如何通过GridSearchCV获取带列名的特征重要性分数？

我想要获取变量的特征重要性分数，并将其与列名关联生成DataFrame。我使用GridSearchCV进行调参，代码如下：

grid_search = GridSearchCV(model, parameters, cv=10) #tuning
grid_search.fit(X_train2, y_train2.values.ravel())

通过grid_search.best_estimator_.named_steps["regressor"].feature_importances_得到的分数数组长度大于数据集列数，无法匹配列名。请问能否直接从GridSearchCV输出带有对应列名的特征重要性分数？以下是我的模型Pipeline供参考：

# Define the pipeline
numeric_transformer = Pipeline(steps=[
    #('imputer', KNNImputer()), # for missing values
    ('scaler', StandardScaler()), # standardizing
    #('scaler', MinMaxScaler()), # normalizing
    #('to_df', FunctionTransformer(lambda x: pd.DataFrame(x, columns=X_train2.select_dtypes(include=['int64', 'float64']).columns)))
])

categorical_transformer = Pipeline(steps=[
    ('onehot', OneHotEncoder(handle_unknown='ignore'))
])

preprocessor = ColumnTransformer(
    transformers=[
        ('num', numeric_transformer, X_train2.select_dtypes(include=['int64', float64']).columns.tolist()),
        ('cat', categorical_transformer, categorical_cols),
        # ("pca", PCA(random_state=548, n_components=25), indices_pca),
        # ('smotenc', SmoteNCWrapper(categorical_features=[1, 2], random_state=548)),

    ])

model = Pipeline(steps=[
    # ('over', SMOTE(random_state=548)),
    # ('smotenc', SmoteNCWrapper(categorical_features=[1, 2], random_state=548)),
    ('preprocessor', preprocessor),
    ('regressor', XGBClassifier(random_state=548))
])

parameters = {  'regressor__n_estimators': [1000],
    'regressor__max_depth': [5],
    'regressor__learning_rate': [0.01],
    # 'regressor__num_leaves': [31],
    # 'regressor__min_child_samples': [20],
    'regressor__reg_alpha': [0.1], 
    'regressor__reg_lambda': [0.1]
}

问题根源是分类特征经过OneHotEncoder后会被拆分成多个哑变量，导致预处理后的特征总数远大于原始数据集列数，直接用原始列名匹配必然出错。可以通过以下步骤获取带对应名称的特征重要性：

1. 提取预处理后的所有特征名称
   从GridSearchCV的最佳模型中取出预处理模块，分别获取数值特征和独热编码后的分类特征名称：

import pandas as pd

# 获取最佳模型的预处理步骤
best_preprocessor = grid_search.best_estimator_.named_steps['preprocessor']

# 获取数值特征列名
num_features = best_preprocessor.transformers_[0][2]

# 获取独热编码后的分类特征列名
cat_encoder = best_preprocessor.named_transformers_['cat'].named_steps['onehot']
cat_features = cat_encoder.get_feature_names_out(categorical_cols)

# 合并所有特征名称
all_features = list(num_features) + list(cat_features)

2. 提取特征重要性并生成DataFrame

# 获取特征重要性分数
feature_importances = grid_search.best_estimator_.named_steps['regressor'].feature_importances_

# 生成DataFrame
importance_df = pd.DataFrame({
    'feature': all_features,
    'importance': feature_importances
}).sort_values(by='importance', ascending=False)

生成的DataFrame会准确对应每个预处理后的特征及其重要性分数。如果需要对同一原始分类特征的哑变量分数进行聚合，可根据特征名称前缀分组求和。

内容的提问来源于stack exchange，提问作者Tudor Orosz