c].mean() for c in X],
            index=X.columns)

        return self

    def transform(self, X, y=None):
        return X.fillna(self.fill)

data = [
    ['a', 1, 2],
    ['b', 1, 1],
    ['b', 2, 2],
    [np.nan, np.nan, np.nan]
]

X = pd.DataFrame(data)
xt = DataFrameImputer().fit_transform(X)

print('before...')
print(X)
print('after...')
print(xt)

which prints,

before...
     0   1   2
0    a   1   2
1    b   1   1
2    b   2   2
3  NaN NaN NaN
after...
   0         1         2
0  a  1.000000  2.000000
1  b  1.000000  1.000000
2  b  2.000000  2.000000
3  b  1.333333  1.666667

Solution 2:

You can use sklearn_pandas.CategoricalImputer for the categorical columns. Details:

First, (from the book Hands-On Machine Learning with Scikit-Learn and TensorFlow) you can have subpipelines for numerical and string/categorical features, where each subpipeline's first transformer is a selector that takes a list of column names (and the full_pipeline.fit_transform() takes a pandas DataFrame):

class DataFrameSelector(BaseEstimator, TransformerMixin):
    def __init__(self, attribute_names):
        self.attribute_names = attribute_names
    def fit(self, X, y=None):
        return self
    def transform(self, X):
        return X[self.attribute_names].values

You can then combine these sub pipelines with sklearn.pipeline.FeatureUnion, for example:

full_pipeline = FeatureUnion(transformer_list=[
    ("num_pipeline", num_pipeline),
    ("cat_pipeline", cat_pipeline)
])

Now, in the num_pipeline you can simply use sklearn.preprocessing.Imputer(), but in the cat_pipline, you can use CategoricalImputer() from the sklearn_pandas package.

note: sklearn-pandas package can be installed with pip install sklearn-pandas, but it is imported as import sklearn_pandas

Solution 3:

There is a package sklearn-pandas which has option for imputation for categorical variable https://github.com/scikit-learn-contrib/sklearn-pandas#categoricalimputer

>>> from sklearn_pandas import CategoricalImputer
>>> data = np.array(['a', 'b', 'b', np.nan], dtype=object)
>>> imputer = CategoricalImputer()
>>> imputer.fit_transform(data)
array(['a', 'b', 'b', 'b'], dtype=object)