Machine Learning basico: el primer modelo de verdad

import numpy as np
import pandas as pd
from sklearn.datasets import load_breast_cancer
from sklearn.model_selection import train_test_split, cross_val_score, StratifiedKFold
from sklearn.preprocessing import StandardScaler
from sklearn.pipeline import Pipeline
from sklearn.linear_model import LogisticRegression
from sklearn.neighbors import KNeighborsClassifier
from sklearn.tree import DecisionTreeClassifier
from sklearn.ensemble import RandomForestClassifier, GradientBoostingClassifier
from sklearn.metrics import classification_report, roc_auc_score, f1_score

# Cargar datos
data = load_breast_cancer()
X = pd.DataFrame(data.data, columns=data.feature_names)
y = data.target  # 0=maligno, 1=benigno
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42, stratify=y)

# Modelos con la misma estructura de Pipeline
# - StandardScaler solo donde tiene sentido (LR y KNN)
# - Random_state fijo para reproducibilidad
modelos = {
    'Logistic Regression': Pipeline([
        ('scaler', StandardScaler()),
        ('model', LogisticRegression(max_iter=1000, random_state=42))
    ]),
    'KNN (k=5)': Pipeline([
        ('scaler', StandardScaler()),
        ('model', KNeighborsClassifier(n_neighbors=5))
    ]),
    'Decision Tree': Pipeline([
        ('model', DecisionTreeClassifier(max_depth=5, random_state=42))
    ]),
    'Random Forest': Pipeline([
        ('model', RandomForestClassifier(n_estimators=100, max_depth=10, random_state=42, n_jobs=-1))
    ]),
    'Gradient Boosting': Pipeline([
        ('model', GradientBoostingClassifier(n_estimators=100, max_depth=3, random_state=42))
    ]),
}

# Validacion cruzada ESTRATIFICADA (mantiene proporcion de clases)
cv = StratifiedKFold(n_splits=5, shuffle=True, random_state=42)
print(f"{'Modelo':<22} {'Accuracy':<12} {'F1':<10} {'ROC-AUC':<10}")
print("=" * 60)

for nombre, modelo in modelos.items():
    acc = cross_val_score(modelo, X_train, y_train, cv=cv, scoring='accuracy').mean()
    f1  = cross_val_score(modelo, X_train, y_train, cv=cv, scoring='f1').mean()
    auc = cross_val_score(modelo, X_train, y_train, cv=cv, scoring='roc_auc').mean()
    print(f"{nombre:<22} {acc:.3f}        {f1:.3f}     {auc:.3f}")

# NOTA: se entrena 5 modelos * 3 metricas * 5 folds = 75 fits por ejecucion
# En datasets grandes, considera usar n_jobs=-1

from sklearn.model_selection import cross_val_score, KFold
from sklearn.pipeline import Pipeline
from sklearn.impute import SimpleImputer
from sklearn.preprocessing import StandardScaler
from sklearn.linear_model import LogisticRegression

# FORMA INCORRECTA (leakage)
# imp = SimpleImputer()
# X_imp = imp.fit_transform(X)  # usa la media de TODO el dataset
# cross_val_score(model, X_imp, y)  # cada fold usa stats de test

# FORMA CORRECTA (sin leakage)
model = Pipeline([
    ('imputer', SimpleImputer()),  # aprende la media SOLO en train de cada fold
    ('scaler',   StandardScaler()),
    ('clf',      LogisticRegression())
])

cv = KFold(n_splits=5, shuffle=True, random_state=42)
scores = cross_val_score(model, X, y, cv=cv, scoring='roc_auc')
print(f"AUC: {scores.mean():.3f} ± {scores.std():.3f}")

import matplotlib.pyplot as plt
import numpy as np
from sklearn.model_selection import learning_curve
from sklearn.ensemble import RandomForestClassifier

train_sizes, train_scores, val_scores = learning_curve(
    RandomForestClassifier(n_estimators=100, max_depth=10, random_state=42),
    X_train, y_train,
    cv=5,
    train_sizes=np.linspace(0.1, 1.0, 10),
    scoring='roc_auc',
    n_jobs=-1,
)

train_mean = train_scores.mean(axis=1)
val_mean   = val_scores.mean(axis=1)

plt.figure(figsize=(8, 5))
plt.plot(train_sizes, train_mean, 'o-', label='Train AUC', color='#06B6D4')
plt.plot(train_sizes, val_mean, 'o-', label='Validation AUC', color='#8B5CF6')
plt.fill_between(train_sizes, val_mean - val_scores.std(axis=1), val_mean + val_scores.std(axis=1), alpha=0.2, color='#8B5CF6')
plt.xlabel('Tamanio del training set')
plt.ylabel('ROC-AUC')
plt.title('Learning Curves: diagnostico de bias-variance')
plt.legend()
plt.grid(alpha=0.3)
import seaborn as sns; sns.despine()
plt.tight_layout()
plt.show()

# Diagnostico:
# - Train alto, val bajo, curvas NO convergen → overfitting. Mas datos o regularizacion.
# - Train bajo, val bajo, curvas convergen → underfitting. Modelo mas complejo.
# - Train alto, val alto, curvas convergen → punto optimo.