如何在含ColumnTransformer的sklearn Pipeline中正确实现LightGBM早停验证

问题背景 #

需要搭建集成早停验证的LightGBM模型，并用Pipeline整合预处理与模型拟合、预测流程，但运行代码时触发数据类型错误。

原始代码 #

coltransformer = ColumnTransformer([
    ('cat', OneHotEncoder(sparse_output = False), cat_invar),
    ('num', 'passthrough', num_invar)])

lgbPipe = Pipeline([
    ('preprocess', coltransformer),
    ('lgb', LGBMClassifier())])

X_learn, X_val, Y_learn, Y_val = train_test_split(X, y, test_size = 0.2)

lgbPipe.fit(X_learn, Y_learn, lgb__eval_set = (X_val, Y_val))

错误信息 #

ValueError: DataFrame.dtypes for data must be int, float or bool.
Did not expect the data types in the following fields: Geography, Gender, Surname

问题根源 #

直接将原始验证集X_val传入lgb__eval_set时，LightGBM会读取未经过Pipeline预处理的原始数据——其中的分类特征未完成编码，不符合模型对输入数据类型的要求。

解决方案 #

必须先用Pipeline中已拟合的预处理组件转换验证集，再传入早停验证参数，以下是两种规范实现方式：

方式一：分步处理验证集（直观易懂） #

coltransformer = ColumnTransformer([
    ('cat', OneHotEncoder(sparse_output = False), cat_invar),
    ('num', 'passthrough', num_invar)])

lgbPipe = Pipeline([
    ('preprocess', coltransformer),
    ('lgb', LGBMClassifier())])

X_learn, X_val, Y_learn, Y_val = train_test_split(X, y, test_size = 0.2)

# 先拟合预处理组件，转换训练集和验证集
lgbPipe.named_steps['preprocess'].fit(X_learn)
X_learn_proc = lgbPipe.named_steps['preprocess'].transform(X_learn)
X_val_proc = lgbPipe.named_steps['preprocess'].transform(X_val)

# 传入处理后的验证集开启早停训练
lgbPipe.named_steps['lgb'].fit(
    X_learn_proc,
    Y_learn,
    eval_set=[(X_val_proc, Y_val)],
    early_stopping_rounds=50,  # 连续50轮无提升则停止训练，可按需调整
    verbose=10  # 可选，打印训练进度
)

方式二：保留Pipeline完整性（更符合工程规范） #

from sklearn.pipeline import Pipeline
from sklearn.compose import ColumnTransformer
from sklearn.preprocessing import OneHotEncoder
from sklearn.model_selection import train_test_split
import lightgbm as lgb

coltransformer = ColumnTransformer([
    ('cat', OneHotEncoder(sparse_output = False), cat_invar),
    ('num', 'passthrough', num_invar)])

lgbPipe = Pipeline([
    ('preprocess', coltransformer),
    ('lgb', LGBMClassifier())])

X_learn, X_val, Y_learn, Y_val = train_test_split(X, y, test_size = 0.2)

# 先拟合预处理组件
lgbPipe.named_steps['preprocess'].fit(X_learn)
# 转换验证集
X_val_proc = lgbPipe.named_steps['preprocess'].transform(X_val)

# 拟合整个Pipeline，传入处理后的验证集
lgbPipe.fit(
    X_learn,
    Y_learn,
    lgb__eval_set=[(X_val_proc, Y_val)],
    lgb__early_stopping_rounds=50,
    lgb__verbose=10
)

关键注意事项 #

• early_stopping_rounds是早停功能的核心参数，需根据任务需求调整阈值；
• 确保cat_invar和num_invar分别准确对应分类特征、数值特征的列名列表，避免预处理漏项或错项。

内容的提问来源于stack exchange，提问作者PingPong