Why Scikit-learn Is Still Relevant in 2026
Deep learning gets the headlines, but for tabular data — the majority of business ML use cases — scikit-learn and gradient boosting still win. Sklearn gives you consistent APIs, excellent documentation, and decades of battle-tested algorithms in one package.
The rule of thumb: if your dataset has fewer than 1M rows and structured features, start with sklearn. Always use sklearn for your baseline before trying anything more complex.
What Is New in v1.4+
HDBSCAN clustering (v1.3+): hierarchical density-based clustering, better than DBSCAN for varying-density clusters.
from sklearn.cluster import HDBSCAN
hdbscan = HDBSCAN(min_cluster_size=10, min_samples=5)
labels = hdbscan.fit_predict(X)
TunedThresholdClassifierCV (v1.5+): automatically tunes the decision threshold for binary classifiers to optimize a business metric.
from sklearn.calibration import TunedThresholdClassifierCV
from sklearn.ensemble import GradientBoostingClassifier
clf = GradientBoostingClassifier()
tuned_clf = TunedThresholdClassifierCV(
clf, scoring="f1", cv=5
)
tuned_clf.fit(X_train, y_train)
print(tuned_clf.best_threshold_) # e.g. 0.37 instead of default 0.5
The Essential Pipeline Pattern
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import StandardScaler, OneHotEncoder
from sklearn.compose import ColumnTransformer
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import cross_val_score
numeric_features = ["age", "income", "tenure"]
categorical_features = ["country", "plan", "device"]
preprocessor = ColumnTransformer([
("num", StandardScaler(), numeric_features),
("cat", OneHotEncoder(handle_unknown="ignore", sparse_output=False), categorical_features),
])
pipeline = Pipeline([
("preprocessor", preprocessor),
("classifier", RandomForestClassifier(n_estimators=100, random_state=42)),
])
scores = cross_val_score(pipeline, X, y, cv=5, scoring="roc_auc")
print(f"AUC: {scores.mean():.3f} ± {scores.std():.3f}")
Pipelines prevent data leakage — the scaler fits on training data only in each CV fold.
Hyperparameter Search
from sklearn.model_selection import GridSearchCV
param_grid = {
"classifier__n_estimators": [100, 300],
"classifier__max_depth": [None, 5, 10],
"classifier__min_samples_split": [2, 5],
}
search = GridSearchCV(pipeline, param_grid, cv=5, scoring="roc_auc", n_jobs=-1)
search.fit(X_train, y_train)
print(search.best_params_)
For larger search spaces, use Optuna instead of GridSearchCV.
SHAP for Model Explanation
import shap
explainer = shap.TreeExplainer(search.best_estimator_["classifier"])
shap_values = explainer.shap_values(X_test)
shap.summary_plot(shap_values, X_test, feature_names=numeric_features + categorical_features)
Sklearn vs XGBoost vs LightGBM for Tabular
Sklearn's GradientBoostingClassifier is slower than XGBoost/LightGBM but good for baselines. For competition-level performance on tabular data, XGBoost or LightGBM outperform sklearn's tree methods. The sklearn API is consistent across all: fit(), predict(), predict_proba().
Resources: Scikit-learn, GitHub, v1.4 changelog.
