集成学习
import numpy as np
import matplotlib.pyplot as plt
from sklearn import datasets
X, y = datasets.make_moons(n_samples=500, noise=0.3, random_state=42)
plt.scatter(X[y==0,0], X[y==0,1])
plt.scatter(X[y==1,0], X[y==1,1])
plt.show()
from sklearn.model_selection import train_test_split
X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=42)
from sklearn.linear_model import LogisticRegression
log_clf = LogisticRegression()
log_clf.fit(X_train, y_train)
log_clf.score(X_test, y_test)# ->0.86399999999
from sklearn.svm import SVC
svm_clf = SVC()
svm_clf.fit(X_train, y_train)
svm_clf.score(X_test, y_test)#->0.8880000000001
from sklearn.tree import DecisionTreeClassifier
dt_clf = DecisionTreeClassifier(random_state=666)
dt_clf.fit(X_train, y_train)
dt_clf.score(X_test, y_test)#->0.86399999999999999
y_predict1 = log_clf.predict(X_test)
y_predict2 = svm_clf.predict(X_test)
y_predict3 = dt_clf.predict(X_test)
y_predict = np.array((y_predict1 + y_predict2 + y_predict3) >= 2, dtype='int')
y_predict[:10] #->array([1, 0, 0, 1, 1, 1, 0, 0, 0, 0])
from sklearn.metrics import accuracy_score
accuracy_score(y_test, y_predict)#->0.89600000000000002
使用Voting Classifier
hard-voting-classifier : 少数服从多数
from sklearn.ensemble import VotingClassifier
voting_clf = VotingClassifier(estimators=[
('log_clf', LogisticRegression()),
('svm_clf', SVC()),
('dt_clf', DecisionTreeClassifier(random_state=666))], voting='hard')
voting_clf.fit(X_train, y_train)
voting_clf.score(X_test, y_test)#->0.89600000000000002
soft-voting-classifier:投票需要有权重
使用 Soft Voting Classifier
voting_clf2 = VotingClassifier(estimators=[
('log_clf', LogisticRegression()),
('svm_clf', SVC(probability=True)), #SVC默认不计算概率
('dt_clf', DecisionTreeClassifier(random_state=666))],
voting='soft')
voting_clf2.fit(X_train, y_train)
voting_clf2.score(X_test, y_test) #->0.91200000000000003
Bagging 和 Pasting
- 虽然有很多机器学习方法,但是从投票的角度看,仍然不够多。
- 创建更多的子模型,集成更多子模型的意见。
- 子模型之间不能一致,子模型之间要有差异性。
- 创建差异性 1
- 每个子模型只看样本数据的一部分。例如:一共有500个样本数据,每个子模型只看100个样本数据(可以是同样的分类器)
- 每个子模型不需要太高的准确率——集成学习的威力所在
- 创造差异性 2
- 取样:放回取样,不放回取样
- 放回取样更常用(可以选择的样本数多,并且不是很依赖于取样的随机)
- 令Bagging分类器中的每一个子模型都使用决策树(非参数的学习方式,不对目标函数的形式作出强烈假设,通过不做假设,它们可以从训练数据中自由地学习任何函数形式)更能产生差异较大的子模型
使用 Bagging(放回取样,统计学中叫做bootstrap)
from sklearn.tree import DecisionTreeClassifier
from sklearn.ensemble import BaggingClassifier
bagging_clf = BaggingClassifier(DecisionTreeClassifier(),
n_estimators=500, max_samples=100,
bootstrap=True)#n_estimators=500集成的决策树模型数 max_samples=100每一个子模型样本个数
bagging_clf.fit(X_train, y_train)
bagging_clf.score(X_test, y_test)#->0.91200000000000003
from sklearn.tree import DecisionTreeClassifier
from sklearn.ensemble import BaggingClassifier
bagging_clf = BaggingClassifier(DecisionTreeClassifier(),
n_estimators=5000, max_samples=100,
bootstrap=True)
bagging_clf.fit(X_train, y_train)
bagging_clf.score(X_test, y_test)#->0.92000000000000004
oob和更多Bagging相关
oob(out of bag)
做放回取样时,有一定的概率会出现一部分样本取不到的情况(平均大约有37%的样本取不到)
from sklearn.tree import DecisionTreeClassifier
from sklearn.ensemble import BaggingClassifier
bagging_clf = BaggingClassifier(DecisionTreeClassifier(),
n_estimators=500, max_samples=100,
bootstrap=True, oob_score=True)
bagging_clf.fit(X, y)
Out[4]:
BaggingClassifier(base_estimator=DecisionTreeClassifier(class_weight=None, criterion='gini', max_depth=None,
max_features=None, max_leaf_nodes=None,
min_impurity_decrease=0.0, min_impurity_split=None,
min_samples_leaf=1, min_samples_split=2,
min_weight_fraction_leaf=0.0, presort=False, random_state=None,
splitter='best'),
bootstrap=True, bootstrap_features=False, max_features=1.0,
max_samples=100, n_estimators=500, n_jobs=1, oob_score=True,
random_state=None, verbose=0, warm_start=False)
使用oob_score_
bagging_clf.oob_score_ #->0.91800000000000004
n_jobs
- 集成学习极易进行并行化处理。
- 样本选取是独立的
- 独立地训练若干个子模型
%%time
bagging_clf = BaggingClassifier(DecisionTreeClassifier(),
n_estimators=500, max_samples=100,
bootstrap=True, oob_score=True)
bagging_clf.fit(X, y)
CPU times: user 1.81 s, sys: 27.2 ms, total: 1.84 s
Wall time: 2.95 s
%%time
bagging_clf = BaggingClassifier(DecisionTreeClassifier(),
n_estimators=500, max_samples=100,
bootstrap=True, oob_score=True,
n_jobs=-1)
bagging_clf.fit(X, y)
CPU times: user 385 ms, sys: 56.1 ms, total: 441 ms
Wall time: 1.83 s
bootstrap_features
- 针对特征进行随机取样 Random Subspaces(随机子空间)
- 既针对样本,又针对特征进行随机取样 Random Patches
- Random Patches在二维图像上的体现就是即在行维度随机,又在列维度随机
random subspaces
random_subspaces_clf = BaggingClassifier(DecisionTreeClassifier(),
n_estimators=500, max_samples=500,
bootstrap=True, oob_score=True,
max_features=1, bootstrap_features=True)
#对样本的随机取样关闭:max_samples=500 令每次取样的最大样本个数等于样本总数
#max_features对特征随机取样,每次看1一个特征;
#对特征取样的方式为放回取样
random_subspaces_clf.fit(X, y)
random_subspaces_clf.oob_score_#->0.83399999999999996
random patches
random_patches_clf = BaggingClassifier(DecisionTreeClassifier(),
n_estimators=500, max_samples=100,
bootstrap=True, oob_score=True,
max_features=1, bootstrap_features=True)
#对特征随机取样的同时,对样本随机取样max_samples=100
random_patches_clf.fit(X, y)
random_patches_clf.oob_score_#->0.85799999999999998
随机森林
- Base Estimator:Decision Tree 全部使用决策树作为集成学习的基础分类器
- 决策树在节点划分上,在随机的特征子集上寻找最优划分特征
from sklearn.ensemble import RandomForestClassifier
rf_clf = RandomForestClassifier(n_estimators=500, oob_score=True, random_state=666, n_jobs=-1)
rf_clf.fit(X, y)
Out[4]:
RandomForestClassifier(bootstrap=True, class_weight=None, criterion='gini',
max_depth=None, max_features='auto', max_leaf_nodes=None,
min_impurity_decrease=0.0, min_impurity_split=None,
min_samples_leaf=1, min_samples_split=2,
min_weight_fraction_leaf=0.0, n_estimators=500, n_jobs=-1,
oob_score=True, random_state=666, verbose=0, warm_start=False)
rf_clf.oob_score_ #->0.89200000000000002
rf_clf2 = RandomForestClassifier(n_estimators=500, max_leaf_nodes=16, oob_score=True, random_state=666, n_jobs=-1)#max_leaf_nodes每棵决策树最多的叶子节点数
rf_clf2.fit(X, y)
rf_clf2.oob_score_#->0.90600000000000003
随机森林拥有决策树和BaggingClassifier的所有参数
Extra-Trees极其随机森林
- 决策树在节点划分上,使用随机的特征和随机的阈值
- 提供额外的随机性,抑制过拟合(因为每棵决策树都极其随机,抑制了方差),但增大了bias(偏差)
- 节点的划分毫不费劲,因此有更加快的训练速度
from sklearn.ensemble import ExtraTreesClassifier
et_clf = ExtraTreesClassifier(n_estimators=500, bootstrap=True, oob_score=True, random_state=666, n_jobs=-1)
et_clf.fit(X, y)
et_clf.oob_score_ #->0.89200000000000002
集成学习解决回归问题
from sklearn.ensemble import BaggingRegressor
from sklearn.ensemble import RandomForestRegressor
from sklearn.ensemble import ExtraTreesRegressor
另一种集成学习:Boosting
- 集成多个模型
- 每个模型都在尝试增强(Boosting)整体的效果
AdaBoosting
- 在新的一次学习中,提高上一次学习过程中和模型差距大的点的权重,产生一个新的子模型
- 最终让所有的子模型进行投票
from sklearn.tree import DecisionTreeClassifier
from sklearn.ensemble import AdaBoostClassifier
ada_clf = AdaBoostClassifier(
DecisionTreeClassifier(max_depth=2), n_estimators=500)
ada_clf.fit(X_train, y_train)
Out[5]:
AdaBoostClassifier(algorithm='SAMME.R',
base_estimator=DecisionTreeClassifier(class_weight=None, criterion='gini', max_depth=2,
max_features=None, max_leaf_nodes=None,
min_impurity_decrease=0.0, min_impurity_split=None,
min_samples_leaf=1, min_samples_split=2,
min_weight_fraction_leaf=0.0, presort=False, random_state=None,
splitter='best'),
learning_rate=1.0, n_estimators=500, random_state=None)
ada_clf.score(X_test, y_test)#->0.85599999999999998
Gradient Boosting
- 训练一个模型m1,产生错误e1
- 针对e1训练第二个模型m2,产生错误m2
- 针对e2训练第三个模型m3,产生错误e3……
- 最终预测结果是m1+m2+m3+…
from sklearn.ensemble import GradientBoostingClassifier
gb_clf = GradientBoostingClassifier(max_depth=2, n_estimators=30)
gb_clf.fit(X_train, y_train)
GradientBoostingClassifier(criterion='friedman_mse', init=None,
learning_rate=0.1, loss='deviance', max_depth=2,
max_features=None, max_leaf_nodes=None,
min_impurity_decrease=0.0, min_impurity_split=None,
min_samples_leaf=1, min_samples_split=2,
min_weight_fraction_leaf=0.0, n_estimators=30,
presort='auto', random_state=None, subsample=1.0, verbose=0,
warm_start=False)
gb_clf.score(X_test, y_test)#->0.90400000000000003
Boosting 解决回归问题
from sklearn.ensemble import AdaBoostRegressor
from sklearn.ensemble import GradientBoostingRegressor
### Stacking
- Stacking中,在Layer1中,用三个模型的结果作为Layer2的输入,添加一个模型再训练一次新模型。
- Layer2中同样可以设置多个模型,再汇集到Layer3…
- 每次训练需要用不同的训练数据集。因此,开始要把训练集分成3份,第一份训练Layer1,第二份和第一份的结果训练Layer2
- 回归问题中,令每一个模型得出一个概率值