XGBoost最终模型参数和预测

如何解决XGBoost最终模型参数和预测

我在dtrain上训练XGBoost并在dtest上验证：

params = {'max_depth': 8,'min_child_weight': 5,'eta': 0.3,'subsample': 1,'colsample_bytree': 1,'objective': 'binary:logistic','disable_default_eval_metric': 1,'seed': 42,'tree_method': 'hist'}

def f1_eval(predt: np.ndarray,dtrain: xgboost.DMatrix):
    y = dtrain.get_label()
    predt_binary = np.where(predt > 0.5,1,0)
    return "F1_score",metrics.f1_score(y_true=y,y_pred=predt_binary)

xgboost_tuned_undersampled = xgboost.train(
    params,dtrain,num_boost_round=num_boost_round,evals=[(dtrain,'dtrain'),(dtest,'dtest')],feval=f1_eval,# customized F1
    early_stopping_rounds = 10,maximize=True)

它在迭代4处停止，F1为0.46471：

[0] dtrain-F1_score:0.34354 dtest-F1_score:0.25433
Multiple eval metrics have been passed: 'dtest-F1_score' will be used for early stopping.

Will train until dtest-F1_score hasn't improved in 10 rounds.
[1] dtrain-F1_score:0.59165 dtest-F1_score:0.45522
[2] dtrain-F1_score:0.63360 dtest-F1_score:0.42282
[3] dtrain-F1_score:0.66448 dtest-F1_score:0.44860
[4] dtrain-F1_score:0.67808 dtest-F1_score:0.46471
[5] dtrain-F1_score:0.69250 dtest-F1_score:0.45481
[6] dtrain-F1_score:0.69586 dtest-F1_score:0.45217
[7] dtrain-F1_score:0.70366 dtest-F1_score:0.45882
[8] dtrain-F1_score:0.71691 dtest-F1_score:0.45930
[9] dtrain-F1_score:0.72272 dtest-F1_score:0.45014
[10]    dtrain-F1_score:0.72622 dtest-F1_score:0.45014
[11]    dtrain-F1_score:0.73168 dtest-F1_score:0.44633
[12]    dtrain-F1_score:0.73809 dtest-F1_score:0.44944
[13]    dtrain-F1_score:0.74786 dtest-F1_score:0.44321
[14]    dtrain-F1_score:0.75239 dtest-F1_score:0.43836
Stopping. Best iteration:
[4] dtrain-F1_score:0.67808 dtest-F1_score:0.46471

为什么当我使用最终模型对SAME dtest进行预测并估计F1时，它会更低？

xgboost_tuned_undersampled_predt = np.where(xgboost_tuned_undersampled.predict(dtest) > 0.5,0)
print('F1:',round(metrics.f1_score(dtest.get_label(),xgboost_tuned_undersampled_predt),2))

F1: 0.37

最终的XGBoost模型具有哪些参数？

p.s。我知道我应该使用验证集而不是测试集，但我只是注意到了这一点，无法弄清为什么F1不相似。