Introduction
Most ML algorithms are fussy eaters. They want numbers only, no missing values, and often features on similar scales. Real data rarely look like that. Preparing the data — often called preprocessing — is where a big share of real project time goes.
A key principle: write the preparation steps as code, not as manual edits. Then you can:
- apply exactly the same steps to the test set and to future data in production;
- try different options easily (e.g. median vs mean filling);
- rerun everything when new data arrive.
In scikit-learn, each preparation step is a transformer: an object that learns something from the training data with fit() (like a median) and applies it with transform(). Many have fit_transform() to do both at once.
First, separate the features from the labels:
PYTHON
from sklearn.datasets import fetch_california_housing
from sklearn.model_selection import train_test_split
df = fetch_california_housing(as_frame=True).frame
train_set, test_set = train_test_split(df, test_size=0.2, random_state=42)
X_train = train_set.drop(columns="MedHouseVal") # features
y_train = train_set["MedHouseVal"].copy() # labels