Scikit-Learn (sklearn) is Python's foundational open-source library for statistical modeling and traditional machine learning, built on top of NumPy, SciPy, and Matplotlib. Its architectural strength rests on a strictly unified three-interface design pattern:
- Estimator: Learns model parameters from data via
fit(X, y). - Transformer: Preprocesses and projects features via
transform(X)andfit_transform(X). - Predictor: Makes predictions on unseen samples via
predict(X),predict_proba(X), andscore(X, y).
🔗 Related Cheat Sheets: NumPy Cheat Sheet · Pandas Cheat Sheet · Supervised Learning Algorithms · Data Analysis Methods · Python Cheat Sheet
1. End-to-End Quickstart Example
The following self-contained script walks through loading a dataset, stratified splitting, feature scaling, model fitting, and evaluation:
from sklearn.datasets import load_iris
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler
from sklearn.neighbors import KNeighborsClassifier
from sklearn.metrics import accuracy_score, classification_report
# 1. Load data
iris = load_iris()
X, y = iris.data, iris.target
# 2. Stratified train/test split
X_train, X_test, y_train, y_test = train_test_split(
X, y, test_size=0.2, random_state=42, stratify=y
)
# 3. Feature scaling (fit on training set ONLY to prevent leakage)
scaler = StandardScaler()
X_train_scaled = scaler.fit_transform(X_train)
X_test_scaled = scaler.transform(X_test)
# 4. Fit estimator
knn = KNeighborsClassifier(n_neighbors=5)
knn.fit(X_train_scaled, y_train)
# 5. Predict and evaluate
y_pred = knn.predict(X_test_scaled)
print(f"Accuracy: {accuracy_score(y_test, y_pred):.4f}")
print(classification_report(y_test, y_pred, target_names=iris.target_names))
2. Data Preparation & Splitting
Scikit-Learn expects 2D numerical structures: feature matrix $X$ of shape (n_samples, n_features) and target vector $y$ of shape (n_samples,). Inputs can be NumPy ndarrays, SciPy sparse matrices (scipy.sparse), or Pandas DataFrames/Series.
2.1 Built-in & Synthetic Datasets
from sklearn import datasets
# Small baseline toy datasets (returns Bunch object)
iris = datasets.load_iris() # Multi-class classification
digits = datasets.load_digits() # 8x8 handwritten digit images
diabetes = datasets.load_diabetes() # Regression benchmark
wine = datasets.load_wine() # 3-class chemical analysis
# Real-world benchmark datasets (cached locally)
housing = datasets.fetch_california_housing()
# Synthetic data generation for algorithm stress-testing
X, y = datasets.make_classification(
n_samples=1000, n_features=20, n_informative=15, n_classes=2, random_state=42
)
X_reg, y_reg = datasets.make_regression(n_samples=500, n_features=10, noise=0.1, random_state=42)
X_blob, y_blob = datasets.make_blobs(n_samples=300, centers=3, n_features=2, random_state=42)
2.2 Dataset Splitting (train_test_split)
from sklearn.model_selection import train_test_split
X_train, X_test, y_train, y_test = train_test_split(
X, y,
test_size=0.25, # Proportion of test set (default: 0.25)
random_state=42, # Fixed seed for reproducibility
shuffle=True, # Shuffle samples before splitting
stratify=y # Highly recommended for classification: preserves class ratios
)
3. Feature Engineering & Preprocessing (sklearn.preprocessing & sklearn.impute)
| Transformer | Use Case | Mathematical Transformation / Logic | Key Parameters |
|---|---|---|---|
StandardScaler |
Numerical feature Z-score standardization | $z = \frac{x - \mu}{\sigma}$ (zero mean, unit variance) | with_mean=True, with_std=True |
MinMaxScaler |
Bounding features to a fixed closed interval | $x' = \frac{x - x_{\min}}{x_{\max} - x_{\min}} \cdot (\max - \min) + \min$ | feature_range=(0, 1) |
RobustScaler |
Outlier-resilient feature scaling | Interquartile range scaling: $x' = \frac{x - Q_2}{Q_3 - Q_1}$ | quantile_range=(25.0, 75.0) |
Normalizer |
Sample-wise row vector normalization | Scales each sample to have unit $L_1$, $L_2$, or max norm | norm='l2' ('l1', 'max') |
Binarizer |
Hard thresholding continuous values | $x > \text{threshold} \Rightarrow 1$, else $0$ | threshold=0.0 |
SimpleImputer |
Missing value replacement | Imputes missing values with mean, median, most frequent, or constant | missing_values=np.nan, strategy='mean' |
OneHotEncoder |
Nominal categorical feature encoding | Converts categorical variables into binary dummy columns | drop='first', handle_unknown='ignore' |
OrdinalEncoder |
Ordinal categorical feature encoding | Encodes categorical variables into ordered integers $0, \dots, K-1$ | categories='auto', encoded_missing_value |
LabelEncoder |
Target variable $y$ label encoding | Encodes string labels to integers $0 \dots C-1$ | For target $y$ only, NEVER for feature $X$ |
PolynomialFeatures |
High-order interaction terms | Generates polynomial features $[1, a, b, a^2, ab, b^2]$ | degree=2, include_bias=False |
3.1 Preprocessing Code Recipes
import numpy as np
from sklearn.preprocessing import (
StandardScaler, MinMaxScaler, RobustScaler, Normalizer,
Binarizer, OneHotEncoder, OrdinalEncoder, LabelEncoder, PolynomialFeatures
)
from sklearn.impute import SimpleImputer
# 1. Numerical standardization (never fit on test sets!)
scaler = StandardScaler()
X_train_std = scaler.fit_transform(X_train)
X_test_std = scaler.transform(X_test)
# 2. Missing value imputation
imputer = SimpleImputer(missing_values=np.nan, strategy='median')
X_train_imp = imputer.fit_transform(X_train)
X_test_imp = imputer.transform(X_test)
# 3. One-hot encoding categorical columns (handles unseen test categories gracefully)
ohe = OneHotEncoder(handle_unknown='ignore', sparse_output=False)
X_train_ohe = ohe.fit_transform(X_train_cat)
X_test_ohe = ohe.transform(X_test_cat)
# 4. Target variable encoding
le = LabelEncoder()
y_train_enc = le.fit_transform(y_train)
y_test_enc = le.transform(y_test)
# 5. Polynomial feature expansion
poly = PolynomialFeatures(degree=2, include_bias=False)
X_train_poly = poly.fit_transform(X_train)
4. Common Estimators Master Index
4.1 Supervised Learning: Classification
from sklearn.linear_model import LogisticRegression
from sklearn.neighbors import KNeighborsClassifier
from sklearn.svm import SVC
from sklearn.naive_bayes import GaussianNB
from sklearn.ensemble import RandomForestClassifier, HistGradientBoostingClassifier
# Logistic Regression baseline
clf_lr = LogisticRegression(C=1.0, max_iter=200, random_state=42)
# K-Nearest Neighbors
clf_knn = KNeighborsClassifier(n_neighbors=5, metric='minkowski', p=2)
# Support Vector Machine (RBF kernel)
clf_svm = SVC(C=1.0, kernel='rbf', gamma='scale', probability=True, random_state=42)
# Gaussian Naive Bayes
clf_gnb = GaussianNB()
# Random Forest (Bagging ensemble)
clf_rf = RandomForestClassifier(n_estimators=100, max_depth=None, n_jobs=-1, random_state=42)
# Histogram-based Gradient Boosting (fast GBDT matching LightGBM)
clf_hgb = HistGradientBoostingClassifier(learning_rate=0.1, max_iter=100, random_state=42)
4.2 Supervised Learning: Regression
from sklearn.linear_model import LinearRegression, Ridge, Lasso
from sklearn.svm import SVR
from sklearn.ensemble import RandomForestRegressor, HistGradientBoostingRegressor
# Ordinary Least Squares Linear Regression
reg_lr = LinearRegression()
# Ridge Regression (L2 regularization to combat multicollinearity)
reg_ridge = Ridge(alpha=1.0)
# Lasso Regression (L1 regularization for sparse feature selection)
reg_lasso = Lasso(alpha=0.1)
# Support Vector Regression
reg_svr = SVR(kernel='rbf', C=1.0, epsilon=0.1)
# Random Forest Regressor
reg_rf = RandomForestRegressor(n_estimators=100, n_jobs=-1, random_state=42)
4.3 Unsupervised Learning: Clustering & Decomposition
from sklearn.cluster import KMeans, DBSCAN
from sklearn.decomposition import PCA
# K-Means Clustering
kmeans = KMeans(n_clusters=3, init='k-means++', n_init=10, random_state=42)
kmeans.fit(X)
cluster_labels = kmeans.labels_ # Cluster assignments
cluster_centers = kmeans.cluster_centers_
# Density-Based Spatial Clustering (noise-resistant, arbitrary shapes)
dbscan = DBSCAN(eps=0.5, min_samples=5)
db_labels = dbscan.fit_predict(X) # -1 indicates noise outliers
# Principal Component Analysis (PCA)
pca = PCA(n_components=0.95) # Retain 95% cumulative explained variance
X_pca = pca.fit_transform(X)
print("Retained components:", pca.n_components_)
print("Explained variance ratio:", pca.explained_variance_ratio_)
5. Unified API Operations Cheat Sheet
| Operation | Canonical Method | Target Object Type | Semantics & Return Value |
|---|---|---|---|
| Fit | estimator.fit(X, y) |
All Estimators | Estimates parameters from training data; returns self |
| Predict | predictor.predict(X) |
All Predictors | Generates hard class labels or continuous targets |
| Probability | predictor.predict_proba(X) |
Classifiers | Outputs posterior class probability distribution matrix |
| Margin Score | predictor.decision_function(X) |
Linear / SVM | Distance of samples to the separating hyperplane |
| Transform | transformer.transform(X) |
All Transformers | Projects input features using learned parameters |
| Fit & Transform | transformer.fit_transform(X) |
All Transformers | Computes parameters and transforms in one optimized step |
| Default Score | estimator.score(X, y) |
Classifiers / Regressors | Accuracy for classification; $R^2$ coefficient of determination for regression |
6. Evaluation Metrics (sklearn.metrics)
6.1 Classification Metrics
from sklearn.metrics import (
accuracy_score, precision_score, recall_score, f1_score,
roc_auc_score, confusion_matrix, classification_report,
ConfusionMatrixDisplay
)
# 1. Hard label classification metrics (specify average for multi-class)
acc = accuracy_score(y_test, y_pred)
prec = precision_score(y_test, y_pred, average='macro')
rec = recall_score(y_test, y_pred, average='macro')
f1 = f1_score(y_test, y_pred, average='macro')
# 2. Continuous ranking metrics (ROC-AUC requires predicted probabilities)
y_prob = clf.predict_proba(X_test)[:, 1] # Binary positive class probabilities
auc = roc_auc_score(y_test, y_prob)
# 3. Confusion matrix & comprehensive text report
cm = confusion_matrix(y_test, y_pred)
print(classification_report(y_test, y_pred))
# 4. Display confusion matrix visualization
disp = ConfusionMatrixDisplay(confusion_matrix=cm)
disp.plot()
6.2 Regression Metrics
from sklearn.metrics import mean_absolute_error, mean_squared_error, r2_score
mae = mean_absolute_error(y_test, y_pred)
mse = mean_squared_error(y_test, y_pred)
rmse = mean_squared_error(y_test, y_pred, squared=False) # Root Mean Squared Error
r2 = r2_score(y_test, y_pred) # R² goodness-of-fit
6.3 Clustering Metrics
from sklearn.metrics import (
adjusted_rand_score, homogeneity_score, v_measure_score, silhouette_score
)
# External validation (when ground truth labels exist)
ari = adjusted_rand_score(y_true, cluster_labels) # Range: [-1, 1], higher is better
homo = homogeneity_score(y_true, cluster_labels) # Range: [0, 1]
vm = v_measure_score(y_true, cluster_labels) # Harmonic mean of homogeneity & completeness
# Internal validation (geometry-based, no ground truth needed)
sil = silhouette_score(X, cluster_labels) # Range: [-1, 1], near 1 indicates clean clusters
7. Cross-Validation & Hyperparameter Tuning (sklearn.model_selection)
7.1 Cross-Validation
from sklearn.model_selection import cross_val_score, cross_validate, StratifiedKFold
# Standard K-fold cross-validation
scores = cross_val_score(clf, X, y, cv=5, scoring='accuracy')
print(f"5-Fold CV Accuracy: {scores.mean():.4f} (±{scores.std():.4f})")
# Stratified K-Fold generator
skf = StratifiedKFold(n_splits=5, shuffle=True, random_state=42)
scores_skf = cross_val_score(clf, X, y, cv=skf, scoring='f1_macro')
# Multi-metric joint evaluation
multi_scores = cross_validate(
clf, X, y, cv=5,
scoring=['accuracy', 'roc_auc'],
return_train_score=False
)
print("Mean Validation AUC:", multi_scores['test_roc_auc'].mean())
7.2 Grid Search & Randomized Search
from sklearn.model_selection import GridSearchCV, RandomizedSearchCV
from scipy.stats import randint
# 1. Exhaustive Grid Search
param_grid = {
'n_neighbors': [3, 5, 7, 11],
'weights': ['uniform', 'distance'],
'metric': ['euclidean', 'manhattan']
}
grid = GridSearchCV(
estimator=KNeighborsClassifier(),
param_grid=param_grid,
cv=5,
scoring='accuracy',
n_jobs=-1
)
grid.fit(X_train, y_train)
print("Best Params:", grid.best_params_)
best_model = grid.best_estimator_
# 2. Randomized Parameter Search (for large parameter search spaces)
param_dist = {
'n_estimators': randint(50, 300),
'max_depth': [None, 5, 10, 20],
'min_samples_split': randint(2, 11)
}
rand_search = RandomizedSearchCV(
estimator=RandomForestClassifier(random_state=42),
param_distributions=param_dist,
n_iter=20, # Number of random parameter candidates
cv=5,
scoring='accuracy',
random_state=42,
n_jobs=-1
)
rand_search.fit(X_train, y_train)
print("Best Random Search Params:", rand_search.best_params_)
8. Industrial Best Practice: Pipeline & ColumnTransformer
8.1 Why Pipelines are Mandatory
Executing feature scaling or imputation globally across an entire dataset prior to splitting or cross-validation causes Data Leakage: information from test folds bleeds into training statistics, resulting in deceptively optimistic offline metrics that fail in production.
Pipeline binds preprocessing transformations and estimators into an atomic, leakage-free unit. In cross-validation, transformers are fitted strictly on each training fold and applied to validation folds automatically.
8.2 Production ColumnTransformer Recipe
Real-world tabular data mixes continuous numerical columns and categorical variables:
from sklearn.compose import ColumnTransformer
from sklearn.pipeline import Pipeline
from sklearn.impute import SimpleImputer
from sklearn.preprocessing import StandardScaler, OneHotEncoder
from sklearn.ensemble import RandomForestClassifier
# 1. Explicitly partition feature column names
numeric_features = ['age', 'income', 'credit_score']
categorical_features = ['gender', 'occupation', 'city']
# 2. Numerical sub-pipeline: median impute -> standard scale
numeric_transformer = Pipeline(steps=[
('imputer', SimpleImputer(strategy='median')),
('scaler', StandardScaler())
])
# 3. Categorical sub-pipeline: constant impute -> one-hot encode
categorical_transformer = Pipeline(steps=[
('imputer', SimpleImputer(strategy='constant', fill_value='missing')),
('onehot', OneHotEncoder(handle_unknown='ignore', sparse_output=False))
])
# 4. Bind columns into a unified ColumnTransformer
preprocessor = ColumnTransformer(transformers=[
('num', numeric_transformer, numeric_features),
('cat', categorical_transformer, categorical_features)
])
# 5. Assemble end-to-end full model pipeline
full_pipeline = Pipeline(steps=[
('preprocessor', preprocessor),
('classifier', RandomForestClassifier(n_estimators=100, random_state=42))
])
# 6. Atomic fit and inference (zero leakage, elegant deployment)
full_pipeline.fit(X_train_df, y_train)
y_pred = full_pipeline.predict(X_test_df)
print("Pipeline Test Accuracy:", full_pipeline.score(X_test_df, y_test))