解决Sklearn Pipeline中XGBoost模型sample_weight参数未生效问题

问题重现 #

在使用Sklearn Pipeline封装StandardScaler和XGBDistribution时，通过model__sample_weight传递样本权重，出现XGBoost learner提示参数未被使用的警告：

示例代码：

from xgboost_distribution import XGBDistribution
from scipy.stats import nbinom
from sklearn.preprocessing import StandardScaler
from sklearn.pipeline import Pipeline

# 假设已完成数据集划分：X_train, X_test, y_train, y_test 及 sample_wgt

model = XGBDistribution(
    distribution="negative-binomial"
)

pipe = Pipeline(
  [
    ('scaler', StandardScaler()),
    ('model', model)
  ]
)

pipe.fit(X_train, y_train, model__sample_weight=sample_wgt)

执行后警告：

WARNING: /workspace/src/learner.cc:742: 
Parameters: { "sample_weight" } are not used.

解决方案 #

问题根源在于XGBDistribution的fit方法未正确接收Pipeline传递的sample_weight参数，导致底层XGBoost learner无法获取该参数。可以通过自定义包装类，明确处理样本权重的传递：

步骤1：创建XGBDistribution包装器 #

from sklearn.base import BaseEstimator, RegressorMixin
from xgboost_distribution import XGBDistribution

class XGBDistributionWrapper(BaseEstimator, RegressorMixin):
    def __init__(self, **kwargs):
        # 传递XGBDistribution的初始化参数
        self.model = XGBDistribution(**kwargs)
    
    def fit(self, X, y, sample_weight=None):
        # 明确接收sample_weight并传递给模型的fit方法
        if sample_weight is not None:
            self.model.fit(X, y, sample_weight=sample_weight)
        else:
            self.model.fit(X, y)
        return self
    
    def predict(self, X):
        # 复用原模型的predict方法
        return self.model.predict(X)
    
    def predict_distribution(self, X):
        # 复用原模型的分布预测方法
        return self.model.predict_distribution(X)

步骤2：在Pipeline中使用包装器 #

from sklearn.preprocessing import StandardScaler
from sklearn.pipeline import Pipeline

# 使用包装器替代原XGBDistribution
model = XGBDistributionWrapper(distribution="negative-binomial")

pipe = Pipeline(
  [
    ('scaler', StandardScaler()),
    ('model', model)
  ]
)

# 按原方式传递样本权重
pipe.fit(X_train, y_train, model__sample_weight=sample_wgt)

包装器会明确接收sample_weight参数并传递给底层的XGBDistribution模型，确保样本权重被正确使用，不会再出现参数未使用的警告。

内容的提问来源于stack exchange，提问作者a.powell