StackingClassifier+GridSearchCV报错：Pipeline的C参数无效求助

问题背景 #

运行堆叠分类器结合网格搜索的代码时触发ValueError，提示Invalid parameter C for estimator Pipeline，但单独检查LogisticRegression确实包含C参数，参考示例调整逻辑后仍无法解决。

原代码 #

df_final = df.copy()

X, y = df_final.drop("Churn", axis=1), df_final.Churn

lr_pipeline = make_pipeline(preprocessor, under_sampling, LogisticRegression())
dt_pipeline = make_pipeline(preprocessor, under_sampling, DecisionTreeClassifier())
rf_pipeline = make_pipeline(preprocessor, under_sampling, RandomForestClassifier())
gb_pipeline = make_pipeline(preprocessor, under_sampling, XGBClassifier())

estimators = [
    ("lr", lr_pipeline),
    ("dt", dt_pipeline),
    ("rf", rf_pipeline),
    ("gb", gb_pipeline)
]

classifiers = StackingClassifier(estimators=estimators, final_estimator=RandomForestClassifier())

params = {
    'final_estimator__n_estimators': [10, 100, 1000],
    'rf__max_depth': [5, 10, 15],
    'rf__min_samples_split': [2, 5, 10],
    'lr__C': [0.1, 1, 10],
}

cv = KFold(n_splits=5, shuffle=True, random_state=42)

models = GridSearchCV(estimator=classifiers, param_grid=params, cv=cv, scoring='accuracy', n_jobs=-1)

models.fit(X, y)

print(f"Melhores parâmetros: {models.best_params_}")
print(f"Acurácia: {models.best_score_:.4f}") 

报错信息 #

ValueError: Invalid parameter C for estimator Pipeline(steps=[('columntransformer',
                 ColumnTransformer(remainder='passthrough',
                                   transformers=[('minmaxscaler',
                                                  MinMaxScaler(),
                                                  <sklearn.compose._column_transformer.make_column_selector object at 0x7f1dec598c10>),
                                                 ('onehotencoder',
                                                  OneHotEncoder(),
                                                  <sklearn.compose._column_transformer.make_column_selector object at 0x7f1dec598f40>),
                                                 ('functiontransformer',
                                                  FunctionTransformer(func=<function identity at 0x7f1dec58fc10>),
                                                  ['SeniorCitizen'])])),
                ('randomundersampler',
                 RandomUnderSampler(random_state=42,
                                    sampling_strategy='not minority')),
                ('logisticregression', LogisticRegression())]). Check the list of available parameters with `estimator.get_params().keys()`.

排查确认 #

单独检查LogisticRegression参数，确认包含C：

dict_keys(['C', 'class_weight', 'dual', 'fit_intercept', 'intercept_scaling', 'l1_ratio', 'max_iter', 'multi_class', 'n_jobs', 'penalty', 'random_state', 'solver', 'tol', 'verbose', 'warm_start'])

解决方案 #

问题出在GridSearchCV的参数命名路径不完整：
当estimator是嵌套结构（StackingClassifier包含Pipeline，Pipeline又包含模型）时，参数名需要完整的层级路径：

1. 使用make_pipeline创建的Pipeline，步骤名称是自动生成的小写类名（如LogisticRegression对应logisticregression，RandomForestClassifier对应randomforestclassifier）
2. 参数路径需遵循：堆叠内的estimator名称__Pipeline内的模型步骤名__模型参数

修改后的params字典：

params = {
    'final_estimator__n_estimators': [10, 100, 1000],
    # 补充RF在Pipeline内的步骤名
    'rf__randomforestclassifier__max_depth': [5, 10, 15],
    'rf__randomforestclassifier__min_samples_split': [2, 5, 10],
    # 补充LR在Pipeline内的步骤名
    'lr__logisticregression__C': [0.1, 1, 10],
}

验证方法 #

可通过打印StackingClassifier的所有参数键，确认正确的参数路径：

print(classifiers.get_params().keys())

输出中会包含lr__logisticregression__C、rf__randomforestclassifier__max_depth这类完整路径，即为正确参数名。

内容的提问来源于stack exchange，提问作者leandro.starke