objectif

Entraîner CatBoostClassifier avec auto-encodage des catégories.

code minimal

import pandas as pd
from catboost import CatBoostClassifier, Pool

df = pd.DataFrame({
    "cat1": ["a","b","a","c"],
    "cat2": ["x","x","y","y"],
    "num": [1.0, 2.0, 3.0, 4.0],
    "y": [0,1,0,1],
})
X = df[["cat1","cat2","num"]]
y = df["y"]
cat_idx = [0,1]

train_pool = Pool(X, y, cat_features=cat_idx)
clf = CatBoostClassifier(
    depth=6,
    learning_rate=0.1,
    iterations=500,
    loss_function="Logloss",
    random_seed=0,
    verbose=False,
)
clf.fit(train_pool)
print(clf.predict_proba(train_pool).shape[1] == 2)

utilisation

# Prédiction sur DataFrame brut (mêmes colonnes/ordre)
proba = clf.predict_proba(X)[:,1]
print(len(proba) == len(X))

variante(s) utile(s)

# Importance des features (PredictionValuesChange)
imp = clf.get_feature_importance(train_pool, type="PredictionValuesChange")
print(len(imp) == X.shape[1])

notes

• CatBoost gère nativement les catégories (encodage target statistics avec régularisation).