sklearn.impute SimpleImputer：为什么transform（）首先需要fit_transform（）？

Question

sklearn provides transform() method to Apply one-hot encoder.

to use transform() method, fit_transform() is needed before calling transform() method, otherwise

np.array([[1, 1], [2, 1], [3, 2], [np.nan, 2]])
from sklearn.impute import SimpleImputer
my_imputer = SimpleImputer()
my_imputer.transform(df)

error shows up

NotFittedError: This SimpleImputer instance is not fitted yet. Call 'fit' with appropriate arguments before using this method.

calling fit_transform() before transform()

my_imputer.fit_transform(df)
my_imputer.transform(df)

fix this error.

the question is, why does transform() need fit_transform() ?

Answer 1

During fit() the imputer learns about the mean, median etc of the data, which is then applied to the missing values during transform() .

fit_transform() is just a shorthand for combining the two methods. So essentially:

• fit(X, y) :- Learns about the required aspects of the supplied data and returns the new object with the learned parameters. It does not change the supplied data in any way.

• transform() :- Actually transform the supplied data to the new form.

fit_transform(df) is not required to be called before transform. Only fit() is needed to be called. Generally the sequence you described is done with train and test split of data. Something like:

# Combining the learning of parameters from training data and transforming into a single step.
X_train_new = my_imputer.fit_transform(X_train)

# We dont want to learn about test data, only change it according to previously learnt information
X_test_new = my_imputer.transform(X_test)

The above code snippet can be broken into:

# It learns about the data and does nothing else
my_imputer.fit(X_train)

# Calling transform to apply the learnt information on supplied data
X_train_new = my_imputer.transform(X_train)
X_test_new = my_imputer.transform(X_test)