Chapitre 4 — Machine Learning avec scikit-learn 🤖

🤖 scikit-learn : ML accessible

scikit-learn fournit une API unifiée pour 50+ algorithmes. Apprenez le pipeline classique et entraînez votre premier modèle en 10 lignes.

🔄 Pipeline ML classique

1. Data : charger le dataset (pandas)
2. Preprocessing : nettoyer, encoder, scaler
3. Split : train/test (70/15/15)
4. Model : entraîner
5. Evaluate : métriques + courbes
6. Tune : optimiser hyperparamètres
7. Save : exporter le modèle

✂️ Train/test split

from sklearn.model_selection import train_test_split

X = df.drop(columns=['prix'])
y = df['prix']

X_train, X_test, y_train, y_test = train_test_split(
    X, y, test_size=0.30, random_state=42
)
X_val, X_test, y_val, y_test = train_test_split(
    X_test, y_test, test_size=0.50, random_state=42
)
# Final : 70% train, 15% val, 15% test

🔢 Encodage variables catégorielles

from sklearn.preprocessing import LabelEncoder, OneHotEncoder

# LabelEncoder : pour cibles ordinales (low/medium/high)
le = LabelEncoder()
df['niveau_enc'] = le.fit_transform(df['niveau'])

# OneHotEncoder : pour features non ordinales (région, produit)
df_encoded = pd.get_dummies(df, columns=['region', 'produit'], drop_first=True)

⚖️ Scaling numérique

from sklearn.preprocessing import StandardScaler, MinMaxScaler

scaler = StandardScaler()  # moyenne 0, écart-type 1
X_train_scaled = scaler.fit_transform(X_train)
X_test_scaled = scaler.transform(X_test)  # PAS de fit_transform sur test !

📈 Régression : prédire un nombre

from sklearn.linear_model import LinearRegression
from sklearn.ensemble import RandomForestRegressor
from xgboost import XGBRegressor

# Régression linéaire (baseline)
lr = LinearRegression()
lr.fit(X_train, y_train)
pred = lr.predict(X_test)

# Random Forest
rf = RandomForestRegressor(n_estimators=100, random_state=42)
rf.fit(X_train, y_train)

# XGBoost (souvent gagnant en compétition)
xgb = XGBRegressor(n_estimators=200, learning_rate=0.05)
xgb.fit(X_train, y_train)

🎯 Classification : prédire une classe

from sklearn.linear_model import LogisticRegression
from sklearn.neighbors import KNeighborsClassifier
from sklearn.tree import DecisionTreeClassifier

logreg = LogisticRegression(max_iter=1000)
knn = KNeighborsClassifier(n_neighbors=5)
tree = DecisionTreeClassifier(max_depth=5)

📐 Métriques régression

from sklearn.metrics import mean_squared_error, r2_score
import numpy as np

rmse = np.sqrt(mean_squared_error(y_test, pred))
r2 = r2_score(y_test, pred)
print(f'RMSE : {rmse:.0f} | R² : {r2:.3f}')

🏷️ Métriques classification

from sklearn.metrics import (accuracy_score, f1_score,
                              roc_auc_score, confusion_matrix,
                              classification_report)

print(classification_report(y_test, pred))
print(confusion_matrix(y_test, pred))
print(f'AUC-ROC : {roc_auc_score(y_test, proba):.3f}')

🔁 Cross-validation

from sklearn.model_selection import cross_val_score

scores = cross_val_score(rf, X, y, cv=5, scoring='r2')
print(f'R² moyen : {scores.mean():.3f} ± {scores.std():.3f}')

🎛️ Hyperparameter tuning

from sklearn.model_selection import GridSearchCV

params = {
    'n_estimators': [100, 200, 500],
    'max_depth': [5, 10, 20, None]
}
grid = GridSearchCV(RandomForestRegressor(), params, cv=5, n_jobs=-1)
grid.fit(X_train, y_train)
print(grid.best_params_)

# Optuna (plus moderne, recherche bayésienne)
import optuna
def objective(trial):
    n = trial.suggest_int('n_estimators', 100, 1000)
    d = trial.suggest_int('max_depth', 3, 30)
    model = RandomForestRegressor(n_estimators=n, max_depth=d)
    return cross_val_score(model, X, y, cv=3).mean()

study = optuna.create_study(direction='maximize')
study.optimize(objective, n_trials=50)

🧠 Feature importance

importances = pd.Series(rf.feature_importances_, index=X.columns)
importances.sort_values().plot(kind='barh', figsize=(8,6))
plt.title('Feature importance')

💾 Sauvegarder le modèle

import joblib

joblib.dump(rf, 'modele_loyer.joblib')
# Plus tard, en prod
model = joblib.load('modele_loyer.joblib')
prediction = model.predict([[80, 3, 5, 'Bastos']])