اهمیت‌های تهی (null importances): اهمیت واقعاً معناداره؟

import numpy as np
import pandas as pd
from lightgbm import LGBMClassifier

def importances(X, y_arr, seed=42):
    m = LGBMClassifier(n_estimators=200, random_state=seed, verbose=-1)
    return m.fit(X, y_arr).feature_importances_

imp_reelle = importances(X_train, y_train.values)

rng = np.random.default_rng(0)
null_imps = np.array([
    importances(X_train, rng.permutation(y_train.values), seed=i)
    for i in range(50)
])  # 50 modèles sur cible mélangée

p95_null = np.percentile(null_imps, 95, axis=0)
bilan = pd.DataFrame({
    "imp_reelle": imp_reelle,
    "p95_null": p95_null,
    "significative": imp_reelle > p95_null,
}, index=X_train.columns).sort_values("imp_reelle", ascending=False)
print(bilan.head(6).round(1).to_string())
print("features significatives :", int(bilan["significative"].sum()))