1

tree_method = 'exact'这真的意味着使用精确的xgboost贪心算法进行拆分查找吗？

我问这个问题是因为xgboost运行速度不合理。这是我用于运行测试的脚本

from xgboost import XGBRegressor as rr
import numpy as np
from sklearn.model_selection import train_test_split
import pickle 
import sys
from time import time

t1 = time()
data = sys.argv[1]
with open(data, 'rb') as source:
    data = pickle.load(source)
np.random.shuffle(data)
x = [item[0] for item in data]
y = [item[1] for item in data]
x_train, x_val, y_train, y_val = train_test_split(x, y, test_size = 0.2, random_state = 100)
eval_set = [(x_train, y_train), (x_val, y_val)]

# Exact model
model_exact = rr(max_depth = 5, 
                 n_estimators = 1,
                 slient = False, 
                 min_child_weight = 0, 
                 tree_method = 'exact')
model_exact.fit(x_train, 
                y_train, 
                eval_set=eval_set,
                eval_metric="mae",
                early_stopping_rounds=30)
t2 = time()
print(f"Time used: {t2 - t1}")

用于测试的腌制数据已上传至此处。

每个实例都有96特征，并且总共有11450实例。

xgboost0.9804270267486572s通过在单核 ( 1.3 GHz Intel Core i5)上运行找到第一个拆分。如果xgboost实际上对所有可能的拆分执行了贪婪搜索，则意味着只xgboost评估11450 x 96 = 1099200拆分0.9804270267486572s！

真的有xgboost这么厉害吗？或者，我误解了tree_method = exact吗？

4

0 回答 0