LightGBM算法 & XGBoost算法对比分析

2020-09-21  本文已影响0人  ShowMeCoding

官方文档
https://lightgbm.readthedocs.io/en/latest/index.html(微软开源)

参考文档:机器学习算法之LightGBM
https://www.biaodianfu.com/lightgbm.html

1 简介

LightGBM算法属于GBDT模型的进阶,和XGBoost一样采用损失函数的负梯度作为当前决策树的残差近似值,去拟合新的决策树。
因为GBDT在每一次迭代的时候,都需要遍历整个训练数据多次。如果把整个训练数据装进内存则会限制训练数据的大小;如果不装进内存,反复地读写训练数据又会消耗非常大的时间。因此,为了弥补普通的GBDT算法无法处理海量数据的问题,提出了LightGBM算法。

2 算法核心思想

2.1 XGBoost算法

XGBoost使用的是pre-sorted算法,能够更精确的找到数据分隔点

决策树的生长方式
XGBoost采用的是按层生长level(depth)-wise生长策略,能够同时分裂同一层的叶子,从而进行多线程优化,不容易过拟合;但不加区分的对待同一层的叶子,带来了很多没必要的开销。因为实际上很多叶子的分裂增益较低,没必要进行搜索和分裂。

2.2 LightGBM算法

LightGBM使用的是histogram算法(直方图算法),首先将连续的浮点数据转换为bin数据,具体过程是首先确定对于每一个特征需要多少的桶bin,然后均分,将属于该桶的样本数据更新为bin的值,最后用直方图表示。(看起来很高大上,其实就是直方图统计,最后我们将大规模的数据放在了直方图中)。
占用的内存更低,数据分隔的复杂度更低。其思想是将连续的浮点特征离散成k个离散值,并构造宽度为k的Histogram。然后遍历训练数据,统计每个离散值在直方图中的累计统计量。在进行特征选择时,只需要根据直方图的离散值,遍历寻找最优的分割点。
使用直方图算法的优点

决策树的生长方式
LightGBM采用leaf-wise生长策略,每次从当前所有叶子中找到分裂增益最大(一般也是数据量最大)的一个叶子,然后分裂,如此循环。因此同Level-wise相比,在分裂次数相同的情况下,Leaf-wise可以降低更多的误差,得到更好的精度。Leaf-wise的缺点是可能会长出比较深的决策树,产生过拟合。因此LightGBM在Leaf-wise之上增加了一个最大深度的限制,在保证高效率的同时防止过拟合。

3 优势——相比较XGBoost

4 在Scikit-Learn与XGBoost对比

https://scikit-learn.org/stable/modules/generated/sklearn.ensemble.HistGradientBoostingClassifier.html?highlight=lightgbm
histogram-based gradient boosting classification tree:基于直方图的梯度提升分类树

4.1 导入所需要的包和库

from xgboost import XGBRegressor
import lightgbm as lgb
import numpy as np
from sklearn.metrics import mean_squared_error
from sklearn.model_selection import GridSearchCV
from sklearn.datasets import load_iris
from sklearn.model_selection import train_test_split

4.2 加载鸢尾花数据

iris = load_iris()
data = iris.data
target = iris.target
X_train, X_test, y_train, y_test = train_test_split(data, target, test_size=0.2)

4.3 分别创建XGB和LGB模型进行训练

4.3.1 训练XGBoost

reg = XGBRegressor(
    n_estimators = 20,   # 迭代次数
    learning_rate = 0.1, # 学习率
    max_depth=5
)
reg.fit(X_train, y_train)
# 测试集预测
y_pred = reg.predict(X_test)
# 模型评估
print('The rmse of prediction is:', mean_squared_error(y_test, y_pred) ** 0.5)
The rmse of prediction is: 0.2579017768469883
# feature importances
print('Feature importances:', list(reg.feature_importances_))
Feature importances: [0.008026785, 0.025713025, 0.7764279, 0.1898323]

网格搜索

estimator = XGBRegressor()
param_grid = {
    'learning_rate':[0.2,0.5,0.8],
    'n_estimators': np.arange(0,100,20)
}
reg = GridSearchCV(estimator, param_grid)
clf = reg.fit(X_train, y_train)
y_pred = clf.predict(X_test)
mean_squared_error(y_test, y_pred)
0.05894615685616801

4.3.2 LightGBM训练

gbm = lgb.LGBMRegressor(objective='regression', num_leaves=31, learning_rate=0.1, n_estimators=20)
gbm.fit(X_train, y_train, eval_set=[(X_test, y_test)], eval_metric='l1', early_stopping_rounds=5)
[1] valid_0's l1: 0.558965  valid_0's l2: 0.507722
Training until validation scores don't improve for 5 rounds
[2] valid_0's l1: 0.508315  valid_0's l2: 0.422938
[3] valid_0's l1: 0.469458  valid_0's l2: 0.354078
[4] valid_0's l1: 0.428637  valid_0's l2: 0.297129
[5] valid_0's l1: 0.397107  valid_0's l2: 0.251408
[6] valid_0's l1: 0.363512  valid_0's l2: 0.210417
[7] valid_0's l1: 0.334591  valid_0's l2: 0.179052
[8] valid_0's l1: 0.311293  valid_0's l2: 0.151726
[9] valid_0's l1: 0.288382  valid_0's l2: 0.130775
[10]    valid_0's l1: 0.269027  valid_0's l2: 0.111078
[11]    valid_0's l1: 0.250241  valid_0's l2: 0.097117
[12]    valid_0's l1: 0.231755  valid_0's l2: 0.0847149
[13]    valid_0's l1: 0.219546  valid_0's l2: 0.0744859
[14]    valid_0's l1: 0.206336  valid_0's l2: 0.0668118
[15]    valid_0's l1: 0.193833  valid_0's l2: 0.0604602
[16]    valid_0's l1: 0.183163  valid_0's l2: 0.0551445
[17]    valid_0's l1: 0.173473  valid_0's l2: 0.0503368
[18]    valid_0's l1: 0.164622  valid_0's l2: 0.0465006
[19]    valid_0's l1: 0.156789  valid_0's l2: 0.0433829
[20]    valid_0's l1: 0.153458  valid_0's l2: 0.0419514
Did not meet early stopping. Best iteration is:
[20]    valid_0's l1: 0.153458  valid_0's l2: 0.0419514

LGBMRegressor(n_estimators=20, objective='regression')
# 测试集预测
y_pred = gbm.predict(X_test, num_iteration=gbm.best_iteration_)
# 模型评估
print('The rmse of prediction is:', mean_squared_error(y_test, y_pred) ** 0.5)
The rmse of prediction is: 0.20482033985501885
# feature importances
print('Feature importances:', list(gbm.feature_importances_))
Feature importances: [15, 9, 27, 21]

网格搜索,参数优化

estimator = lgb.LGBMRegressor(num_leaves=31)
param_grid = {
    'learning_rate':[0.2,0.5,0.8],
    'n_estimators': np.arange(0,100,20)
}
gbm = GridSearchCV(estimator, param_grid)
clf = gbm.fit(X_train, y_train)
y_pred = clf.predict(X_test)
mean_squared_error(y_test, y_pred)
 0.03318734265620019
​print('Best parameters found by grid search are:', reg.best_params_)
print('Best parameters found by grid search are:', gbm.best_params_)
Best parameters found by grid search are: {'learning_rate': 0.2, 'n_estimators': 20}
Best parameters found by grid search are: {'learning_rate': 0.2, 'n_estimators': 20}

4.4 总结

model 默认参数 网格搜索 网格搜索参数
XGBoost 0.2579017768469883 0.05894615685616801 'learning_rate': 0.2, 'n_estimators': 20
LightGBM 0.20482033985501885 0.03318734265620019 'learning_rate': 0.2, 'n_estimators': 20

在评价标准为RMSE的情况下,从预测的精度来看,LightGBM明显优于XGBoost。

上一篇下一篇

猜你喜欢

热点阅读