Machine Learning with scikit-learn — Complete Pipeline

The scikit-learn Design Philosophy

scikit-learn is built around one elegant interface: the Estimator. Every object — preprocessor, model, or pipeline — implements the same methods:

• fit(X, y) — learn from data (returns self)
• transform(X) — apply a learned transformation (transformers only)
• predict(X) — generate predictions (predictors only)
• fit_transform(X, y) — fit then transform in one step (more efficient)
• score(X, y) — evaluate performance

This uniformity means you can swap any transformer or model without changing surrounding code. A Pipeline chains them so fit(X, y) trains every step and predict(X) runs them all.

Datasets — Built-in and Synthetic

Data Preprocessing — Every Transformer

Scalers

Encoders for Categorical Variables

Train-Test Split and Cross-Validation

Key Algorithms — Deep Dive with Examples

Pipelines — The Right Way to Build ML Systems

Model Evaluation — Complete Metrics

Hyperparameter Tuning

Feature Importance and Selection

Project: Complete Customer Churn Predictor

Exercises

Exercise 1 — Preprocessing Pipeline

Exercise 2 — Cross-Validation Comparison

Exercise 3 — Build a Full Pipeline

Exercise 4 — Clustering

Key Takeaways

• scikit-learn's Estimator interface (fit, transform, predict) applies to every object — swap components without changing surrounding code
• Always split data before fitting any preprocessor — fitting the scaler on all data leaks test statistics into training
• Pipeline prevents data leakage and makes model serialization, cross-validation, and grid search clean and correct
• StratifiedKFold preserves class proportions — always use it for classification, especially with imbalanced classes
• cross_val_score gives a more reliable performance estimate than a single train/test split
• For imbalanced classes: accuracy is misleading — use ROC-AUC, F1, or precision/recall; also consider class_weight="balanced"
• RandomizedSearchCV is almost always better than GridSearchCV for large hyperparameter spaces — same quality, fraction of the time
• Feature importance from trees is fast but biased toward high-cardinality features — complement with permutation importance or SHAP
• ColumnTransformer is the right way to apply different preprocessing to different column types — it composes cleanly inside a Pipeline
• The best model depends on your data: linear models for interpretability, trees for non-linearity without scaling, SVM for high-dimensional data, ensemble methods for best raw performance