比賽得分公式如下：

其中，P為Precision , R為 Recall。

GBDT訓練基於驗證集評價，此時會調用評價函數，XGBoost的best_iteration和best_score均是基於評價函數得出。

評價函數：

input: preds和dvalid，即為驗證集和驗證集上的預測值，

return string 類型的名稱 和一個flaot類型的fevalerror值表示評價值的大小，其是以error的形式定義，即當此值越大是認為模型效果越差。

1 from sklearn.metrics import confusion_matrix
2 def customedscore(preds, dtrain):
 
3     label = dtrain.get_label()
4     pred = [int(i>=0.5) for i in preds]
5     confusion_matrixs = confusion_matrix(label, pred)
6     recall =float(confusion_matrixs[0][0]) / float(confusion_matrixs[0][1]+confusion_matrixs[0][0])
7     precision = float(confusion_matrixs[0][0]) / float(confusion_matrixs[1][0]+confusion_matrixs[0][0])
 
8     F = 5*precision* recall/(2*precision+3*recall)*100
9     return ‘FSCORE‘,float(F)

應用：

訓練時要傳入參數：feval = customedscore,

 1    params = { ‘silent‘: 1,  ‘objective‘: ‘binary:logistic‘ , ‘gamma‘:0.1,
 2         ‘min_child_weight‘:5,
 3         ‘max_depth‘:5,
 4         ‘lambda‘:10,
 5         ‘subsample
 
‘:0.7,
 6         ‘colsample_bytree‘:0.7,
 7         ‘colsample_bylevel‘:0.7,
 8         ‘eta‘: 0.01,
 9         ‘tree_method‘:‘exact‘}
10     model = xgb.train(params, trainsetall, num_round,verbose_eval=10, feval = customedscore,maximize=False)

自定義 目標函數，這個我沒有具體使用

1 # user define objective function, given prediction, return gradient and second order gradient
2 # this is log likelihood loss
3 def logregobj(preds, dtrain):
4     labels = dtrain.get_label()
5     preds = 1.0 / (1.0 + np.exp(-preds))
6     grad = preds - labels
7     hess = preds * (1.0-preds)
8     return grad, hess

# training with customized objective, we can also do step by step training
# simply look at xgboost.py‘s implementation of train
bst = xgb.train(param, dtrain, num_round, watchlist, logregobj, evalerror)

參考：

https://github.com/dmlc/xgboost/blob/master/demo/guide-python/custom_objective.py

http://blog.csdn.net/lujiandong1/article/details/52791117

xgboost 自定義評價函數（metric）與目標函數