数据挖掘 - xgboost 中用于回归的自定义目标函数 - 吾爱随笔录

我正在研究一个回归问题，我想修改 xgboost 库中的损失函数，以便我的预测永远不会小于实际值。我写了这段代码：

def custom_loss(preds, dtrain):
    labels = dtrain.get_label()
    df = preds - labels
    df = pd.DataFrame(df, columns=['val'])
    df['valg'] = df['val'].apply(lambda x: 10*abs(x) if x<0 else x)
    grad = df['valg'].as_matrix()
    return preds-labels, grad

这本质上意味着我想更多地惩罚那些小于我的实际值的预测。但是，这不起作用，我的预测也没有改善。谁能帮我弄清楚我哪里出错了？谢谢。

整个 Python 脚本：

params = {"booster" : "gbtree",
  "eta": 0.20,
  "max_depth": 4,
  "subsample": 0.75,
  "colsample_bytree": 0.65,
  "silent": 1,
  "eval_metric": "rmse",
  }
num_round = 400

def custom_loss(preds, dtrain):
    labels = dtrain.get_label()
    df = preds - labels
    df = pd.DataFrame(df, columns=['val'])
    df['valg'] = df['val'].apply(lambda x: 5*abs(x) if x<0 else x)
    grad = df['valg'].as_matrix()
    return preds-labels, grad


dtrain = xgb.DMatrix(X_train.drop('price_act', axis=1), 
label=X_train['price_act'])
dtest = xgb.DMatrix(X_test.drop('price_act',axis=1), 
label=X_test['price_act'])

watchlist = [(dtrain,'train'), (dtest,'eval')]

bst = xgb.train(params, dtrain, num_round, watchlist, custom_loss)