XGBoost中的MAPE评估指标

Question

我正在尝试在xgboost中使用MAPE作为评估指标，但结果很奇怪：

def xgb_mape(preds, dtrain):
   labels = dtrain.get_label()
   return('mape', np.mean(np.abs((labels - preds) / (labels+1))))

xgp = {"colsample_bytree": 0.9, 
   "min_child_weight": 24, 
   "subsample": 0.9, 
   "eta": 0.05, 
   "objective": "reg:linear", 
   "seed": 70}

cv = xgb.cv(params = xgp, 
        dtrain = xgb.DMatrix(train_set[cols_to_use], label=train_set.y),
        folds = KFold(n = len(train_set), n_folds=4, random_state = 707, shuffle=True),
        feval = xgb_mape,
        early_stopping_rounds=10,
        num_boost_round=1000,
        verbose_eval=10,
        maximize=False
        )

它返回：

[0]     train-mape:0.780683+0.00241932  test-mape:0.779896+0.0024619
[10]    train-mape:0.84939+0.0196102    test-mape:0.858054+0.0184669
[20]    train-mape:1.0778+0.0313676     test-mape:1.10751+0.0293785
[30]    train-mape:1.26066+0.0343771    test-mape:1.30707+0.0323237
[40]    train-mape:1.37713+0.0347438    test-mape:1.43339+0.030565
[50]    train-mape:1.45653+0.042433     test-mape:1.52176+0.0383677
[60]    train-mape:1.52268+0.0386395    test-mape:1.5909+0.0353497
[70]    train-mape:1.5636+0.0383622     test-mape:1.63482+0.0301809
[80]    train-mape:1.59408+0.0378158    test-mape:1.66748+0.0315529
[90]    train-mape:1.61712+0.0403532    test-mape:1.69134+0.0325177
[100]   train-mape:1.63028+0.0389446    test-mape:1.70578+0.0316045
[110]   train-mape:1.63556+0.0375842    test-mape:1.71153+0.031564
[120]   train-mape:1.63509+0.0393198    test-mape:1.7117+0.0320471

当maximize=False时，训练和测试结果会提高，但是早期停止无法正常工作。错误在哪里？

更新：在xgb_mape中添加-1*后，问题得到解决。看起来自定义评估函数的maximize参数无法正常工作。

- Slavka

Answer 1

根据xgboost实现平均精度度量的示例，由于xgb优化器只进行最小化操作，如果您想要实现一个最大化的度量，就需要在其前面加上负号(-)。

def pr_auc_metric(y_predicted, y_true):
    return 'pr_auc', -skmetrics.average_precision_score(y_true.get_label(), y_predicted)

那么你的就是：

def xgb_mape(preds, dtrain):
   labels = dtrain.get_label()
   return('mape', -np.mean(np.abs((labels - preds) / (labels + 1))))

- data_steve