XGBRegressor使用sklearn的mean_squared_error时样本数不一致问题

问题背景 #

用户尝试训练多输出XGBRegressor，代码如下：

import xgboost as xgb
from sklearn.metrics import mean_squared_error

def xgboost(): 
    model = xgb.XGBRegressor(n_estimators=200,
                             max_depth=4,
                             subsample=1,
                             min_child_weight=1,
                             objective='reg:squarederror',
                             tree_method='hist',
                             eval_metric=mean_squared_error, # 使用sklearn的MSE
                             early_stopping_rounds=50)
    return model

num_amostras = x_train.shape[0]
val_size = 0.2
num_amostras_train = int(num_amostras * (1-val_size))
x_train_xgb = x_train[:num_amostras_train]
y_train_xgb = y_train[:num_amostras_train]
x_val_xgb = x_train[num_amostras_train:]
y_val_xgb = y_train[num_amostras_train:]
model_xgb = xgboost()
model_xgb.fit(x_train_xgb, y_train_xgb, eval_set=[(x_train_xgb, y_train_xgb), (x_val_xgb, y_val_xgb)])
resultados = model_xgb.evals_result()

数据形状：

• x_train: (1458, 55)
• x_train_xgb: (1166, 55)，y_train_xgb: (1166, 24)
• x_val_xgb: (292, 55)，y_val_xgb: (292, 24)

运行时报错：

Traceback (most recent call last):
  File ~\PeDFurnas\lib\site-packages\spyder_kernels\py3compat.py:356 in compat_exec
    exec(code, globals, locals)
  File c:\users\ldsp_\sipredvs\scripts\treinamento_demanda.py:201
    model_xgb.fit(x_train_xgb, y_train_xgb, eval_set=[(x_train_xgb, y_train_xgb),(x_val_xgb, y_val_xgb)])
  File ~\PeDFurnas\lib\site-packages\xgboost\core.py:729 in inner_f
    return func(**kwargs)
  File ~\PeDFurnas\lib\site-packages\xgboost\sklearn.py:1086 in fit
    self._Booster = train(
  File ~\PeDFurnas\lib\site-packages\xgboost\core.py:729 in inner_f
    return func(**kwargs)
  File ~\PeDFurnas\lib\site-packages\xgboost\training.py:182 in train
    if cb_container.after_iteration(bst, i, dtrain, evals):
  File ~\PeDFurnas\lib\site-packages\xgboost\callback.py:238 in after_iteration
    score: str = model.eval_set(evals, epoch, self.metric, self._output_margin)
  File ~\PeDFurnas\lib\site-packages\xgboost\core.py:2138 in eval_set
    feval_ret = feval(
  File ~\PeDFurnas\lib\site-packages\xgboost\sklearn.py:139 in inner
    return func.__name__, func(y_true, y_score)
  File ~\PeDFurnas\lib\site-packages\sklearn\metrics\_regression.py:442 in mean_squared_error
    y_type, y_true, y_pred, multioutput = _check_reg_targets(
  File ~\PeDFurnas\lib\site-packages\sklearn\metrics\_regression.py:100 in _check_reg_targets
    check_consistent_length(y_true, y_pred)
  File ~\PeDFurnas\lib\site-packages\sklearn\utils\validation.py:397 in check_consistent_length
    raise ValueError(
ValueError: Found input variables with inconsistent numbers of samples: [27984, 1166]

其中27984=1166*24，是y_train_xgb形状的乘积；改用XGBoost默认的'rmse'指标可正常运行。

问题原因 #

1. 形状不匹配：在多输出回归场景下，XGBoost返回的预测值y_score是扁平化的一维数组（形状为(n_samples * n_outputs,)），而sklearn的mean_squared_error默认接收的y_true是二维数组（(n_samples, n_outputs)），两者样本数检查时不匹配，导致报错。
2. 内置指标适配性：XGBoost内置的'rmse'/'mse'指标已经针对多输出场景做了适配，会自动处理形状对齐，因此可以正常运行。

修复方法 #

方法1：使用XGBoost内置的MSE指标 #

直接将eval_metric指定为'mse'，效果和sklearn的mean_squared_error一致，代码修改如下：

def xgboost(): 
    model = xgb.XGBRegressor(n_estimators=200,
                             max_depth=4,
                             subsample=1,
                             min_child_weight=1,
                             objective='reg:squarederror',
                             tree_method='hist',
                             eval_metric='mse',  # 使用XGBoost内置MSE
                             early_stopping_rounds=50)
    return model

方法2：自定义适配多输出的MSE函数 #

如果必须使用sklearn的mean_squared_error，可以自定义函数手动还原预测值的形状：

from sklearn.metrics import mean_squared_error
import numpy as np

def multioutput_mse(y_true, y_pred):
    # 将扁平化的预测值还原为和真实值一致的形状
    y_pred_reshaped = y_pred.reshape(y_true.shape)
    # 返回指标名称和计算结果
    return 'mse', mean_squared_error(y_true, y_pred_reshaped)

# 修改模型定义中的eval_metric
def xgboost(): 
    model = xgb.XGBRegressor(n_estimators=200,
                             max_depth=4,
                             subsample=1,
                             min_child_weight=1,
                             objective='reg:squarederror',
                             tree_method='hist',
                             eval_metric=multioutput_mse,  # 使用自定义函数
                             early_stopping_rounds=50)
    return model

内容的提问来源于stack exchange，提问作者Murilo