sklearn.impute

impute

What to do about the blanks.

1 Get data Something to practise on

2 Split & tune An honest test set

3 Prepare Into numbers a model accepts

4 Compose One object, no leakage

5 Model The algorithm itself

6 Unsupervised No labels available

7 Measure Is it any good, and why

8 Utilities Glue and extension

What it is for

Filling missing values. Worth pausing on the choice, because in financial data the fact that a field is blank is very often signal in its own right — an unanswered question on an application form means something different from a zero.

`SimpleImputer` common	Mean, median, most frequent or constant. Start here
`KNNImputer`	Fill from similar rows
`IterativeImputer`	Model each column from the others. Powerful, slow, needs the experimental import
`MissingIndicator` bank	Adds a boolean 'was missing' column. Often the missingness is the signal

Worked examples 2

Runnable, with the arguments that change the outcome called out. Output shapes are shown as comments where they help.

Impute, and keep the fact that it was missing

In application data, a blank field usually means something. Throwing that signal away to fill the value is a real loss.

from sklearn.impute import SimpleImputer
import numpy as np

X = np.array([[45_000, 3.0],
              [np.nan,  1.0],
              [62_000,  np.nan],
              [51_000,  2.0]])

# add_indicator appends a boolean column per feature that had NaNs
imp = SimpleImputer(strategy="median", add_indicator=True)
out = imp.fit_transform(X)

print(out)
# [[4.5e4 3.0 0.0 0.0]
#  [5.1e4 1.0 1.0 0.0]     <- income was missing
#  [6.2e4 2.5 0.0 1.0]     <- dependants was missing
#  [5.1e4 2.0 0.0 0.0]]

# The model can now learn "applications with no stated income behave
# differently", which is often more predictive than the imputed value.

When you can skip imputation entirely

HistGradientBoosting handles NaN natively — it learns which side of each split missing values should go. Often better than any imputation you would choose.

from sklearn.ensemble import HistGradientBoostingClassifier
import numpy as np

X = np.array([[45_000, 3.0], [np.nan, 1.0], [62_000, np.nan], [51_000, 2.0]])
y = np.array([0, 1, 0, 1])

clf = HistGradientBoostingClassifier(max_iter=50)
clf.fit(X, y)          # no imputer, no error
print(clf.predict(X))

# Also true of XGBoost and LightGBM. NOT true of LogisticRegression,
# SVC, KNeighbors or RandomForest, all of which raise on NaN.

Worth knowing

HistGradientBoosting handles missing values natively — no imputation step needed at all.