Closed
Description
Currently, the Imputer works on columns. Let's assume I have a dataset with mixed categorical and numerical data points and missing values and I want to use the pipeline object:
import numpy as np
import scipy.stats
import sklearn.pipeline
import sklearn.preprocessing
import sklearn.ensemble
categorical_data = np.random.randint(0, 5, size=(10, 4))
numerical_data = np.random.randn(10, 4)
data = np.hstack((categorical_data, numerical_data))
# Add missing values
data[0, 1] = np.NaN
data[1, 5] = np.NaN
X = data
y = np.array([0, 0, 0, 0, 0, 1, 1, 1, 1, 1])If I now try to use imputation, I can only choose a single strategy for both data types:
pipeline = sklearn.pipeline.Pipeline((('imp',
sklearn.preprocessing.Imputer(
strategy='most_frequent')),
('ohe',
sklearn.preprocessing.OneHotEncoder(
categorical_features=[0, 1, 2, 3])),
('rf',
sklearn.ensemble.RandomForestClassifier())))which would result in a more or less random value being picked for the missing value of the continuous feature:
Xt, _ = pipeline._pre_transform(X, y)
# Now uses the return value of scipy.stats.mode which makes only very little
# sense for a continuous value
print(scipy.stats.mode(data)[5])
print(Xt.toarray()[1])To overcome this, I propose to add a new attribute to the Imputer which allows to specify the columns to be imputed in order to allow something like this:
pipeline = sklearn.pipeline.Pipeline((('imp_cat',
sklearn.preprocessing.Imputer(
strategy='most_frequent',
columns=[0, 1, 2, 3])),
('imp_num',
sklearn.preprocessing.Imputer(
strategy='median',
columns=[4, 5, 6, 7])),
('ohe',
sklearn.preprocessing.OneHotEncoder(
categorical_features=[0, 1, 2, 3])),
('rf',
sklearn.ensemble.RandomForestClassifier())))I could not find a related issue and am willing to work on this if this is considered worth adding to scikit-learn.