The three interfaces

Almost every object in scikit-learn is one of three shapes. This uniformity is the library's central design decision and the reason it has lasted.

**Estimator** — learns from data, then predicts.

    model.fit(X, y)            learn
    model.predict(X)           apply
    model.predict_proba(X)     apply, with probabilities (classifiers)
    model.score(X, y)          a quick default metric

**Transformer** — changes data rather than predicting it.

    scaler.fit(X)              learn the parameters, e.g. mean and std
    scaler.transform(X)        apply them
    scaler.fit_transform(X)    both at once - TRAINING DATA ONLY

**Meta-estimator** — wraps other estimators. `Pipeline`, `GridSearchCV`, `ColumnTransformer`, `VotingClassifier`. These expose `fit` and `predict` themselves, so they can go anywhere a plain model can, including inside one another.

Because the interface is uniform, swapping algorithms is a one-line change, and tooling written against it works with anything that follows it — which is why XGBoost, LightGBM and CatBoost all adopted the same method names despite being separate projects by different organisations.

from sklearn.linear_model import LogisticRegression
from sklearn.ensemble import RandomForestClassifier
from xgboost import XGBClassifier          # not part of sklearn

# Identical calls. GridSearchCV and Pipeline neither know nor care
# which library a model came from.
for m in [LogisticRegression(max_iter=1000),
          RandomForestClassifier(n_jobs=-1),
          XGBClassifier()]:
    m.fit(X_train, y_train)
    print(type(m).__name__, m.score(X_test, y_test))