pip install openpyxl

Requirement already satisfied: openpyxl in c:\users\alvar\downloads\datasetsclase\.venv\lib\site-packages (3.1.5)
Requirement already satisfied: et-xmlfile in c:\users\alvar\downloads\datasetsclase\.venv\lib\site-packages (from openpyxl) (2.0.0)
Note: you may need to restart the kernel to use updated packages.

[notice] A new release of pip is available: 25.3 -> 26.0.1
[notice] To update, run: python.exe -m pip install --upgrade pip

import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
from pathlib import Path

pd.set_option('display.max_columns', 20)
plt.style.use('ggplot')
sns.set_palette('husl')

# Ruta al dataset (ajusta si es necesario)
DATA_PATH = r'C:\Users\alvar\Downloads\online_retail_II.xlsx'

# Si el Excel tiene varias hojas, podemos cargar la primera o especificar por nombre
#df = pd.read_excel(DATA_PATH, sheet_name=1)
xl = pd.ExcelFile(DATA_PATH)
df = pd.concat([pd.read_excel(DATA_PATH, sheet_name=s) for s in xl.sheet_names], ignore_index=True)
print(f'Filas: {len(df):,} | Columnas: {len(df.columns)}')

Filas: 1,067,371 | Columnas: 8

df.head(10)

df.info()

<class 'pandas.core.frame.DataFrame'>
RangeIndex: 1067371 entries, 0 to 1067370
Data columns (total 8 columns):
 #   Column       Non-Null Count    Dtype         
---  ------       --------------    -----         
 0   Invoice      1067371 non-null  object        
 1   StockCode    1067371 non-null  object        
 2   Description  1062989 non-null  object        
 3   Quantity     1067371 non-null  int64         
 4   InvoiceDate  1067371 non-null  datetime64[ns]
 5   Price        1067371 non-null  float64       
 6   Customer ID  824364 non-null   float64       
 7   Country      1067371 non-null  object        
dtypes: datetime64[ns](1), float64(2), int64(1), object(4)
memory usage: 65.1+ MB

df.dtypes

Invoice                object
StockCode              object
Description            object
Quantity                int64
InvoiceDate    datetime64[ns]
Price                 float64
Customer ID           float64
Country                object
dtype: object

# Nulos y duplicados
print("Nulos por columna:")
display(df.isnull().sum().sort_values(ascending=False))

print("\nDuplicados exactos:", df.duplicated().sum())

print("\nResumen rápido:")
print("Clientes únicos (raw):", df['Customer ID'].nunique())
print("Facturas únicas (raw):", df['Invoice'].nunique())
print("Periodo (raw):", df['InvoiceDate'].min(), "->", df['InvoiceDate'].max())

Nulos por columna:

Customer ID    243007
Description      4382
StockCode           0
Invoice             0
Quantity            0
InvoiceDate         0
Price               0
Country             0
dtype: int64

Duplicados exactos: 34335

Resumen rápido:
Clientes únicos (raw): 5942
Facturas únicas (raw): 53628
Periodo (raw): 2009-12-01 07:45:00 -> 2011-12-09 12:50:00

df.describe(include='all')

# Distribuciones básicas antes de modelar
fig, axes = plt.subplots(1, 2, figsize=(14, 5))

sns.boxplot(x=df['Quantity'], ax=axes[0])
axes[0].set_title('Boxplot Quantity')

sns.boxplot(x=df['Price'], ax=axes[1])
axes[1].set_title('Boxplot Price')

plt.tight_layout()
plt.show()

def clean_retail_data(df):
    """
    Limpieza reutilizable para Online Retail II.

    Qué hace:
    - estandariza nombres de columnas
    - elimina duplicados exactos
    - convierte tipos
    - marca cancelaciones
    - elimina filas sin Customer ID o sin fecha
    - deja solo compras válidas para el modelado principal
    - crea el importe por línea
    """

    df = df.copy()
    df.columns = df.columns.str.strip()

    # Duplicados exactos
    df = df.drop_duplicates()

    # Tipos
    df['InvoiceDate'] = pd.to_datetime(df['InvoiceDate'], errors='coerce')
    df['Customer ID'] = pd.to_numeric(df['Customer ID'], errors='coerce')

    # Marcar cancelaciones
    df['is_cancellation'] = df['Invoice'].astype(str).str.startswith('C').astype(int)

    # Filas críticas
    df = df.dropna(subset=['Customer ID', 'InvoiceDate'])

    # Tipo de cliente
    df['Customer ID'] = df['Customer ID'].astype(int)

    # Para el modelo principal nos quedamos con compras reales
    df = df[(df['Quantity'] > 0) & (df['Price'] > 0)]

    # Revenue por línea
    df['line_total'] = df['Quantity'] * df['Price']

    return df

df_clean = clean_retail_data(df)

print("Dimensión original:", df.shape)
print("Dimensión limpia:", df_clean.shape)
print("Clientes únicos (clean):", df_clean['Customer ID'].nunique())
print("Facturas únicas (clean):", df_clean['Invoice'].nunique())
print("Periodo (clean):", df_clean['InvoiceDate'].min(), "->", df_clean['InvoiceDate'].max())

df_clean.head()

Dimensión original: (1067371, 8)
Dimensión limpia: (779425, 10)
Clientes únicos (clean): 5878
Facturas únicas (clean): 36969
Periodo (clean): 2009-12-01 07:45:00 -> 2011-12-09 12:50:00

# Resumen de negocio tras la limpieza
print("Nº clientes (raw):", df['Customer ID'].nunique())
print("Periodo raw:", df['InvoiceDate'].min(), "-", df['InvoiceDate'].max())
print("Total transacciones (raw):", df['Invoice'].nunique())
print("-" * 50)
print("Nº clientes (clean):", df_clean['Customer ID'].nunique())
print("Periodo clean:", df_clean['InvoiceDate'].min(), "-", df_clean['InvoiceDate'].max())
print("Total transacciones (clean):", df_clean['Invoice'].nunique())

Nº clientes (raw): 5942
Periodo raw: 2009-12-01 07:45:00 - 2011-12-09 12:50:00
Total transacciones (raw): 53628
--------------------------------------------------
Nº clientes (clean): 5878
Periodo clean: 2009-12-01 07:45:00 - 2011-12-09 12:50:00
Total transacciones (clean): 36969

# Se reservan los últimos 3 meses del dataset para construir el target
max_date = df_clean['InvoiceDate'].max()
cutoff_date = max_date - pd.DateOffset(months=3)

df_train = df_clean[df_clean['InvoiceDate'] < cutoff_date].copy()
df_future = df_clean[df_clean['InvoiceDate'] >= cutoff_date].copy()

future_customers = df_future['Customer ID'].unique()

print("Fecha máxima del dataset:", max_date)
print("Cutoff (3 últimos meses):", cutoff_date)
print("Periodo train:", df_train['InvoiceDate'].min(), "->", df_train['InvoiceDate'].max())
print("Periodo target:", df_future['InvoiceDate'].min(), "->", df_future['InvoiceDate'].max())
print("Clientes en train:", df_train['Customer ID'].nunique())
print("Clientes que recompran en target:", len(future_customers))

Fecha máxima del dataset: 2011-12-09 12:50:00
Cutoff (3 últimos meses): 2011-09-09 12:50:00
Periodo train: 2009-12-01 07:45:00 -> 2011-09-09 12:41:00
Periodo target: 2011-09-09 13:01:00 -> 2011-12-09 12:50:00
Clientes en train: 5279
Clientes que recompran en target: 2893

# Variables RFM clásicas a nivel cliente
snapshot_date = df_train['InvoiceDate'].max() + pd.Timedelta(days=1)

rfm = df_train.groupby('Customer ID').agg({
    'InvoiceDate': lambda x: (snapshot_date - x.max()).days,
    'Invoice': 'nunique',
    'line_total': 'sum'
}).reset_index()

rfm.columns = ['Customer ID', 'Recency', 'Frequency', 'Monetary']
rfm.head()

# ============================================================
# Feature engineering a nivel cliente
# ============================================================

# Ticket medio
rfm['avg_ticket'] = rfm['Monetary'] / rfm['Frequency']

# Número de productos distintos comprados
products = df_train.groupby('Customer ID')['StockCode'].nunique().reset_index()
products.columns = ['Customer ID', 'num_products']
rfm = rfm.merge(products, on='Customer ID', how='left')

# Número de descripciones distintas
descriptions = df_train.groupby('Customer ID')['Description'].nunique().reset_index()
descriptions.columns = ['Customer ID', 'num_descriptions']
rfm = rfm.merge(descriptions, on='Customer ID', how='left')

# Número de países distintos
countries = df_train.groupby('Customer ID')['Country'].nunique().reset_index()
countries.columns = ['Customer ID', 'num_countries']
rfm = rfm.merge(countries, on='Customer ID', how='left')

# Cantidad total de artículos comprados
qty = df_train.groupby('Customer ID')['Quantity'].sum().reset_index()
qty.columns = ['Customer ID', 'total_quantity']
rfm = rfm.merge(qty, on='Customer ID', how='left')

# Dataset a nivel factura para crear variables de ticket
invoice_level = (
    df_train.groupby(['Customer ID', 'Invoice'], as_index=False)
    .agg(
        invoice_date=('InvoiceDate', 'min'),
        invoice_amount=('line_total', 'sum'),
        invoice_items=('Quantity', 'sum'),
        invoice_products=('StockCode', 'nunique')
    )
)

# Importe de la última factura
last_invoice = (
    invoice_level.sort_values(['Customer ID', 'invoice_date'])
    .groupby('Customer ID')
    .tail(1)[['Customer ID', 'invoice_amount']]
    .rename(columns={'invoice_amount': 'last_invoice_amount'})
)
rfm = rfm.merge(last_invoice, on='Customer ID', how='left')

# Importe medio por factura
mean_invoice_amount = invoice_level.groupby('Customer ID')['invoice_amount'].mean().reset_index()
mean_invoice_amount.columns = ['Customer ID', 'mean_invoice_amount']
rfm = rfm.merge(mean_invoice_amount, on='Customer ID', how='left')

# Relación entre la última factura y el ticket medio
rfm['ratio_last_invoice_mean'] = rfm['last_invoice_amount'] / rfm['mean_invoice_amount']

# Variables temporales simples
rfm['freq_per_recency'] = rfm['Frequency'] / (rfm['Recency'] + 1)
rfm['monetary_per_recency'] = rfm['Monetary'] / (rfm['Recency'] + 1)
rfm['is_recent_customer'] = (rfm['Recency'] <= 30).astype(int)

# Target: 1 si el cliente reaparece en los últimos 3 meses
rfm['target'] = rfm['Customer ID'].isin(future_customers).astype(int)

# Limpieza final
rfm = rfm.replace([np.inf, -np.inf], np.nan).fillna(0)

print("Shape final del dataset de modelado:", rfm.shape)
display(rfm.head())

Shape final del dataset de modelado: (5279, 16)

from sklearn.model_selection import train_test_split

X = rfm.drop(['Customer ID', 'target'], axis=1)
y = rfm['target']

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

print("X_train:", X_train.shape)
print("X_test:", X_test.shape)
print("Tasa target train:", round(y_train.mean(), 4))
print("Tasa target test:", round(y_test.mean(), 4))

X_train: (3695, 14)
X_test: (1584, 14)
Tasa target train: 0.4346
Tasa target test: 0.4343

from sklearn.preprocessing import StandardScaler
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import roc_auc_score

scaler = StandardScaler()

X_train_scaled = scaler.fit_transform(X_train)
X_test_scaled = scaler.transform(X_test)

lr = LogisticRegression(
    max_iter=2000,
    class_weight='balanced',
    random_state=42
)

lr.fit(X_train_scaled, y_train)
y_pred_lr = lr.predict_proba(X_test_scaled)[:, 1]

print("AUC Logistic:", roc_auc_score(y_test, y_pred_lr))

AUC Logistic: 0.7915314835963455

from sklearn.ensemble import RandomForestClassifier

rf = RandomForestClassifier(
    n_estimators=1200,
    max_depth=16,
    min_samples_split=4,
    min_samples_leaf=2,
    max_features='sqrt',
    class_weight='balanced_subsample',
    random_state=42,
    n_jobs=-1
)

rf.fit(X_train, y_train)
y_pred_rf = rf.predict_proba(X_test)[:, 1]
auc_rf = roc_auc_score(y_test, y_pred_rf)

print("AUC Random Forest:", auc_rf)

AUC Random Forest: 0.790160727263289

from xgboost import XGBClassifier

xgb = XGBClassifier(
    n_estimators=600,
    max_depth=4,
    learning_rate=0.03,
    subsample=0.85,
    colsample_bytree=0.85,
    min_child_weight=5,
    reg_lambda=1.0,
    objective='binary:logistic',
    eval_metric='auc',
    random_state=42,
    n_jobs=-1
)

xgb.fit(X_train, y_train)
y_pred_xgb = xgb.predict_proba(X_test)[:, 1]
auc_xgb = roc_auc_score(y_test, y_pred_xgb)

print("AUC XGBoost:", auc_xgb)

AUC XGBoost: 0.7926572882059799

from lightgbm import LGBMClassifier

lgbm = LGBMClassifier(
    n_estimators=500,
    learning_rate=0.03,
    num_leaves=31,
    max_depth=-1,
    min_child_samples=5,
    subsample=0.8,
    colsample_bytree=0.8,
    reg_lambda=1.0,
    random_state=42,
    verbosity=-1
)

lgbm.fit(X_train, y_train)
y_pred_lgbm = lgbm.predict_proba(X_test)[:, 1]
auc_lgbm = roc_auc_score(y_test, y_pred_lgbm)

print("AUC LightGBM:", auc_lgbm)

AUC LightGBM: 0.7828478638911961

from catboost import CatBoostClassifier

cat = CatBoostClassifier(
    iterations=700,
    depth=5,
    learning_rate=0.03,
    loss_function='Logloss',
    eval_metric='AUC',
    verbose=0,
    random_seed=42
)

cat.fit(X_train, y_train)
y_pred_cat = cat.predict_proba(X_test)[:, 1]
auc_cat = roc_auc_score(y_test, y_pred_cat)

print("AUC CatBoost:", auc_cat)

AUC CatBoost: 0.7945666138911961

# Ensemble ponderado de probabilidades
# Damos más peso a los modelos con mejor desempeño individual.
y_pred_ens = (
    0.20 * y_pred_rf +
    0.25 * y_pred_xgb +
    0.20 * y_pred_lgbm +
    0.35 * y_pred_cat
)
auc_ens = roc_auc_score(y_test, y_pred_ens)

results = pd.DataFrame({
    'Modelo': ['Logistic Regression', 'Random Forest', 'XGBoost', 'LightGBM', 'CatBoost', 'Ensemble ponderado'],
    'AUC': [roc_auc_score(y_test, y_pred_lr), auc_rf, auc_xgb, auc_lgbm, auc_cat, auc_ens]
}).sort_values('AUC', ascending=False)

display(results)
print("Mejor AUC base:", results.iloc[0]['AUC'])

Mejor AUC base: 0.7945666138911961

from sklearn.model_selection import GridSearchCV

rf_grid = RandomForestClassifier(
    class_weight='balanced_subsample',
    random_state=42,
    n_jobs=-1
)

param_grid = {
    'n_estimators': [400, 800],
    'max_depth': [10, 16, None],
    'min_samples_split': [2, 4, 8],
    'min_samples_leaf': [1, 2, 4],
    'max_features': ['sqrt', 'log2']
}

grid_search = GridSearchCV(
    estimator=rf_grid,
    param_grid=param_grid,
    scoring='roc_auc',
    cv=3,
    n_jobs=-1,
    verbose=1
)

grid_search.fit(X_train, y_train)

best_rf_grid = grid_search.best_estimator_
y_pred_rf_grid = best_rf_grid.predict_proba(X_test)[:, 1]
auc_rf_grid = roc_auc_score(y_test, y_pred_rf_grid)

print("Mejores parámetros RF (GridSearch):")
print(grid_search.best_params_)
print("AUC Random Forest + GridSearch:", auc_rf_grid)

Fitting 3 folds for each of 108 candidates, totalling 324 fits

c:\Users\alvar\Downloads\datasetsClase\.venv\Lib\site-packages\joblib\externals\loky\process_executor.py:782: UserWarning: A worker stopped while some jobs were given to the executor. This can be caused by a too short worker timeout or by a memory leak.
  warnings.warn(

Mejores parámetros RF (GridSearch):
{'max_depth': 10, 'max_features': 'sqrt', 'min_samples_leaf': 4, 'min_samples_split': 2, 'n_estimators': 400}
AUC Random Forest + GridSearch: 0.7964110517026579

!pip install optuna -q

import optuna
from sklearn.model_selection import StratifiedKFold, cross_val_score

[notice] A new release of pip is available: 25.3 -> 26.0.1
[notice] To update, run: python.exe -m pip install --upgrade pip
c:\Users\alvar\Downloads\datasetsClase\.venv\Lib\site-packages\tqdm\auto.py:21: TqdmWarning: IProgress not found. Please update jupyter and ipywidgets. See https://ipywidgets.readthedocs.io/en/stable/user_install.html
  from .autonotebook import tqdm as notebook_tqdm

def objective_lgbm(trial):
    params = {
        'n_estimators': trial.suggest_int('n_estimators', 300, 900),
        'learning_rate': trial.suggest_float('learning_rate', 0.01, 0.08, log=True),
        'num_leaves': trial.suggest_int('num_leaves', 15, 63),
        'max_depth': trial.suggest_int('max_depth', 3, 10),
        'min_child_samples': trial.suggest_int('min_child_samples', 5, 50),
        'subsample': trial.suggest_float('subsample', 0.7, 1.0),
        'colsample_bytree': trial.suggest_float('colsample_bytree', 0.7, 1.0),
        'reg_lambda': trial.suggest_float('reg_lambda', 0.1, 5.0),
        'random_state': 42,
        'verbosity': -1
    }

    model = LGBMClassifier(**params)
    cv = StratifiedKFold(n_splits=3, shuffle=True, random_state=42)

    scores = cross_val_score(
        model,
        X_train,
        y_train,
        scoring='roc_auc',
        cv=cv,
        n_jobs=-1
    )
    return scores.mean()

study = optuna.create_study(direction='maximize')
study.optimize(objective_lgbm, n_trials=20, show_progress_bar=False)

best_params_optuna = study.best_params
best_lgbm_optuna = LGBMClassifier(**best_params_optuna, random_state=42, verbosity=-1)

best_lgbm_optuna.fit(X_train, y_train)
y_pred_lgbm_optuna = best_lgbm_optuna.predict_proba(X_test)[:, 1]
auc_lgbm_optuna = roc_auc_score(y_test, y_pred_lgbm_optuna)

print("Mejores parámetros LGBM (Optuna):")
print(best_params_optuna)
print("Mejor AUC CV (Optuna):", study.best_value)
print("AUC LightGBM + Optuna en test:", auc_lgbm_optuna)

[I 2026-03-25 23:04:12,591] A new study created in memory with name: no-name-872f2309-cc99-4724-a0a2-0923918d9059
[I 2026-03-25 23:04:16,308] Trial 0 finished with value: 0.7761556033771383 and parameters: {'n_estimators': 333, 'learning_rate': 0.021693909142691793, 'num_leaves': 53, 'max_depth': 8, 'min_child_samples': 27, 'subsample': 0.7942663983418382, 'colsample_bytree': 0.9789797640832767, 'reg_lambda': 1.680759296232529}. Best is trial 0 with value: 0.7761556033771383.
[I 2026-03-25 23:04:18,413] Trial 1 finished with value: 0.7786905861709452 and parameters: {'n_estimators': 525, 'learning_rate': 0.025674731810447295, 'num_leaves': 15, 'max_depth': 6, 'min_child_samples': 23, 'subsample': 0.9283674705582567, 'colsample_bytree': 0.9834200345865146, 'reg_lambda': 3.8351031701781406}. Best is trial 1 with value: 0.7786905861709452.
[I 2026-03-25 23:04:20,698] Trial 2 finished with value: 0.7739142969190764 and parameters: {'n_estimators': 783, 'learning_rate': 0.03450314385829416, 'num_leaves': 45, 'max_depth': 4, 'min_child_samples': 13, 'subsample': 0.8826626840973919, 'colsample_bytree': 0.9486943930337538, 'reg_lambda': 3.8497443270777763}. Best is trial 1 with value: 0.7786905861709452.
[I 2026-03-25 23:04:24,478] Trial 3 finished with value: 0.7592569076643403 and parameters: {'n_estimators': 820, 'learning_rate': 0.06346491474367172, 'num_leaves': 37, 'max_depth': 5, 'min_child_samples': 36, 'subsample': 0.8084656106792252, 'colsample_bytree': 0.8251059086666744, 'reg_lambda': 2.9509926986878554}. Best is trial 1 with value: 0.7786905861709452.
[I 2026-03-25 23:04:25,419] Trial 4 finished with value: 0.7516532087217618 and parameters: {'n_estimators': 669, 'learning_rate': 0.07555101206655318, 'num_leaves': 55, 'max_depth': 9, 'min_child_samples': 45, 'subsample': 0.7416845991965556, 'colsample_bytree': 0.7250202517845193, 'reg_lambda': 2.498286954118494}. Best is trial 1 with value: 0.7786905861709452.
[I 2026-03-25 23:04:26,041] Trial 5 finished with value: 0.7639503634132369 and parameters: {'n_estimators': 733, 'learning_rate': 0.052004226901292444, 'num_leaves': 47, 'max_depth': 4, 'min_child_samples': 18, 'subsample': 0.8615207768311244, 'colsample_bytree': 0.7347884776460767, 'reg_lambda': 1.931818605587583}. Best is trial 1 with value: 0.7786905861709452.
[I 2026-03-25 23:04:29,696] Trial 6 finished with value: 0.7752942660374952 and parameters: {'n_estimators': 889, 'learning_rate': 0.016868037698458224, 'num_leaves': 15, 'max_depth': 9, 'min_child_samples': 22, 'subsample': 0.7807044217572258, 'colsample_bytree': 0.9483478542685527, 'reg_lambda': 3.337810394841639}. Best is trial 1 with value: 0.7786905861709452.
[I 2026-03-25 23:04:30,903] Trial 7 finished with value: 0.761584810344098 and parameters: {'n_estimators': 527, 'learning_rate': 0.03402389876799954, 'num_leaves': 62, 'max_depth': 10, 'min_child_samples': 26, 'subsample': 0.8038416603129885, 'colsample_bytree': 0.7698104325755993, 'reg_lambda': 0.9398564486937853}. Best is trial 1 with value: 0.7786905861709452.
[I 2026-03-25 23:04:31,320] Trial 8 finished with value: 0.7808878783613782 and parameters: {'n_estimators': 422, 'learning_rate': 0.02542901915342043, 'num_leaves': 21, 'max_depth': 5, 'min_child_samples': 27, 'subsample': 0.8904969599076645, 'colsample_bytree': 0.8282993388784089, 'reg_lambda': 2.782608450766556}. Best is trial 8 with value: 0.7808878783613782.
[I 2026-03-25 23:04:31,992] Trial 9 finished with value: 0.7739937701393685 and parameters: {'n_estimators': 414, 'learning_rate': 0.02669230416946984, 'num_leaves': 30, 'max_depth': 9, 'min_child_samples': 30, 'subsample': 0.912443238367347, 'colsample_bytree': 0.9343493470344374, 'reg_lambda': 4.233492591623715}. Best is trial 8 with value: 0.7808878783613782.
[I 2026-03-25 23:04:32,277] Trial 10 finished with value: 0.7915676600627773 and parameters: {'n_estimators': 330, 'learning_rate': 0.011127768135349133, 'num_leaves': 25, 'max_depth': 3, 'min_child_samples': 5, 'subsample': 0.9975575004334847, 'colsample_bytree': 0.8716798292897097, 'reg_lambda': 4.843425958424428}. Best is trial 10 with value: 0.7915676600627773.
[I 2026-03-25 23:04:32,537] Trial 11 finished with value: 0.7915991667294865 and parameters: {'n_estimators': 317, 'learning_rate': 0.010773599115626163, 'num_leaves': 26, 'max_depth': 3, 'min_child_samples': 6, 'subsample': 0.9819308713826748, 'colsample_bytree': 0.8654603945175494, 'reg_lambda': 4.962261438384296}. Best is trial 11 with value: 0.7915991667294865.
[I 2026-03-25 23:04:32,799] Trial 12 finished with value: 0.7915341353332437 and parameters: {'n_estimators': 303, 'learning_rate': 0.01041687929199374, 'num_leaves': 28, 'max_depth': 3, 'min_child_samples': 5, 'subsample': 0.9707110131324157, 'colsample_bytree': 0.8759339826059921, 'reg_lambda': 4.54375128121747}. Best is trial 11 with value: 0.7915991667294865.
[I 2026-03-25 23:04:33,091] Trial 13 finished with value: 0.7916634216986104 and parameters: {'n_estimators': 393, 'learning_rate': 0.01017310601481108, 'num_leaves': 28, 'max_depth': 3, 'min_child_samples': 5, 'subsample': 0.9859434699660597, 'colsample_bytree': 0.8787731878387044, 'reg_lambda': 4.962078509299207}. Best is trial 13 with value: 0.7916634216986104.
[I 2026-03-25 23:04:34,070] Trial 14 finished with value: 0.7813009513763335 and parameters: {'n_estimators': 446, 'learning_rate': 0.014943697314221443, 'num_leaves': 35, 'max_depth': 7, 'min_child_samples': 12, 'subsample': 0.9523154404936991, 'colsample_bytree': 0.908356909891674, 'reg_lambda': 4.865283497961779}. Best is trial 13 with value: 0.7916634216986104.
[I 2026-03-25 23:04:34,453] Trial 15 finished with value: 0.7887864349625363 and parameters: {'n_estimators': 581, 'learning_rate': 0.0140778352439255, 'num_leaves': 33, 'max_depth': 3, 'min_child_samples': 11, 'subsample': 0.9913416871721442, 'colsample_bytree': 0.7961482813847338, 'reg_lambda': 4.9941664929359435}. Best is trial 13 with value: 0.7916634216986104.
[I 2026-03-25 23:04:34,955] Trial 16 finished with value: 0.7886673022744892 and parameters: {'n_estimators': 385, 'learning_rate': 0.010162577727625095, 'num_leaves': 22, 'max_depth': 5, 'min_child_samples': 17, 'subsample': 0.9409792456856579, 'colsample_bytree': 0.8628385353249751, 'reg_lambda': 3.592800926738331}. Best is trial 13 with value: 0.7916634216986104.
[I 2026-03-25 23:04:35,335] Trial 17 finished with value: 0.7848328770671689 and parameters: {'n_estimators': 492, 'learning_rate': 0.018567606987487, 'num_leaves': 42, 'max_depth': 4, 'min_child_samples': 49, 'subsample': 0.9656398057829962, 'colsample_bytree': 0.9014329475638707, 'reg_lambda': 4.257141922322578}. Best is trial 13 with value: 0.7916634216986104.
[I 2026-03-25 23:04:36,176] Trial 18 finished with value: 0.7796551204422402 and parameters: {'n_estimators': 607, 'learning_rate': 0.013451153872346532, 'num_leaves': 22, 'max_depth': 6, 'min_child_samples': 10, 'subsample': 0.8351174998576998, 'colsample_bytree': 0.8331146871382538, 'reg_lambda': 0.3692061053550084}. Best is trial 13 with value: 0.7916634216986104.
[I 2026-03-25 23:04:36,402] Trial 19 finished with value: 0.7917877365215421 and parameters: {'n_estimators': 356, 'learning_rate': 0.012993922539651297, 'num_leaves': 31, 'max_depth': 3, 'min_child_samples': 35, 'subsample': 0.7085778091755379, 'colsample_bytree': 0.7778303405256466, 'reg_lambda': 4.236217992563103}. Best is trial 19 with value: 0.7917877365215421.

Mejores parámetros LGBM (Optuna):
{'n_estimators': 356, 'learning_rate': 0.012993922539651297, 'num_leaves': 31, 'max_depth': 3, 'min_child_samples': 35, 'subsample': 0.7085778091755379, 'colsample_bytree': 0.7778303405256466, 'reg_lambda': 4.236217992563103}
Mejor AUC CV (Optuna): 0.7917877365215421
AUC LightGBM + Optuna en test: 0.8028633396490865

# Comparativa ampliada con modelos optimizados
results_opt = pd.DataFrame({
    'Modelo': [
        'Logistic Regression',
        'Random Forest',
        'XGBoost',
        'LightGBM',
        'CatBoost',
        'Ensemble ponderado',
        'Random Forest + GridSearch',
        'LightGBM + Optuna'
    ],
    'AUC': [
        roc_auc_score(y_test, y_pred_lr),
        auc_rf,
        auc_xgb,
        auc_lgbm,
        auc_cat,
        auc_ens,
        auc_rf_grid,
        auc_lgbm_optuna
    ]
}).sort_values('AUC', ascending=False)

display(results_opt)
print("Mejor modelo tras optimización:", results_opt.iloc[0]['Modelo'])
print("Mejor AUC tras optimización:", results_opt.iloc[0]['AUC'])

pred_map = {
    'Logistic Regression': y_pred_lr,
    'Random Forest': y_pred_rf,
    'XGBoost': y_pred_xgb,
    'LightGBM': y_pred_lgbm,
    'CatBoost': y_pred_cat,
    'Ensemble ponderado': y_pred_ens,
    'Random Forest + GridSearch': y_pred_rf_grid,
    'LightGBM + Optuna': y_pred_lgbm_optuna
}

best_model_name = results_opt.iloc[0]['Modelo']
best_pred = pred_map[best_model_name]
print("Se utilizará este modelo para el análisis post-modelado:", best_model_name)

Mejor modelo tras optimización: LightGBM + Optuna
Mejor AUC tras optimización: 0.8028633396490865
Se utilizará este modelo para el análisis post-modelado: LightGBM + Optuna

# Importancia de variables del Random Forest
importances = pd.Series(rf.feature_importances_, index=X.columns)
importances = importances.sort_values(ascending=False)

display(importances.head(15))

plt.figure(figsize=(10, 6))
importances.head(15).sort_values().plot(kind='barh')
plt.title('Top 15 variables más importantes - Random Forest')
plt.xlabel('Importancia')
plt.show()

monetary_per_recency       0.138391
freq_per_recency           0.135664
Recency                    0.125484
Monetary                   0.092936
total_quantity             0.081358
num_descriptions           0.065894
num_products               0.065649
last_invoice_amount        0.061170
mean_invoice_amount        0.060220
avg_ticket                 0.060220
Frequency                  0.054921
ratio_last_invoice_mean    0.051016
is_recent_customer         0.007075
num_countries              0.000001
dtype: float64

from sklearn.metrics import roc_curve

fpr_rf, tpr_rf, _ = roc_curve(y_test, y_pred_rf)
fpr_best, tpr_best, _ = roc_curve(y_test, best_pred)

plt.figure(figsize=(8, 6))
plt.plot(fpr_rf, tpr_rf, label=f'Random Forest base (AUC = {auc_rf:.3f})')
plt.plot(fpr_best, tpr_best, label=f'{best_model_name} (AUC = {results_opt.iloc[0]["AUC"]:.3f})')
plt.plot([0, 1], [0, 1], linestyle='--')
plt.title("Curvas ROC")
plt.xlabel("False Positive Rate")
plt.ylabel("True Positive Rate")
plt.legend()
plt.show()

from sklearn.metrics import precision_recall_curve

precision, recall, pr_thresholds = precision_recall_curve(y_test, best_pred)

plt.figure(figsize=(8, 6))
plt.plot(recall, precision)
plt.title(f"Curva Precision-Recall - {best_model_name}")
plt.xlabel("Recall")
plt.ylabel("Precision")
plt.show()

from sklearn.metrics import f1_score

thresholds = np.arange(0.10, 0.90, 0.01)
f1_scores = [f1_score(y_test, (best_pred >= t).astype(int)) for t in thresholds]

best_threshold = thresholds[int(np.argmax(f1_scores))]
best_f1 = max(f1_scores)

print("Threshold óptimo según F1:", round(float(best_threshold), 2))
print("F1 asociado:", round(float(best_f1), 4))

plt.figure(figsize=(8, 4))
plt.plot(thresholds, f1_scores)
plt.axvline(best_threshold, linestyle='--')
plt.title("F1 según threshold")
plt.xlabel("Threshold")
plt.ylabel("F1")
plt.show()

Threshold óptimo según F1: 0.34
F1 asociado: 0.7139

import dalex as dx

# Elegimos como modelo a explicar el Random Forest base
explainer_rf = dx.Explainer(
    rf,
    X_train,
    y_train,
    label='Random Forest'
)

# Importancia de variables con DALEX
vi_rf = explainer_rf.model_parts()
vi_rf.result.head()

Preparation of a new explainer is initiated

  -> data              : 3695 rows 14 cols
  -> target variable   : Parameter 'y' was a pandas.Series. Converted to a numpy.ndarray.
  -> target variable   : 3695 values
  -> model_class       : sklearn.ensemble._forest.RandomForestClassifier (default)
  -> label             : Random Forest
  -> predict function  : <function yhat_proba_default at 0x000001EBB27AA8D0> will be used (default)
  -> predict function  : Accepts pandas.DataFrame and numpy.ndarray.
  -> predicted values  : min = 0.00284, mean = 0.449, max = 0.999
  -> model type        : classification will be used (default)
  -> residual function : difference between y and yhat (default)
  -> residuals         : min = -0.657, mean = -0.0148, max = 0.643
  -> model_info        : package sklearn

A new explainer has been created!

# Visualización de importancia de variables
vi_rf.plot()

# Explicación local de un cliente concreto del test
cliente_idx = 0
obs = X_test.iloc[[cliente_idx]]

bd_rf = explainer_rf.predict_parts(obs, type='break_down')
bd_rf.result.head()

bd_rf.plot()

df_post = X_test.copy()
df_post['target_real'] = y_test.values
df_post['proba'] = best_pred

df_post['segment'] = pd.cut(
    df_post['proba'],
    bins=[0, 0.4, 0.7, 1.0],
    labels=['Low', 'Medium', 'High'],
    include_lowest=True
)

display(df_post.head())

segment_counts = df_post['segment'].value_counts().sort_index()

plt.figure(figsize=(6, 4))
segment_counts.plot(kind='bar')
plt.title("Número de clientes por segmento")
plt.xlabel("Segmento")
plt.ylabel("Número de clientes")
plt.show()

repurchase_rate = df_post.groupby('segment')['target_real'].mean()

plt.figure(figsize=(6, 4))
repurchase_rate.plot(kind='bar')
plt.title("Tasa real de recompra por segmento")
plt.xlabel("Segmento")
plt.ylabel("Proporción de recompra")
plt.show()

C:\Users\alvar\AppData\Local\Temp\ipykernel_25304\4035356772.py:1: FutureWarning: The default of observed=False is deprecated and will be changed to True in a future version of pandas. Pass observed=False to retain current behavior or observed=True to adopt the future default and silence this warning.
  repurchase_rate = df_post.groupby('segment')['target_real'].mean()

avg_revenue_segment = df_post.groupby('segment')['Monetary'].mean()

plt.figure(figsize=(6, 4))
avg_revenue_segment.plot(kind='bar')
plt.title("Revenue medio por segmento")
plt.xlabel("Segmento")
plt.ylabel("Revenue medio")
plt.show()

C:\Users\alvar\AppData\Local\Temp\ipykernel_25304\848586204.py:1: FutureWarning: The default of observed=False is deprecated and will be changed to True in a future version of pandas. Pass observed=False to retain current behavior or observed=True to adopt the future default and silence this warning.
  avg_revenue_segment = df_post.groupby('segment')['Monetary'].mean()

plt.figure(figsize=(7, 4))
plt.hist(df_post['proba'], bins=30)
plt.title("Distribución de probabilidades predichas")
plt.xlabel("Probabilidad de recompra")
plt.ylabel("Número de clientes")
plt.show()

df_clean['year'] = df_clean['InvoiceDate'].dt.year

revenue_year = df_clean.groupby('year')['line_total'].sum().reset_index()
revenue_year['prev_year'] = revenue_year['line_total'].shift(1)
revenue_year['growth_%'] = (
    (revenue_year['line_total'] - revenue_year['prev_year']) / revenue_year['prev_year']
) * 100

display(revenue_year)

plt.figure(figsize=(7, 4))
plt.bar(revenue_year['year'].astype(str), revenue_year['line_total'])
plt.title("Revenue por año")
plt.xlabel("Año")
plt.ylabel("Revenue")
plt.show()

# ============================================
# 9.1 Calidad de datos avanzada
# ============================================

import pandas as pd
import numpy as np

print("=== Duplicados ===")

dup_full = df.duplicated().sum()
print(f"Filas duplicadas (registro entero): {dup_full:,} ({100*dup_full/len(df):.2f}%)")

key_cols = ['Invoice', 'StockCode', 'Description']
subset_key_cols = [c for c in key_cols if c in df.columns]
if subset_key_cols:
    dup_key = df.duplicated(subset=subset_key_cols).sum()
    print(f"Filas duplicadas por {subset_key_cols}: {dup_key:,} ({100*dup_key/len(df):.2f}%)")

print("\n=== Valores negativos ===")
for col in ['Quantity', 'Price']:
    if col in df.columns:
        n_neg = int((df[col] < 0).sum())
        print(f"{col}: {n_neg:,} valores negativos ({100*n_neg/len(df):.2f}%)")

numeric_cols = df.select_dtypes(include='number').columns
negatives = {c: int((df[c] < 0).sum()) for c in numeric_cols}
negatives = {k: v for k, v in negatives.items() if v > 0}

if negatives:
    display(
        pd.Series(negatives, name='negativos')
        .sort_values(ascending=False)
        .to_frame()
    )
else:
    print("No hay negativos en columnas numéricas.")

# Muestra rápida de filas con algún negativo
if negatives:
    cols_neg = list(negatives.keys())
    mask_neg = (df[cols_neg] < 0).any(axis=1)
    print(f"\nMuestra de filas con al menos un negativo ({mask_neg.sum():,} filas):")
    display(df.loc[mask_neg, ['Invoice', 'StockCode', 'Description', 'Quantity', 'Price', 'Customer ID', 'InvoiceDate']].head(10))

=== Duplicados ===
Filas duplicadas (registro entero): 34,335 (3.22%)
Filas duplicadas por ['Invoice', 'StockCode', 'Description']: 45,945 (4.30%)

=== Valores negativos ===
Quantity: 22,950 valores negativos (2.15%)
Price: 5 valores negativos (0.00%)

Muestra de filas con al menos un negativo (22,955 filas):

# ============================================
# 9.2 Distribución de Quantity (P95 / P99)
# ============================================

import matplotlib.pyplot as plt

qty_col = 'Quantity'

if qty_col in df.columns:
    qty_pos = df.loc[df[qty_col] > 0, qty_col].dropna()

    q95 = qty_pos.quantile(0.95)
    q99 = qty_pos.quantile(0.99)

    fig, axes = plt.subplots(1, 2, figsize=(12, 4))

    data_p95 = qty_pos.clip(upper=q95)
    bins_p95 = min(60, max(20, int(data_p95.max())))
    axes[0].hist(data_p95, bins=bins_p95, range=(0, q95), edgecolor='black', alpha=0.7)
    axes[0].set_title(f"Distribución de Quantity (0 a P95 = {q95:.0f})")
    axes[0].set_xlabel("Quantity")
    axes[0].set_ylabel("Frecuencia")

    data_p99 = qty_pos.clip(upper=q99)
    bins_p99 = min(80, max(20, int(data_p99.max())))
    axes[1].hist(data_p99, bins=bins_p99, range=(0, q99), edgecolor='black', alpha=0.7)
    axes[1].set_title(f"Quantity (0 a P99 = {q99:.0f})")
    axes[1].set_xlabel("Quantity")
    axes[1].set_ylabel("Frecuencia")

    plt.tight_layout()
    plt.show()
else:
    print("No existe la columna 'Quantity'.")

# ============================================
# 9.3 Distribución de Price (P95)
# ============================================

price_col = 'Price'

if price_col in df.columns:
    price_pos = df.loc[df[price_col] > 0, price_col].dropna()
    q95_price = price_pos.quantile(0.95)

    data_price = price_pos.clip(upper=q95_price)
    bins_price = min(60, max(20, int(data_price.max() * 2)))

    plt.figure(figsize=(8, 4))
    plt.hist(data_price, bins=bins_price, range=(0, q95_price), edgecolor='black', alpha=0.7)
    plt.title(f"Distribución de Price (0 a P95 = {q95_price:.2f})")
    plt.xlabel("Price")
    plt.ylabel("Frecuencia")
    plt.tight_layout()
    plt.show()
else:
    print("No existe la columna 'Price'.")

# ============================================
# 9.4 Top 20 productos más vendidos
# ============================================

if all(c in df.columns for c in ['Description', 'Quantity']):
    top_products = (
        df.loc[(df['Quantity'] > 0) & (df['Description'].notna())]
        .groupby('Description', as_index=False)['Quantity']
        .sum()
        .sort_values('Quantity', ascending=False)
        .head(20)
    )

    plt.figure(figsize=(10, 7))
    plt.barh(top_products['Description'][::-1], top_products['Quantity'][::-1])
    plt.title("Top 20 productos más vendidos")
    plt.xlabel("Cantidad vendida")
    plt.ylabel("Description")
    plt.tight_layout()
    plt.show()

    display(top_products)
else:
    print("Faltan columnas necesarias: 'Description' y/o 'Quantity'.")

# ============================================
# 9.5 Transacciones por mes
# ============================================

if 'InvoiceDate' in df.columns:
    df_ts = df.dropna(subset=['InvoiceDate']).copy()
    df_ts['InvoiceDate'] = pd.to_datetime(df_ts['InvoiceDate'], errors='coerce')
    df_ts = df_ts.dropna(subset=['InvoiceDate'])

    trans_month = (
        df_ts.set_index('InvoiceDate')
        .resample('MS')
        .size()
    )

    plt.figure(figsize=(12, 4))
    plt.plot(trans_month.index, trans_month.values)
    plt.title("Transacciones por mes")
    plt.xlabel("InvoiceDate")
    plt.ylabel("Nº transacciones")
    plt.tight_layout()
    plt.show()

    display(trans_month.to_frame("n_transacciones").tail(12))
else:
    print("No existe la columna 'InvoiceDate'.")

# ============================================
# 9.7 Revenue trimestral: 2010 vs 2011
# ============================================

if all(c in df.columns for c in ['InvoiceDate', 'Quantity', 'Price']):
    v = df.dropna(subset=['InvoiceDate']).copy()
    v['InvoiceDate'] = pd.to_datetime(v['InvoiceDate'], errors='coerce')
    v = v.dropna(subset=['InvoiceDate'])

    # Solo ventas positivas para comparar revenue comercial real
    v = v[(v['Quantity'] > 0) & (v['Price'] > 0)].copy()
    v['Revenue'] = v['Quantity'] * v['Price']
    v['year'] = v['InvoiceDate'].dt.year
    v['quarter'] = 'Q' + v['InvoiceDate'].dt.quarter.astype(str)

    v = v[v['year'].isin([2010, 2011])]

    tot = (
        v.groupby(['quarter', 'year'])['Revenue']
        .sum()
        .unstack(fill_value=0)
        .reindex(index=['Q1', 'Q2', 'Q3', 'Q4'])
        .fillna(0)
    )

    ax = tot.plot(kind='bar', figsize=(9, 5), edgecolor='black', alpha=0.9)
    ax.set_title("Total ventas por trimestre (Revenue): 2010 vs 2011")
    ax.set_xlabel("Trimestre")
    ax.set_ylabel("Ventas totales")
    plt.xticks(rotation=0)
    plt.tight_layout()
    plt.show()

    display(tot.round(2))
else:
    print("Faltan columnas necesarias: 'InvoiceDate', 'Quantity' y/o 'Price'.")

# ============================================
# 9.8 Heatmap top productos por mes
# ============================================

import seaborn as sns

needed_cols = ['InvoiceDate', 'StockCode', 'Description', 'Quantity']
if all(c in df.columns for c in needed_cols):
    hm = df.copy()
    hm['InvoiceDate'] = pd.to_datetime(hm['InvoiceDate'], errors='coerce')
    hm = hm.dropna(subset=['InvoiceDate', 'Description'])
    hm = hm[hm['Quantity'] > 0].copy()

    hm['year'] = hm['InvoiceDate'].dt.year
    hm['month'] = hm['InvoiceDate'].dt.to_period('M').astype(str)

    # Puedes fijar 2010 para que quede más limpio visualmente
    hm_2010 = hm[hm['year'] == 2010].copy()

    top_n = 12
    top_products_hm = (
        hm_2010.groupby(['StockCode', 'Description'])['Quantity']
        .sum()
        .sort_values(ascending=False)
        .head(top_n)
        .reset_index()
    )

    top_keys = set(zip(top_products_hm['StockCode'], top_products_hm['Description']))

    hm_2010['product_key'] = list(zip(hm_2010['StockCode'], hm_2010['Description']))
    hm_2010 = hm_2010[hm_2010['product_key'].isin(top_keys)].copy()
    hm_2010['product_label'] = hm_2010['StockCode'].astype(str) + " | " + hm_2010['Description']

    pivot_m = hm_2010.pivot_table(
        index='product_label',
        columns='month',
        values='Quantity',
        aggfunc='sum',
        fill_value=0
    )

    plt.figure(figsize=(14, 7))
    sns.heatmap(pivot_m, cmap='YlOrRd', linewidths=0.5, cbar_kws={'label': 'Cantidad'})
    plt.title("Top productos por cantidad vendida — por mes (2010)")
    plt.xlabel("Mes")
    plt.ylabel("Producto")
    plt.tight_layout()
    plt.show()
else:
    print("Faltan columnas necesarias para el heatmap.")

	Invoice	StockCode	Description	Quantity	InvoiceDate	Price	Customer ID	Country
0	489434	85048	15CM CHRISTMAS GLASS BALL 20 LIGHTS	12	2009-12-01 07:45:00	6.95	13085.0	United Kingdom
1	489434	79323P	PINK CHERRY LIGHTS	12	2009-12-01 07:45:00	6.75	13085.0	United Kingdom
2	489434	79323W	WHITE CHERRY LIGHTS	12	2009-12-01 07:45:00	6.75	13085.0	United Kingdom
3	489434	22041	RECORD FRAME 7" SINGLE SIZE	48	2009-12-01 07:45:00	2.10	13085.0	United Kingdom
4	489434	21232	STRAWBERRY CERAMIC TRINKET BOX	24	2009-12-01 07:45:00	1.25	13085.0	United Kingdom
5	489434	22064	PINK DOUGHNUT TRINKET POT	24	2009-12-01 07:45:00	1.65	13085.0	United Kingdom
6	489434	21871	SAVE THE PLANET MUG	24	2009-12-01 07:45:00	1.25	13085.0	United Kingdom
7	489434	21523	FANCY FONT HOME SWEET HOME DOORMAT	10	2009-12-01 07:45:00	5.95	13085.0	United Kingdom
8	489435	22350	CAT BOWL	12	2009-12-01 07:46:00	2.55	13085.0	United Kingdom
9	489435	22349	DOG BOWL , CHASING BALL DESIGN	12	2009-12-01 07:46:00	3.75	13085.0	United Kingdom

	Invoice	StockCode	Description	Quantity	InvoiceDate	Price	Customer ID	Country
count	1067371.0	1067371	1062989	1.067371e+06	1067371	1.067371e+06	824364.000000	1067371
unique	53628.0	5305	5698	NaN	NaN	NaN	NaN	43
top	537434.0	85123A	WHITE HANGING HEART T-LIGHT HOLDER	NaN	NaN	NaN	NaN	United Kingdom
freq	1350.0	5829	5918	NaN	NaN	NaN	NaN	981330
mean	NaN	NaN	NaN	9.938898e+00	2011-01-02 21:13:55.394028544	4.649388e+00	15324.638504	NaN
min	NaN	NaN	NaN	-8.099500e+04	2009-12-01 07:45:00	-5.359436e+04	12346.000000	NaN
25%	NaN	NaN	NaN	1.000000e+00	2010-07-09 09:46:00	1.250000e+00	13975.000000	NaN
50%	NaN	NaN	NaN	3.000000e+00	2010-12-07 15:28:00	2.100000e+00	15255.000000	NaN
75%	NaN	NaN	NaN	1.000000e+01	2011-07-22 10:23:00	4.150000e+00	16797.000000	NaN
max	NaN	NaN	NaN	8.099500e+04	2011-12-09 12:50:00	3.897000e+04	18287.000000	NaN
std	NaN	NaN	NaN	1.727058e+02	NaN	1.235531e+02	1697.464450	NaN

	Invoice	StockCode	Description	Quantity	InvoiceDate	Price	Customer ID	Country	line_total
0	489434	85048	15CM CHRISTMAS GLASS BALL 20 LIGHTS	12	2009-12-01 07:45:00	6.95	13085	United Kingdom	83.4
1	489434	79323P	PINK CHERRY LIGHTS	12	2009-12-01 07:45:00	6.75	13085	United Kingdom	81.0
2	489434	79323W	WHITE CHERRY LIGHTS	12	2009-12-01 07:45:00	6.75	13085	United Kingdom	81.0
3	489434	22041	RECORD FRAME 7" SINGLE SIZE	48	2009-12-01 07:45:00	2.10	13085	United Kingdom	100.8
4	489434	21232	STRAWBERRY CERAMIC TRINKET BOX	24	2009-12-01 07:45:00	1.25	13085	United Kingdom	30.0

	Customer ID	Recency	Frequency	Monetary
0	12346	235	12	77556.46
1	12347	39	6	3402.39
2	12348	158	4	1709.40
3	12349	317	3	2671.14
4	12350	219	1	334.40

	Customer ID	Recency	Frequency	Monetary	avg_ticket	num_products	num_descriptions	num_countries	total_quantity	last_invoice_amount	mean_invoice_amount	ratio_last_invoice_mean	freq_per_recency	monetary_per_recency	target
0	12346	235	12	77556.46	6463.038333	27	26	1	74285	77183.60	6463.038333	11.942309	0.050847	328.629068	0
1	12347	39	6	3402.39	567.065000	107	107	1	2099	584.91	567.065000	1.031469	0.150000	85.059750	1
2	12348	158	4	1709.40	427.350000	25	28	1	2497	367.00	427.350000	0.858781	0.025157	10.750943	1
3	12349	317	3	2671.14	890.380000	90	94	1	993	1402.62	890.380000	1.575305	0.009434	8.399811	1
4	12350	219	1	334.40	334.400000	17	17	1	197	334.40	334.400000	1.000000	0.004545	1.520000	0

	Modelo	AUC
4	CatBoost	0.794567
5	Ensemble ponderado	0.793958
2	XGBoost	0.792657
0	Logistic Regression	0.791531
1	Random Forest	0.790161
3	LightGBM	0.782848

	Modelo	AUC
7	LightGBM + Optuna	0.802863
6	Random Forest + GridSearch	0.796411
4	CatBoost	0.794567
5	Ensemble ponderado	0.793958
2	XGBoost	0.792657
0	Logistic Regression	0.791531
1	Random Forest	0.790161
3	LightGBM	0.782848

	variable	dropout_loss	label
0	_full_model_	0.002483	Random Forest
1	num_countries	0.002483	Random Forest
2	is_recent_customer	0.003274	Random Forest
3	Frequency	0.010619	Random Forest
4	ratio_last_invoice_mean	0.011417	Random Forest

	variable_name	variable_value	variable	cumulative	contribution	sign	position	label
0	intercept		intercept	0.449403	4.494029e-01	1.0	15	Random Forest
1	mean_invoice_amount	211.5	mean_invoice_amount = 211.5	0.467727	1.832365e-02	1.0	14	Random Forest
2	avg_ticket	211.5	avg_ticket = 211.5	0.482562	1.483512e-02	1.0	13	Random Forest
3	last_invoice_amount	211.5	last_invoice_amount = 211.5	0.485878	3.316097e-03	1.0	12	Random Forest
4	num_countries	1.0	num_countries = 1.0	0.485878	-9.858871e-08	-1.0	11	Random Forest

	Recency	Frequency	Monetary	avg_ticket	num_products	num_descriptions	num_countries	total_quantity	last_invoice_amount	mean_invoice_amount	ratio_last_invoice_mean	freq_per_recency	monetary_per_recency	target_real	proba	segment
2977	174	1	211.50	211.500000	5	5	1	66	211.50	211.500000	1.000000	0.005714	1.208571	0	0.257654	Low
1190	212	2	631.17	315.585000	38	38	1	301	330.79	315.585000	1.048180	0.009390	2.963239	0	0.297453	Low
4050	57	8	1201.12	150.140000	94	94	1	382	127.88	150.140000	0.851738	0.137931	20.708966	1	0.678307	Medium
1840	421	3	1220.90	406.966667	29	29	1	446	34.90	406.966667	0.085756	0.007109	2.893128	0	0.229586	Low
4555	509	1	1047.00	1047.000000	6	6	1	200	1047.00	1047.000000	1.000000	0.001961	2.052941	0	0.109135	Low

	year	line_total	prev_year	growth_%
0	2009	683504.010	NaN	NaN
1	2010	8374496.094	683504.010	1125.229987
2	2011	8316804.164	8374496.094	-0.688900

	Invoice	StockCode	Description	Quantity	Price	Customer ID	InvoiceDate
178	C489449	22087	PAPER BUNTING WHITE LACE	-12	2.95	16321.0	2009-12-01 10:33:00
179	C489449	85206A	CREAM FELT EASTER EGG BASKET	-6	1.65	16321.0	2009-12-01 10:33:00
180	C489449	21895	POTTING SHED SOW 'N' GROW SET	-4	4.25	16321.0	2009-12-01 10:33:00
181	C489449	21896	POTTING SHED TWINE	-6	2.10	16321.0	2009-12-01 10:33:00
182	C489449	22083	PAPER CHAIN KIT RETRO SPOT	-12	2.95	16321.0	2009-12-01 10:33:00
183	C489449	21871	SAVE THE PLANET MUG	-12	1.25	16321.0	2009-12-01 10:33:00
184	C489449	84946	ANTIQUE SILVER TEA GLASS ETCHED	-12	1.25	16321.0	2009-12-01 10:33:00
185	C489449	84970S	HANGING HEART ZINC T-LIGHT HOLDER	-24	0.85	16321.0	2009-12-01 10:33:00
186	C489449	22090	PAPER BUNTING RETRO SPOTS	-12	2.95	16321.0	2009-12-01 10:33:00
196	C489459	90200A	PURPLE SWEETHEART BRACELET	-3	4.25	17592.0	2009-12-01 10:44:00

	Description	Quantity
5305	WORLD WAR 2 GLIDERS ASSTD DESIGNS	110249
5184	WHITE HANGING HEART T-LIGHT HOLDER	96688
307	ASSORTED COLOUR BIRD ORNAMENT	81817
3205	PAPER CRAFT , LITTLE BIRDIE	80995
2455	JUMBO BAG RED RETROSPOT	79285
2765	MEDIUM CERAMIC TOP STORAGE JAR	78033
776	BROCADE RING PURSE	71440
3141	PACK OF 60 PINK PAISLEY CAKE CASES	56794
161	60 TEATIME FAIRY CAKE CASES	54716
4634	SMALL POPCORN HOLDER	49984
3144	PACK OF 72 RETROSPOT CAKE CASES	49805
3143	PACK OF 72 RETRO SPOT CAKE CASES	46755
319	ASSORTED COLOURS SILK FAN	44485
2830	MINI PAINT SET VINTAGE	44405
1142	COLOUR GLASS T-LIGHT HOLDER HANGING	40839
4980	VICTORIAN GLASS HANGING T-LIGHT	40748
3145	PACK OF 72 SKULL CAKE CASES	40733
3106	PACK OF 12 LONDON TISSUES	40367
1982	GIRLS ALPHABET IRON ON PATCHES	38574
3710	RED HARMONICA IN BOX	38191

	n_transacciones
InvoiceDate
2011-01-01	35147
2011-02-01	27707
2011-03-01	36748
2011-04-01	29916
2011-05-01	37030
2011-06-01	36874
2011-07-01	39518
2011-08-01	35284
2011-09-01	50226
2011-10-01	60742
2011-11-01	84711
2011-12-01	25526

year	2010	2011
quarter
Q1	2039991.94	1932635.81
Q2	2093657.99	2070084.54
Q3	2272320.86	2536949.74
Q4	3898355.18	3303286.31

Análisis exploratorio: Online Retail II¶

Análisis exploratorio del dataset¶

1. Carga de datos¶

2. Primera inspección¶

3. Calidad de datos¶

🔹 Valores nulos¶

🔹 Duplicados¶

🔹 Resumen del dataset¶

4. Estadísticas descriptivas¶

5. Visualizaciones¶

Quantity y Price¶

5.1 Limpieza reutilizable¶

Limpieza y preparación de datos¶

Qué hago en este paso¶

Por qué lo hago así¶

6. Preparación del problema supervisado¶

Definición del problema de modelado¶

Enfoque que he seguido¶

Por qué es importante¶

Construcción del target¶

6.1 Construcción de variables a nivel cliente¶

Construcción de variables a nivel cliente¶

Por qué agrego a nivel cliente¶

Variables principales (RFM)¶

Variables adicionales¶

Qué busco con esto¶

7. Modelado¶

Modelado y comparación de algoritmos¶

Separación train/test¶

Modelos utilizados¶

Por qué esta elección¶

Métrica utilizada¶

Ensemble¶

Conclusión de este bloque¶

7.1 Optimización de hiperparámetros:¶

Optimización del modelo con Optuna y LightGBM¶

Por qué LightGBM¶

Por qué Optuna¶

Validación¶

Resultado¶

Valor del bloque¶

7.2 Optimización de hiperparámetros:¶

8. Modelo: explicabilidad¶

Variables más relevantes:¶

Curvas ROC¶

8.1 Evaluación adicional: Precision-Recall y umbral de decisión¶

8.1 Explicabilidad con DALEX¶

Explicabilidad del modelo con DALEX¶

Por qué añado explicabilidad¶

Qué hago¶

Qué aporta¶

Conclusión¶

9. EDA post-modelado orientado a negocio¶

Traducción del modelo a negocio¶

Qué hago¶

Por qué segmentar¶

Qué analizo¶

Qué busco¶

Valor del bloque¶

10. Conclusiones¶

Conclusiones del proyecto¶

Principales conclusiones¶

Reflexión final¶