Questions d'entretien pour Data Scientist Senior : Guide complet

import numpy as np
from sklearn.model_selection import learning_curve
from sklearn.tree import DecisionTreeRegressor
import matplotlib.pyplot as plt

# Générer des données
X = np.random.rand(100, 1) * 10
y = 2 * X + 3 + np.random.randn(100, 1) * 2

# Modèle à biais élevé (max_depth=1)
high_bias = DecisionTreeRegressor(max_depth=1)

# Modèle à variance élevée (max_depth=20)
high_variance = DecisionTreeRegressor(max_depth=20)

# Modèle optimal (max_depth=3)
optimal = DecisionTreeRegressor(max_depth=3)

# Les courbes d'apprentissage montrent le compromis biais-variance
train_sizes, train_scores, val_scores = learning_curve(
    optimal, X, y.ravel(), cv=5, train_sizes=np.linspace(0.1, 1.0, 10)
)

print(f"Score d'entraînement : {train_scores.mean():.2f}")
print(f"Score de validation : {val_scores.mean():.2f}")

from sklearn.linear_model import Lasso, Ridge, ElasticNet
from sklearn.datasets import make_regression
from sklearn.model_selection import train_test_split
import numpy as np

# Générer des données avec de nombreuses caractéristiques
X, y = make_regression(n_samples=100, n_features=20, n_informative=10, random_state=42)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.3)

# Régularisation L1 (Lasso)
lasso = Lasso(alpha=1.0)
lasso.fit(X_train, y_train)
print(f"Coefficients Lasso : {np.sum(lasso.coef_ != 0)} non nuls sur {len(lasso.coef_)}")

# Régularisation L2 (Ridge)
ridge = Ridge(alpha=1.0)
ridge.fit(X_train, y_train)
print(f"Coefficients Ridge : {np.sum(ridge.coef_ != 0)} non nuls sur {len(ridge.coef_)}")

# Elastic Net (L1 + L2)
elastic = ElasticNet(alpha=1.0, l1_ratio=0.5)
elastic.fit(X_train, y_train)

print(f"\nScore Lasso : {lasso.score(X_test, y_test):.3f}")
print(f"Score Ridge : {ridge.score(X_test, y_test):.3f}")
print(f"Score Elastic Net : {elastic.score(X_test, y_test):.3f}")

from sklearn.ensemble import RandomForestClassifier, GradientBoostingClassifier
from sklearn.datasets import load_breast_cancer
from sklearn.model_selection import train_test_split, cross_val_score

# Charger les données
data = load_breast_cancer()
X_train, X_test, y_train, y_test = train_test_split(
    data.data, data.target, test_size=0.3, random_state=42
)

# Bagging - Random Forest
rf = RandomForestClassifier(n_estimators=100, random_state=42)
rf.fit(X_train, y_train)
rf_score = rf.score(X_test, y_test)

# Boosting - Gradient Boosting
gb = GradientBoostingClassifier(n_estimators=100, random_state=42)
gb.fit(X_train, y_train)
gb_score = gb.score(X_test, y_test)

print(f"Précision de Random Forest (Bagging) : {rf_score:.3f}")
print(f"Précision de Gradient Boosting : {gb_score:.3f}")

# Validation croisée
rf_cv = cross_val_score(rf, data.data, data.target, cv=5)
gb_cv = cross_val_score(gb, data.data, data.target, cv=5)

print(f"\nScores CV RF : {rf_cv.mean():.3f} (+/- {rf_cv.std():.3f})")
print(f"Scores CV GB : {gb_cv.mean():.3f} (+/- {gb_cv.std():.3f})")

from sklearn.model_selection import cross_val_score, KFold, StratifiedKFold
from sklearn.linear_model import LogisticRegression
from sklearn.datasets import load_iris

data = load_iris()
X, y = data.data, data.target

model = LogisticRegression(max_iter=200)

# K-Fold standard
kfold = KFold(n_splits=5, shuffle=True, random_state=42)
scores = cross_val_score(model, X, y, cv=kfold)
print(f"Scores CV K-Fold : {scores}")
print(f"Moyenne : {scores.mean():.3f} (+/- {scores.std():.3f})")

# K-Fold Stratifié (préserve la distribution des classes)
stratified_kfold = StratifiedKFold(n_splits=5, shuffle=True, random_state=42)
stratified_scores = cross_val_score(model, X, y, cv=stratified_kfold)
print(f"\nScores CV K-Fold Stratifié : {stratified_scores}")
print(f"Moyenne : {stratified_scores.mean():.3f} (+/- {stratified_scores.std():.3f})")

# Validation croisée personnalisée
from sklearn.model_selection import cross_validate

cv_results = cross_validate(
    model, X, y, cv=5,
    scoring=['accuracy', 'precision_macro', 'recall_macro'],
    return_train_score=True
)

print(f"\nPrécision du test : {cv_results['test_accuracy'].mean():.3f}")
print(f"Précision du test : {cv_results['test_precision_macro'].mean():.3f}")
print(f"Rappel du test : {cv_results['test_recall_macro'].mean():.3f}")

from sklearn.decomposition import PCA
from sklearn.manifold import TSNE
from sklearn.datasets import load_digits
import matplotlib.pyplot as plt

# Charger des données de haute dimension
digits = load_digits()
X, y = digits.data, digits.target

print(f"Forme originale : {X.shape}")

# PCA - réduire à 2 dimensions
pca = PCA(n_components=2)
X_pca = pca.fit_transform(X)
print(f"Forme PCA : {X_pca.shape}")
print(f"Ratio de variance expliquée : {pca.explained_variance_ratio_}")
print(f"Variance totale expliquée : {pca.explained_variance_ratio_.sum():.3f}")

# t-SNE - réduire à 2 dimensions
tsne = TSNE(n_components=2, random_state=42)
X_tsne = tsne.fit_transform(X)
print(f"Forme t-SNE : {X_tsne.shape}")

# PCA pour l'extraction de caractéristiques (conserver 95% de la variance)
pca_95 = PCA(n_components=0.95)
X_reduced = pca_95.fit_transform(X)
print(f"\nComposantes pour 95% de la variance : {pca_95.n_components_}")
print(f"Forme réduite : {X_reduced.shape}")

from sklearn.metrics import roc_curve, roc_auc_score, auc
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import train_test_split
from sklearn.datasets import load_breast_cancer
import matplotlib.pyplot as plt

# Charger les données
data = load_breast_cancer()
X_train, X_test, y_train, y_test = train_test_split(
    data.data, data.target, test_size=0.3, random_state=42
)

# Entraîner le modèle
model = LogisticRegression(max_iter=10000)
model.fit(X_train, y_train)

# Obtenir les prédictions de probabilité
y_pred_proba = model.predict_proba(X_test)[:, 1]

# Calculer la courbe ROC
fpr, tpr, thresholds = roc_curve(y_test, y_pred_proba)
roc_auc = auc(fpr, tpr)

print(f"AUC : {roc_auc:.3f}")

# Alternative: calcul direct de l'AUC
auc_score = roc_auc_score(y_test, y_pred_proba)
print(f"AUC (direct) : {auc_score:.3f}")

# Trouver le seuil optimal (statistique J de Youden)
optimal_idx = np.argmax(tpr - fpr)
optimal_threshold = thresholds[optimal_idx]
print(f"Seuil optimal : {optimal_threshold:.3f}")

from sklearn.preprocessing import StandardScaler, OneHotEncoder, PolynomialFeatures
import pandas as pd
import numpy as np

# Exemple de données
df = pd.DataFrame({
    'age': [25, 30, 35, 40, 45],
    'salary': [50000, 60000, 75000, 80000, 90000],
    'department': ['IT', 'HR', 'IT', 'Finance', 'HR'],
    'date': pd.date_range('2023-01-01', periods=5)
})

# Encodage one-hot
df_encoded = pd.get_dummies(df, columns=['department'], prefix='dept')

# Mise à l'échelle
scaler = StandardScaler()
df_encoded[['age_scaled', 'salary_scaled']] = scaler.fit_transform(
    df_encoded[['age', 'salary']]
)

# Binning
df_encoded['age_group'] = pd.cut(df['age'], bins=[0, 30, 40, 100], labels=['young', 'mid', 'senior'])

# Caractéristiques de date
df_encoded['year'] = df['date'].dt.year
df_encoded['month'] = df['date'].dt.month
df_encoded['day_of_week'] = df['date'].dt.dayofweek

# Caractéristiques polynomiales
poly = PolynomialFeatures(degree=2, include_bias=False)
poly_features = poly.fit_transform(df[['age', 'salary']])

# Caractéristiques d'interaction
df_encoded['age_salary_interaction'] = df['age'] * df['salary']

print(df_encoded.head())

from sklearn.datasets import make_classification
from sklearn.model_selection import train_test_split
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import classification_report, confusion_matrix
from imblearn.over_sampling import SMOTE
from imblearn.under_sampling import RandomUnderSampler

# Créer un ensemble de données déséquilibré
X, y = make_classification(
    n_samples=1000, n_features=20, n_informative=15,
    n_classes=2, weights=[0.9, 0.1], random_state=42
)

print(f"Distribution des classes : {np.bincount(y)}")

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.3, random_state=42)

# 1. Sans gérer le déséquilibre
model_baseline = LogisticRegression()
model_baseline.fit(X_train, y_train)
y_pred_baseline = model_baseline.predict(X_test)
print("\nBaseline (pas de gestion) :")
print(classification_report(y_test, y_pred_baseline))

# 2. SMOTE (Synthetic Minority Over-sampling)
smote = SMOTE(random_state=42)
X_train_smote, y_train_smote = smote.fit_resample(X_train, y_train)
print(f"\nAprès SMOTE : {np.bincount(y_train_smote)}")

model_smote = LogisticRegression()
model_smote.fit(X_train_smote, y_train_smote)
y_pred_smote = model_smote.predict(X_test)
print("\nAvec SMOTE :")
print(classification_report(y_test, y_pred_smote))

# 3. Poids des classes
model_weighted = LogisticRegression(class_weight='balanced')
model_weighted.fit(X_train, y_train)
y_pred_weighted = model_weighted.predict(X_test)
print("\nAvec les poids des classes :")
print(classification_report(y_test, y_pred_weighted))

from sklearn.feature_selection import SelectKBest, chi2, RFE, SelectFromModel
from sklearn.ensemble import RandomForestClassifier
from sklearn.datasets import load_breast_cancer
from sklearn.preprocessing import MinMaxScaler

# Charger les données
data = load_breast_cancer()
X, y = data.data, data.target

# 1. Méthode de filtre - SelectKBest avec chi-carré
X_scaled = MinMaxScaler().fit_transform(X)
selector_chi2 = SelectKBest(chi2, k=10)
X_chi2 = selector_chi2.fit_transform(X_scaled, y)
print(f"Caractéristiques originales : {X.shape[1]}")
print(f"Caractéristiques sélectionnées (chi2) : {X_chi2.shape[1]}")

# 2. Méthode d'encapsulation - Recursive Feature Elimination
rf = RandomForestClassifier(n_estimators=100, random_state=42)
rfe = RFE(estimator=rf, n_features_to_select=10)
X_rfe = rfe.fit_transform(X, y)
print(f"Caractéristiques sélectionnées (RFE) : {X_rfe.shape[1]}")
print(f"Classement des caractéristiques : {rfe.ranking_}")

# 3. Méthode intégrée - Importance des caractéristiques basée sur les arbres
rf.fit(X, y)
importances = rf.feature_importances_
indices = np.argsort(importances)[::-1]

print("\nTop 10 des caractéristiques par importance :")
for i in range(10):
    print(f"{i+1}. {data.feature_names[indices[i]]}: {importances[indices[i]]:.4f}")

# SelectFromModel
selector_model = SelectFromModel(rf, threshold='median', prefit=True)
X_selected = selector_model.transform(X)
print(f"\nCaractéristiques sélectionnées (importance) : {X_selected.shape[1]}")

import pandas as pd
import numpy as np
from sklearn.model_selection import train_test_split

# Exemple de données à cardinalité élevée
df = pd.DataFrame({
    'city': np.random.choice([f'City_{i}' for i in range(100)], 1000),
    'target': np.random.randint(0, 2, 1000)
})

print(f"Villes uniques : {df['city'].nunique()}")

# 1. Target Encoding
target_means = df.groupby('city')['target'].mean()
df['city_target_encoded'] = df['city'].map(target_means)

# 2. Frequency Encoding
freq = df['city'].value_counts()
df['city_frequency'] = df['city'].map(freq)

# 3. Regroupement des catégories rares
freq_threshold = 10
rare_cities = freq[freq < freq_threshold].index
df['city_grouped'] = df['city'].apply(lambda x: 'Other' if x in rare_cities else x)

print(f"\nAprès regroupement : {df['city_grouped'].nunique()} valeurs uniques")

# 4. Hash encoding (en utilisant la bibliothèque category_encoders)
# from category_encoders import HashingEncoder
# encoder = HashingEncoder(cols=['city'], n_components=10)
# df_hashed = encoder.fit_transform(df)

print(df[['city', 'city_target_encoded', 'city_frequency', 'city_grouped']].head())

# 1. Entraîner et enregistrer le modèle
from sklearn.ensemble import RandomForestClassifier
from sklearn.datasets import load_iris
import joblib

# Entraîner le modèle
data = load_iris()
model = RandomForestClassifier()
model.fit(data.data, data.target)

# Enregistrer le modèle
joblib.dump(model, 'model.joblib')

# 2. Créer une API avec FastAPI
from fastapi import FastAPI
import numpy as np

app = FastAPI()

# Charger le modèle
model = joblib.load('model.joblib')

@app.post("/predict")
def predict(features: list):
    # Convertir en tableau numpy
    X = np.array(features).reshape(1, -1)
    prediction = model.predict(X)
    probability = model.predict_proba(X)
    
    return {
        "prediction": int(prediction[0]),
        "probability": probability[0].tolist()
    }

# 3. Dockerfile
"""
FROM python:3.9-slim
WORKDIR /app
COPY requirements.txt .
RUN pip install -r requirements.txt
COPY . .
CMD ["uvicorn", "main:app", "--host", "0.0.0.0", "--port", "8000"]
"""

# 4. Utilisation
# curl -X POST "http://localhost:8000/predict" \
#      -H "Content-Type: application/json" \
#      -d '{"features": [5.1, 3.5, 1.4, 0.2]}'

import numpy as np
from scipy import stats

# Simuler des données de production
training_data = np.random.normal(0, 1, 1000)
production_data = np.random.normal(0.5, 1.2, 1000)  # Dérive

# Détecter la dérive des données en utilisant le test de Kolmogorov-Smirnov
statistic, p_value = stats.ks_2samp(training_data, production_data)

print(f"Statistique KS : {statistic:.4f}")
print(f"P-value : {p_value:.4f}")

if p_value < 0.05:
    print("Dérive des données détectée ! Envisager de réentraîner le modèle.")
else:
    print("Aucune dérive significative détectée.")

# Surveiller les performances du modèle
class ModelMonitor:
    def __init__(self, model):
        self.model = model
        self.predictions = []
        self.actuals = []
        
    def log_prediction(self, X, y_pred, y_true=None):
        self.predictions.append(y_pred)
        if y_true is not None:
            self.actuals.append(y_true)
    
    def get_accuracy(self):
        if len(self.actuals) == 0:
            return None
        return np.mean(np.array(self.predictions) == np.array(self.actuals))
    
    def check_drift(self, new_data, reference_data):
        statistic, p_value = stats.ks_2samp(new_data, reference_data)
        return p_value < 0.05

# Utilisation
monitor = ModelMonitor(model)
# monitor.log_prediction(X, y_pred, y_true)
# accuracy = monitor.get_accuracy()

import numpy as np
from scipy import stats

# Simuler les résultats des tests A/B
# Groupe de contrôle (Modèle A)
control_conversions = 520
control_visitors = 10000

# Groupe de traitement (Modèle B)
treatment_conversions = 580
treatment_visitors = 10000

# Calculer les taux de conversion
control_rate = control_conversions / control_visitors
treatment_rate = treatment_conversions / treatment_visitors

print(f"Taux de conversion du contrôle : {control_rate:.4f}")
print(f"Taux de conversion du traitement : {treatment_rate:.4f}")
print(f"Augmentation : {((treatment_rate - control_rate) / control_rate * 100):.2f}%")

# Test de signification statistique (test z à deux proportions)
pooled_rate = (control_conversions + treatment_conversions) / (control_visitors + treatment_visitors)
se = np.sqrt(pooled_rate * (1 - pooled_rate) * (1/control_visitors + 1/treatment_visitors))
z_score = (treatment_rate - control_rate) / se
p_value = 2 * (1 - stats.norm.cdf(abs(z_score)))

print(f"\nScore Z : {z_score:.4f}")
print(f"P-value : {p_value:.4f}")

if p_value < 0.05:
    print("Le résultat est statistiquement significatif !")
    if treatment_rate > control_rate:
        print("Le traitement (Modèle B) est meilleur.")
    else:
        print("Le contrôle (Modèle A) est meilleur.")
else:
    print("Aucune différence statistiquement significative.")

# Calcul de la taille de l'échantillon
from statsmodels.stats.power import zt_ind_solve_power

required_sample = zt_ind_solve_power(
    effect_size=0.02,  # Effet minimal détectable
    alpha=0.05,
    power=0.8,
    alternative='two-sided'
)
print(f"\nTaille de l'échantillon requise par groupe : {int(required_sample)}")

# Exemple : MLflow pour le suivi des expériences
import mlflow
import mlflow.sklearn
from sklearn.ensemble import RandomForestClassifier
from sklearn.datasets import load_iris
from sklearn.model_selection import train_test_split
from sklearn.metrics import accuracy_score

# Charger les données
data = load_iris()
X_train, X_test, y_train, y_test = train_test_split(
    data.data, data.target, test_size=0.3, random_state=42
)

# Démarrer l'exécution MLflow
with mlflow.start_run():
    # Enregistrer les paramètres
    n_estimators = 100
    max_depth = 5
    mlflow.log_param("n_estimators", n_estimators)
    mlflow.log_param("max_depth", max_depth)
    
    # Entraîner le modèle
    model = RandomForestClassifier(
        n_estimators=n_estimators,
        max_depth=max_depth,
        random_state=42
    )
    model.fit(X_train, y