scikit-learn-1-9-0

name	scikit-learn-1-9-0
description	Comprehensive guide to scikit-learn 1.9.0 — the Python machine learning library. Use when working with ML models, pipelines, preprocessing, model selection, metrics, or any data science task using scikit-learn. Covers classification, regression, clustering, dimensionality reduction, ensemble methods, hyperparameter tuning, cross-validation, feature engineering, and more. Trigger on: sklearn, scikit-learn, machine learning, ML pipeline, model training, cross-validation, GridSearchCV, Random Forest, SVM, logistic regression, PCA, KMeans, train_test_split, metrics.
metadata	{"tags":["ml","data-science","python"]}

scikit-learn 1.9.0

Overview

Scikit-learn 1.9.0 (June 2026) is the Python machine learning library built on NumPy, SciPy, and joblib. All estimators follow a unified API: fit(), predict() (predictors), transform() (transformers). The library ships with ~130+ estimators across classification, regression, clustering, dimensionality reduction, preprocessing, model selection, and metrics.

Key 1.9 additions:

• Callback API (sklearn.callback) — progress bars and scoring monitors during fitting
• FrozenEstimator (sklearn.frozen) — wrap fitted estimators to prevent re-fitting in pipelines
• sparse_interface config — control sparse matrix vs array output
• narwhals dependency — improved pandas/Polars dataframe support
• metric_at_thresholds() for computing metrics across all thresholds
• Tree models now handle missing values (NaN) natively in dense data

Dependencies: Python 3.11+, NumPy >= 1.24.1, SciPy >= 1.10.0, joblib >= 1.4.0, narwhals >= 2.0.1, threadpoolctl >= 3.5.0.

Usage

Estimator Pattern (all models)

from sklearn.ensemble import RandomForestClassifier

model = RandomForestClassifier(n_estimators=100, random_state=42)
model.fit(X_train, y_train)          # Train
y_pred = model.predict(X_test)       # Predict
y_proba = model.predict_proba(X_test)  # Probabilities (classifiers)
score = model.score(X_test, y_test)    # Default metric

Pipeline Pattern (recommended for production)

from sklearn.pipeline import Pipeline
from sklearn.preprocessing import StandardScaler
from sklearn.ensemble import RandomForestClassifier

pipe = Pipeline([
    ("scaler", StandardScaler()),
    ("clf", RandomForestClassifier(random_state=42)),
])
pipe.fit(X_train, y_train)
pipe.predict(X_test)

Model Selection Pattern

from sklearn.model_selection import cross_validate, GridSearchCV

# Quick evaluation
cv_results = cross_validate(pipe, X, y, cv=5,
                            scoring=["accuracy", "f1"])

# Hyperparameter tuning
search = GridSearchCV(
    pipe,
    param_grid={"clf__n_estimators": [50, 100, 200]},
    cv=5, scoring="f1"
)
search.fit(X_train, y_train)
search.best_params_

Column-Wise Preprocessing (mixed feature types)

from sklearn.compose import ColumnTransformer
from sklearn.preprocessing import OneHotEncoder, StandardScaler

preprocessor = ColumnTransformer([
    ("num", StandardScaler(), numeric_cols),
    ("cat", OneHotEncoder(handle_unknown="infrequent_if_exist"), categorical_cols),
])

Callbacks (new in 1.9)

from sklearn.callback import ProgressBar, ScoringMonitor

# During fitting of supported estimators
model.fit(X, y, callbacks=[ProgressBar()])

Gotchas

• load_boston removed since 1.2 — use fetch_california_housing or fetch_openml(name="house_prices") instead. The Boston dataset had ethical issues (racially biased variable).
• SVC(probability=True) deprecated in 1.9 — not thread-safe. Use CalibratedClassifierCV(svc, ensemble=False) for probability estimates instead.
• criterion="friedman_mse" deprecated in tree regressors — it was buggy and identical to "squared_error". Use "squared_error" instead.
• TargetEncoder(shuffle=...) deprecated — pass a CV generator via cv argument instead.
• LogisticRegressionCV(scoring) default changing — will switch from None (accuracy) to "neg_log_loss" in v1.11. Explicitly set scoring now.
• n_alphas deprecated in lasso_path/enet_path — pass alphas directly as int or array-like.
• sample_weight with all zeros raises ValueError — validate weights before passing.
• Pipeline parameter access uses double underscore — pipe.set_params(clf__C=0.1) not pipe.clf.C = 0.1.
• clone() resets fitted state — use sklearn.base.clone() to get a fresh unfitted copy with same hyperparameters.
• RandomForest/ExtraTrees max_samples behavior changed — when float, interpreted as fraction of sample_weight.sum(), not X.shape[0]. Integer max_samples > n_samples now allowed.
• FrozenEstimator.__sklearn_clone__ returns self — cloning a frozen estimator doesn't clone the inner estimator. This is intentional (prevents re-fitting).
• Sparse interface config defaults to "spmatrix" — will change to "sparray" in future releases. Set explicitly if you need sparse arrays now.

References

• 01-core-api — BaseEstimator, mixins, Pipeline, ColumnTransformer, clone, config
• 02-classifiers — All classification estimators and their key parameters
• 03-regressors — All regression estimators and their key parameters
• 04-clustering — Clustering algorithms: KMeans, DBSCAN, Agglomerative, etc.
• 05-preprocessing — Scalers, encoders, imputers, transformers
• 06-model-selection — CV splitters, GridSearchCV, cross_validate, learning curves
• 07-metrics — Classification/regression metrics, scorers, display classes
• 08-feature-selection — RFE, SelectKBest, VarianceThreshold, mutual info
• 09-dimensionality-reduction — PCA, SVD, NMF, manifold methods
• 10-data-io — Built-in datasets, fetch_openml, synthetic data generators
• 11-advanced-topics — Callbacks, metadata routing, FrozenEstimator, experimental features