How to standardize features with scikit-learn

Numeric model features often arrive in different units, such as ages, income amounts, and activity counts. StandardScaler in scikit-learn puts each numeric column on a zero-mean, unit-variance scale so scale-sensitive estimators do not let the largest-unit feature dominate model fitting.

StandardScaler learns the mean and standard deviation during fit() and stores them in mean_ and scale_. Fit the scaler on training rows only, then reuse transform() for validation, test, and future rows so evaluation data does not change the training statistics.

Standardization is a preprocessing choice for numeric columns, not a replacement for handling outliers or categorical values. Extreme values affect the mean and standard deviation, while mixed tabular data usually belongs in a ColumnTransformer or Pipeline so every preprocessing step is fitted with the estimator consistently.

1. Create standardize_features.py with training rows, one held-out row, a scaler, and a small classifier.

   standardize_features.py

   import numpy as np
   import sklearn
   from sklearn.linear_model import LogisticRegression
   from sklearn.preprocessing import StandardScaler
    
    
   feature_names = ["age_years", "annual_income", "monthly_visits"]
    
   X_train = np.array(
       [
           [22, 38000, 4],
           [25, 42000, 5],
           [47, 88000, 7],
           [52, 92000, 8],
           [31, 58000, 6],
           [45, 76000, 7],
           [28, 54000, 5],
           [39, 69000, 6],
       ],
       dtype=float,
   )
   y_train = np.array([0, 0, 1, 1, 0, 1, 0, 1])
    
   X_holdout = np.array([[34, 62000, 6]], dtype=float)
    
   scaler = StandardScaler()
   X_train_scaled = scaler.fit_transform(X_train)
   X_holdout_scaled = scaler.transform(X_holdout)
    
   model = LogisticRegression(max_iter=1000, random_state=0)
   model.fit(X_train_scaled, y_train)
   prediction = model.predict(X_holdout_scaled)
    
    
   def display_zero(value):
       return 0.0 if abs(value) < 5e-12 else value
    
    
   print(f"scikit-learn {sklearn.__version__}")
   print("Scaler learned on training rows only:")
   print(f"- rows_seen: {int(scaler.n_samples_seen_)}")
   print(
       "- mean_: "
       + ", ".join(
           f"{name}={value:.2f}" for name, value in zip(feature_names, scaler.mean_)
       )
   )
   print(
       "- scale_: "
       + ", ".join(
           f"{name}={value:.2f}" for name, value in zip(feature_names, scaler.scale_)
       )
   )
   print()
   print("Scaled training columns:")
   for name, mean, std in zip(
       feature_names, X_train_scaled.mean(axis=0), X_train_scaled.std(axis=0)
   ):
       print(f"- {name}: mean={display_zero(mean):.6f}, std={std:.6f}")
   print()
   print("Scaled holdout row:")
   for name, value in zip(feature_names, X_holdout_scaled[0]):
       print(f"- {name}: {value:.3f}")
   print()
   print(f"Holdout prediction from scaled features: {prediction.tolist()}")

   The scaler uses fit_transform() only on X_train. The held-out row uses transform() so its values are scaled with the training means and standard deviations.

2. Run the standardization script.

   $ python standardize_features.py
   scikit-learn 1.9.0
   Scaler learned on training rows only:
   - rows_seen: 8
   - mean_: age_years=36.12, annual_income=64625.00, monthly_visits=6.00
   - scale_: age_years=10.45, annual_income=18781.22, monthly_visits=1.22
   
   Scaled training columns:
   - age_years: mean=0.000000, std=1.000000
   - annual_income: mean=0.000000, std=1.000000
   - monthly_visits: mean=0.000000, std=1.000000
   
   Scaled holdout row:
   - age_years: -0.203
   - annual_income: -0.140
   - monthly_visits: 0.000
   
   Holdout prediction from scaled features: [0]

3. Confirm the fitted scaler learned statistics from the eight training rows.

   Scaler learned on training rows only:
   - rows_seen: 8
   - mean_: age_years=36.12, annual_income=64625.00, monthly_visits=6.00
   - scale_: age_years=10.45, annual_income=18781.22, monthly_visits=1.22

4. Check that each scaled training column is centered and scaled.

   Scaled training columns:
   - age_years: mean=0.000000, std=1.000000
   - annual_income: mean=0.000000, std=1.000000
   - monthly_visits: mean=0.000000, std=1.000000

   StandardScaler uses the fitted training statistics for each feature, so the transformed training columns print zero means and unit standard deviations.

5. Verify the held-out row was transformed without refitting the scaler.

   Scaled holdout row:
   - age_years: -0.203
   - annual_income: -0.140
   - monthly_visits: 0.000
   
   Holdout prediction from scaled features: [0]

   For cross-validation or production prediction, place StandardScaler inside a Pipeline with the estimator so each split fits preprocessing only on its own training fold.

6. Remove the scratch file when the smoke check is complete.

   $ rm standardize_features.py