九、boosting的原理和案例分析
来源
Datewhle24期__集成学习(中) :
https://github.com/datawhalechina/team-learning-data-mining/tree/master/EnsembleLearning
作者:李祖贤、薛传雨、赵可、杨毅远、陈琰钰
论坛地址:
http://datawhale.club/t/topic/1574
9.1 Boosting
9.1.1 Boosting概念原理
-
Bagging往往通过降低方差来减小预测误差
-
Boosting通过降低偏差来减小预测误差
-
常见的模型有: Adaboost、Gradient Boosting(GBT/GBDT/GBRT)、XGBoost、LightGBM。
-
Boosting基本思路:
- Valiant和 Kearns提出了"强学习"和"弱学习"的概念:
- 在概率近似正确PAC学习的框架下:
- 弱学习 :识别错误率小于1/2(即准确率仅比随机猜测略高的学习算法)
- 强学习:识别准确率很高并能在多项式时间内完成的学习算法
9.1.2 Adaboost基本原理
- 对于Boosting方法来说,有两个问题需要给出答案:
- 1.是每一轮学习应该如何改变数据的概率分布,
- 2.如何将各个弱分类器组合起来。
- Adaboost处理方式:
- 1.提高那些被前一轮分类器错误分类的样本的权重,而降低那些被正确分类的样本的权重。
- 2.加大分类错误率低的弱分类器的权重,因为这些分类器能更好地完成分类任务,而减小分类错误率较大的弱分类器的权重,使其在表决中起较小的作用。
假设给定一个二分类的训练数据集:
T
=
{
(
x
1
,
y
1
)
,
(
x
2
,
y
2
)
,
⋯
,
(
x
N
,
y
N
)
}
T=\left\{\left(x_{1}, y_{1}\right),\left(x_{2}, y_{2}\right), \cdots,\left(x_{N}, y_{N}\right)\right\}
T={(x1,y1),(x2,y2),⋯,(xN,yN)},其中每个样本点由特征与类别组成。特征
x
i
∈
X
⊆
R
n
x_{i} \in \mathcal{X} \subseteq \mathbf{R}^{n}
xi∈X⊆Rn,类别
y
i
∈
Y
=
{
−
1
,
+
1
}
y_{i} \in \mathcal{Y}=\{-1,+1\}
yi∈Y={−1,+1},
X
\mathcal{X}
X是特征空间,
Y
\mathcal{Y}
Y是类别集合,
输出最终分类器
G
(
x
)
G(x)
G(x)。Adaboost算法如下:
(1) 初始化训练数据的分布:
D
1
=
(
w
11
,
⋯
,
w
1
i
,
⋯
,
w
1
N
)
,
w
1
i
=
1
N
,
i
=
1
,
2
,
⋯
,
N
D_{1}=\left(w_{11}, \cdots, w_{1 i}, \cdots, w_{1 N}\right), \quad w_{1 i}=\frac{1}{N}, \quad i=1,2, \cdots, N
D1=(w11,⋯,w1i,⋯,w1N),w1i=N1,i=1,2,⋯,N
(2) 对于m=1,2,…,M
- 使用具有权值分布 D m D_m Dm的训练数据集进行学习,得到基本分类器: G m ( x ) : X → { − 1 , + 1 } G_{m}(x): \mathcal{X} \rightarrow\{-1,+1\} Gm(x):X→{−1,+1}
- 计算 G m ( x ) G_m(x) Gm(x)在训练集上的分类误差率 e m = ∑ i = 1 N P ( G m ( x i ) ≠ y i ) = ∑ i = 1 N w m i I ( G m ( x i ) ≠ y i ) e_{m}=\sum_{i=1}^{N} P\left(G_{m}\left(x_{i}\right) \neq y_{i}\right)=\sum_{i=1}^{N} w_{m i} I\left(G_{m}\left(x_{i}\right) \neq y_{i}\right) em=∑i=1NP(Gm(xi)=yi)=∑i=1NwmiI(Gm(xi)=yi)
- 计算 G m ( x ) G_m(x) Gm(x)的系数 α m = 1 2 log 1 − e m e m \alpha_{m}=\frac{1}{2} \log \frac{1-e_{m}}{e_{m}} αm=21logem1−em,这里的log是自然对数ln
- 更新训练数据集的权重分布
D m + 1 = ( w m + 1 , 1 , ⋯ , w m + 1 , i , ⋯ , w m + 1 , N ) w m + 1 , i = w m i Z m exp ( − α m y i G m ( x i ) ) , i = 1 , 2 , ⋯ , N \begin{array}{c} D_{m+1}=\left(w_{m+1,1}, \cdots, w_{m+1, i}, \cdots, w_{m+1, N}\right) \\ w_{m+1, i}=\frac{w_{m i}}{Z_{m}} \exp \left(-\alpha_{m} y_{i} G_{m}\left(x_{i}\right)\right), \quad i=1,2, \cdots, N \end{array} Dm+1=(wm+1,1,⋯,wm+1,i,⋯,wm+1,N)wm+1,i=Zmwmiexp(−αmyiGm(xi)),i=1,2,⋯,N
这里的 Z m Z_m Zm是规范化因子,使得 D m + 1 D_{m+1} Dm+1称为概率分布, Z m = ∑ i = 1 N w m i exp ( − α m y i G m ( x i ) ) Z_{m}=\sum_{i=1}^{N} w_{m i} \exp \left(-\alpha_{m} y_{i} G_{m}\left(x_{i}\right)\right) Zm=∑i=1Nwmiexp(−αmyiGm(xi))
(3) 构建基本分类器的线性组合 f ( x ) = ∑ m = 1 M α m G m ( x ) f(x)=\sum_{m=1}^{M} \alpha_{m} G_{m}(x) f(x)=∑m=1MαmGm(x),得到最终的分类器
G ( x ) = sign ( f ( x ) ) = sign ( ∑ m = 1 M α m G m ( x ) ) \begin{aligned} G(x) &=\operatorname{sign}(f(x)) \\ &=\operatorname{sign}\left(\sum_{m=1}^{M} \alpha_{m} G_{m}(x)\right) \end{aligned} G(x)=sign(f(x))=sign(m=1∑MαmGm(x))
下面对Adaboost算法做如下说明:
对于步骤(1),假设训练数据的权值分布是均匀分布,是为了使得第一次没有先验信息的条件下每个样本在基本分类器的学习中作用一样。
对于步骤(2),每一次迭代产生的基本分类器
G
m
(
x
)
G_m(x)
Gm(x)在加权训练数据集上的分类错误率
e
m
=
∑
i
=
1
N
P
(
G
m
(
x
i
)
≠
y
i
)
=
∑
G
m
(
x
i
)
≠
y
i
w
m
i
\begin{aligned}e_{m} &=\sum_{i=1}^{N} P\left(G_{m}\left(x_{i}\right) \neq y_{i}\right) =\sum_{G_{m}\left(x_{i}\right) \neq y_{i}} w_{m i}\end{aligned}
em=i=1∑NP(Gm(xi)=yi)=Gm(xi)=yi∑wmi代表了在
G
m
(
x
)
G_m(x)
Gm(x)中分类错误的样本权重和,这点直接说明了权重分布
D
m
D_m
Dm与
G
m
(
x
)
G_m(x)
Gm(x)的分类错误率
e
m
e_m
em有直接关系。同时,在步骤(2)中,计算基本分类器
G
m
(
x
)
G_m(x)
Gm(x)的系数
α
m
\alpha_m
αm,
α
m
=
1
2
log
1
−
e
m
e
m
\alpha_{m}=\frac{1}{2} \log \frac{1-e_{m}}{e_{m}}
αm=21logem1−em,它表示了
G
m
(
x
)
G_m(x)
Gm(x)在最终分类器的重要性程度,
α
m
\alpha_m
αm的取值由基本分类器
G
m
(
x
)
G_m(x)
Gm(x)的分类错误率有直接关系,当
e
m
⩽
1
2
e_{m} \leqslant \frac{1}{2}
em⩽21时,
α
m
⩾
0
\alpha_{m} \geqslant 0
αm⩾0,并且
α
m
\alpha_m
αm随着
e
m
e_m
em的减少而增大,因此分类错误率越小的基本分类器在最终分类器的作用越大!
**最重要的,对于步骤(2)中的样本权重的更新: **
w
m
+
1
,
i
=
{
w
m
i
Z
m
e
−
α
m
,
G
m
(
x
i
)
=
y
i
w
m
i
Z
m
e
α
m
,
G
m
(
x
i
)
≠
y
i
w_{m+1, i}=\left\{\begin{array}{ll} \frac{w_{m i}}{Z_{m}} \mathrm{e}^{-\alpha_{m}}, & G_{m}\left(x_{i}\right)=y_{i} \\ \frac{w_{m i}}{Z_{m}} \mathrm{e}^{\alpha_{m}}, & G_{m}\left(x_{i}\right) \neq y_{i} \end{array}\right.
wm+1,i={Zmwmie−αm,Zmwmieαm,Gm(xi)=yiGm(xi)=yi
因此,从上式可以看到:被基本分类器
G
m
(
x
)
G_m(x)
Gm(x)错误分类的样本的权重扩大,被正确分类的样本权重减少,二者相比相差
e
2
α
m
=
1
−
e
m
e
m
\mathrm{e}^{2 \alpha_{m}}=\frac{1-e_{m}}{e_{m}}
e2αm=em1−em倍。
对于步骤(3),线性组合
f
(
x
)
f(x)
f(x)实现了将M个基本分类器的加权表决,系数
α
m
\alpha_m
αm标志了基本分类器
G
m
(
x
)
G_m(x)
Gm(x)的重要性,值得注意的是:所有的
α
m
\alpha_m
αm之和不为1。
f
(
x
)
f(x)
f(x)的符号决定了样本x属于哪一类。
- 一个例子:
训练数据如下表,假设基本分类器的形式是一个分割 x < v x<v x<v或 x > v x>v x>v表示,阈值v由该基本分类器在训练数据集上分类错误率 e m e_m em最低确定。
序号 1 2 3 4 5 6 7 8 9 10 x 0 1 2 3 4 5 6 7 8 9 y 1 1 1 − 1 − 1 − 1 1 1 1 − 1 \begin{array}{ccccccccccc} \hline \text { 序号 } & 1 & 2 & 3 & 4 & 5 & 6 & 7 & 8 & 9 & 10 \\ \hline x & 0 & 1 & 2 & 3 & 4 & 5 & 6 & 7 & 8 & 9 \\ y & 1 & 1 & 1 & -1 & -1 & -1 & 1 & 1 & 1 & -1 \\ \hline \end{array} 序号 xy10121132143−154−165−1761871981109−1
解:
初始化样本权值分布
D 1 = ( w 11 , w 12 , ⋯ , w 110 ) w 1 i = 0.1 , i = 1 , 2 , ⋯ , 10 \begin{aligned} D_{1} &=\left(w_{11}, w_{12}, \cdots, w_{110}\right) \\ w_{1 i} &=0.1, \quad i=1,2, \cdots, 10 \end{aligned} D1w1i=(w11,w12,⋯,w110)=0.1,i=1,2,⋯,10
对m=1:- 在权值分布
D
1
D_1
D1的训练数据集上,遍历每个结点并计算分类误差率
e
m
e_m
em,阈值取v=2.5时分类误差率最低,那么基本分类器为:
G 1 ( x ) = { 1 , x < 2.5 − 1 , x > 2.5 G_{1}(x)=\left\{\begin{array}{ll} 1, & x<2.5 \\ -1, & x>2.5 \end{array}\right. G1(x)={1,−1,x<2.5x>2.5 - G 1 ( x ) G_1(x) G1(x)在训练数据集上的误差率为 e 1 = P ( G 1 ( x i ) ≠ y i ) = 0.3 e_{1}=P\left(G_{1}\left(x_{i}\right) \neq y_{i}\right)=0.3 e1=P(G1(xi)=yi)=0.3。
- 计算 G 1 ( x ) G_1(x) G1(x)的系数: α 1 = 1 2 log 1 − e 1 e 1 = 0.4236 \alpha_{1}=\frac{1}{2} \log \frac{1-e_{1}}{e_{1}}=0.4236 α1=21loge11−e1=0.4236
- 更新训练数据的权值分布:
D 2 = ( w 21 , ⋯ , w 2 i , ⋯ , w 210 ) w 2 i = w 1 i Z 1 exp ( − α 1 y i G 1 ( x i ) ) , i = 1 , 2 , ⋯ , 10 D 2 = ( 0.07143 , 0.07143 , 0.07143 , 0.07143 , 0.07143 , 0.07143 , 0.16667 , 0.16667 , 0.16667 , 0.07143 ) f 1 ( x ) = 0.4236 G 1 ( x ) \begin{aligned} D_{2}=&\left(w_{21}, \cdots, w_{2 i}, \cdots, w_{210}\right) \\ w_{2 i}=& \frac{w_{1 i}}{Z_{1}} \exp \left(-\alpha_{1} y_{i} G_{1}\left(x_{i}\right)\right), \quad i=1,2, \cdots, 10 \\ D_{2}=&(0.07143,0.07143,0.07143,0.07143,0.07143,0.07143,\\ &0.16667,0.16667,0.16667,0.07143) \\ f_{1}(x) &=0.4236 G_{1}(x) \end{aligned} D2=w2i=D2=f1(x)(w21,⋯,w2i,⋯,w210)Z1w1iexp(−α1yiG1(xi)),i=1,2,⋯,10(0.07143,0.07143,0.07143,0.07143,0.07143,0.07143,0.16667,0.16667,0.16667,0.07143)=0.4236G1(x)
- 在权值分布
D
1
D_1
D1的训练数据集上,遍历每个结点并计算分类误差率
e
m
e_m
em,阈值取v=2.5时分类误差率最低,那么基本分类器为:
对于m=2:
- 在权值分布
D
2
D_2
D2的训练数据集上,遍历每个结点并计算分类误差率
e
m
e_m
em,阈值取v=8.5时分类误差率最低,那么基本分类器为:
G 2 ( x ) = { 1 , x < 8.5 − 1 , x > 8.5 G_{2}(x)=\left\{\begin{array}{ll} 1, & x<8.5 \\ -1, & x>8.5 \end{array}\right. G2(x)={1,−1,x<8.5x>8.5 - G 2 ( x ) G_2(x) G2(x)在训练数据集上的误差率为 e 2 = 0.2143 e_2 = 0.2143 e2=0.2143
- 计算 G 2 ( x ) G_2(x) G2(x)的系数: α 2 = 0.6496 \alpha_2 = 0.6496 α2=0.6496
- 更新训练数据的权值分布:
D 3 = ( 0.0455 , 0.0455 , 0.0455 , 0.1667 , 0.1667 , 0.1667 0.1060 , 0.1060 , 0.1060 , 0.0455 ) f 2 ( x ) = 0.4236 G 1 ( x ) + 0.6496 G 2 ( x ) \begin{aligned} D_{3}=&(0.0455,0.0455,0.0455,0.1667,0.1667,0.1667\\ &0.1060,0.1060,0.1060,0.0455) \\ f_{2}(x) &=0.4236 G_{1}(x)+0.6496 G_{2}(x) \end{aligned} D3=f2(x)(0.0455,0.0455,0.0455,0.1667,0.1667,0.16670.1060,0.1060,0.1060,0.0455)=0.4236G1(x)+0.6496G2(x)
对m=3:
- 在权值分布
D
3
D_3
D3的训练数据集上,遍历每个结点并计算分类误差率
e
m
e_m
em,阈值取v=5.5时分类误差率最低,那么基本分类器为:
G 3 ( x ) = { 1 , x > 5.5 − 1 , x < 5.5 G_{3}(x)=\left\{\begin{array}{ll} 1, & x>5.5 \\ -1, & x<5.5 \end{array}\right. G3(x)={1,−1,x>5.5x<5.5 - G 3 ( x ) G_3(x) G3(x)在训练数据集上的误差率为 e 3 = 0.1820 e_3 = 0.1820 e3=0.1820
- 计算 G 3 ( x ) G_3(x) G3(x)的系数: α 3 = 0.7514 \alpha_3 = 0.7514 α3=0.7514
- 更新训练数据的权值分布:
D 4 = ( 0.125 , 0.125 , 0.125 , 0.102 , 0.102 , 0.102 , 0.065 , 0.065 , 0.065 , 0.125 ) D_{4}=(0.125,0.125,0.125,0.102,0.102,0.102,0.065,0.065,0.065,0.125) D4=(0.125,0.125,0.125,0.102,0.102,0.102,0.065,0.065,0.065,0.125)
于是得到:
f
3
(
x
)
=
0.4236
G
1
(
x
)
+
0.6496
G
2
(
x
)
+
0.7514
G
3
(
x
)
f_{3}(x)=0.4236 G_{1}(x)+0.6496 G_{2}(x)+0.7514 G_{3}(x)
f3(x)=0.4236G1(x)+0.6496G2(x)+0.7514G3(x),分类器
sign
[
f
3
(
x
)
]
\operatorname{sign}\left[f_{3}(x)\right]
sign[f3(x)]在训练数据集上的误分类点的个数为0。
于是得到最终分类器为:
G
(
x
)
=
sign
[
f
3
(
x
)
]
=
sign
[
0.4236
G
1
(
x
)
+
0.6496
G
2
(
x
)
+
0.7514
G
3
(
x
)
]
G(x)=\operatorname{sign}\left[f_{3}(x)\right]=\operatorname{sign}\left[0.4236 G_{1}(x)+0.6496 G_{2}(x)+0.7514 G_{3}(x)\right]
G(x)=sign[f3(x)]=sign[0.4236G1(x)+0.6496G2(x)+0.7514G3(x)]
9.1 Boosting代码例子
- 数据集为UCI中葡萄酒数据,
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
plt.style.use("ggplot")
%matplotlib inline
import seaborn as sns
wine = pd.read_csv("https://archive.ics.uci.edu/ml/machine-learning-databases/wine/wine.data",header=None)
wine.columns = ['Class label', 'Alcohol', 'Malic acid', 'Ash', 'Alcalinity of ash','Magnesium', 'Total phenols','Flavanoids', 'Nonflavanoid phenols',
'Proanthocyanins','Color intensity', 'Hue','OD280/OD315 of diluted wines','Proline']
wine = wine[wine['Class label'] != 1]
y = wine['Class label'].values
X = wine[['Alcohol','OD280/OD315 of diluted wines']].values
# 将分类标签变成二进制编码:
from sklearn.preprocessing import LabelEncoder
le = LabelEncoder() #标签编码 类似one-hot
y = le.fit_transform(y)
# 按8:2分割训练集和测试集
from sklearn.model_selection import train_test_split
X_train,X_test,y_train,y_test = train_test_split(X,y,test_size=0.2,
random_state=1,stratify=y)
# stratify参数代表了按照y的类别等比例抽样
# 使用单一决策树建模
from sklearn.tree import DecisionTreeClassifier #决策树
tree = DecisionTreeClassifier(criterion='entropy',random_state=1,max_depth=1)
from sklearn.metrics import accuracy_score #准确率ACC
tree = tree.fit(X_train,y_train)
y_train_pred = tree.predict(X_train)
y_test_pred = tree.predict(X_test)
tree_train = accuracy_score(y_train,y_train_pred)
tree_test = accuracy_score(y_test,y_test_pred)
print('Decision tree train/test accuracies %.3f/%.3f' % (tree_train,tree_test))
Decision tree train/test accuracies 0.916/0.875
AdaBoostClassifier相关参数:
base_estimator
:基本分类器,默认为DecisionTreeClassifier(max_depth=1)
n_estimators
:终止迭代的次数learning_rate
:学习率algorithm
:训练的相关算法,{‘SAMME’,‘SAMME.R’},默认=SAMME.R
random_state
:随机种子
from sklearn.ensemble import AdaBoostClassifier
ada = AdaBoostClassifier(base_estimator=tree,n_estimators=500,learning_rate=0.1,random_state=1)
ada = ada.fit(X_train,y_train)
y_train_pred = ada.predict(X_train)
y_test_pred = ada.predict(X_test)
ada_train = accuracy_score(y_train,y_train_pred)
ada_test = accuracy_score(y_test,y_test_pred)
print('Adaboost train/test accuracies %.3f/%.3f' % (ada_train,ada_test))
Adaboost train/test accuracies 1.000/0.917
- 决策边界图
# 画出单层决策树与Adaboost的决策边界:
x_min = X_train[:, 0].min() - 1
x_max = X_train[:, 0].max() + 1
y_min = X_train[:, 1].min() - 1
y_max = X_train[:, 1].max() + 1
xx, yy = np.meshgrid(np.arange(x_min, x_max, 0.1),np.arange(y_min, y_max, 0.1))
f, axarr = plt.subplots(nrows=1, ncols=2,sharex='col',sharey='row',figsize=(12, 6))
for idx, clf, tt in zip([0, 1],[tree, ada],['Decision tree', 'Adaboost']):
clf.fit(X_train, y_train)
Z = clf.predict(np.c_[xx.ravel(), yy.ravel()])
Z = Z.reshape(xx.shape)
axarr[idx].contourf(xx, yy, Z, alpha=0.3)
axarr[idx].scatter(X_train[y_train==0, 0],X_train[y_train==0, 1],c='blue', marker='^')
axarr[idx].scatter(X_train[y_train==1, 0],X_train[y_train==1, 1],c='red', marker='o')
axarr[idx].set_title(tt)
axarr[0].set_ylabel('Alcohol', fontsize=12)
plt.tight_layout()
plt.text(0, -0.2,s='OD280/OD315 of diluted wines',ha='center',va='center',fontsize=12,transform=axarr[1].transAxes)
plt.show()
- 显然Adaboost边界更复杂, 即Adaboost通过增加模型复杂度而降低偏差,但是过程中引入了方差,易出现过拟合,因此在训练集和测试集之间的性能存在较大的差距.
- 值的注意的是:与单个分类器相比,Adaboost等Boosting模型增加了计算的复杂度,在实践中需要仔细思考是否愿意为预测性能的相对改善而增加计算成本,而且Boosting方式无法做到现在流行的并行计算的方式进行训练,因为每一步迭代都要基于上一部的基本分类器。
参考资料
- https://blog.csdn.net/starter_____/article/details/79328749机器学习 —— Boosting算法
- <>