一、标准化（Z-score），或者去除均值和方差缩放
公式为：(X-mean)/std 计算时对每个属性/每列分别进行。
将数据按期属性（按列进行）减去其均值，并处以其方差。得到的结果是，对于每个属性/每列来说所有数据都聚集在0附近，方差为1。
实现时，有两种不同的方式：
-
使用sklearn.preprocessing.scale()函数，可以直接将给定数据进行标准化。
|
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
|
>>> from sklearn
import
preprocessing
>>>
numpy
as
np
>>> X = np.array([[
1
.,-
ottom:auto!important; float:none!important; height:auto!important; left:auto!important; line-height:1.8em!important; outline:0px!important; overflow:visible!important; position:static!important; right:auto!important; top:auto!important; vertical-align:baseline!important; width:auto!important; font-family:Consolas,
2
.],
... [
0
ottom:auto!important; float:none!important; height:auto!important; left:auto!important; line-height:1.8em!important; outline:0px!important; overflow:visible!important; position:static!important; right:auto!important; top:auto!important; vertical-align:baseline!important; width:auto!important; font-family:Consolas,
.]])
>>> X_scaled = preprocessing.scale(X)
>>> X_scaled
array([[
. ...,0)!important; background:none!important">1.22
...,0)!important; background:none!important">1.33
...],
[
0.26
ottom:auto!important; float:none!important; height:auto!important; left:auto!important; line-height:1.8em!important; outline:0px!important; overflow:visible!important; position:static!important; right:auto!important; top:auto!important; vertical-align:baseline!important; width:auto!important; font-family:Consolas,
[-
1.06
...]])
>>>#处理后数据的均值和方差
>>> X_scaled.mean(axis=
)
array([
.])
>>> X_scaled.std(axis=
)
.])
|
-
使用sklearn.preprocessing.StandardScaler类，使用该类的好处在于可以保存训练集中的参数（均值、方差）直接使用其对象转换测试集数据。
|
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
|
>>> scaler
=
preprocessing.StandardScaler().fit(X)
>>> scaler
StandardScaler(copy
=
True
,with_mean
ottom:auto!important; float:none!important; height:auto!important; left:auto!important; line-height:1.8em!important; outline:0px!important; overflow:visible!important; position:static!important; right:auto!important; top:auto!important; vertical-align:baseline!important; width:auto!important; font-family:Consolas,with_std
)
>>> scaler.mean_
array([
1.
ottom:auto!important; float:none!important; height:auto!important; left:auto!important; line-height:1.8em!important; outline:0px!important; overflow:visible!important; position:static!important; right:auto!important; top:auto!important; vertical-align:baseline!important; width:auto!important; font-family:Consolas,
0.
0.33
...])
>>> scaler.std_
0.81
1.24
...])
>>> scaler.transform(X)
array([[
ottom:auto!important; float:none!important; height:auto!important; left:auto!important; line-height:1.8em!important; outline:0px!important; overflow:visible!important; position:static!important; right:auto!important; top:auto!important; vertical-align:baseline!important; width:auto!important; font-family:Consolas,
-
1.22
1.33
ottom:auto!important; float:none!important; height:auto!important; left:auto!important; line-height:1.8em!important; outline:0px!important; overflow:visible!important; position:static!important; right:auto!important; top:auto!important; vertical-align:baseline!important; width:auto!important; font-family:Consolas,
[
0.26
ottom:auto!important; float:none!important; height:auto!important; left:auto!important; line-height:1.8em!important; outline:0px!important; overflow:visible!important; position:static!important; right:auto!important; top:auto!important; vertical-align:baseline!important; width:auto!important; font-family:Consolas,
[
1.06
...]])
>>>
#可以直接使用训练集对测试集数据进行转换
>>> scaler.transform([[
]])
array([[
2.44
...]])
|
二、将属性缩放到一个指定范围
除了上述介绍的方法之外，另一种常用的方法是将属性缩放到一个指定的最大和最小值（通常是1-0）之间，这可以通过preprocessing.MinMaxScaler类实现。
1、对于方差非常小的属性可以增强其稳定性。
2、维持稀疏矩阵中为0的条目。
3
4
12
19
20
21
22
23
24
np.array([[
2.
],
... [
ottom:auto!important; float:none!important; height:auto!important; left:auto!important; line-height:1.8em!important; outline:0px!important; overflow:visible!important; position:static!important; right:auto!important; top:auto!important; vertical-align:baseline!important; width:auto!important; font-family:Consolas,
]])
...
>>> min_max_scaler
preprocessing.MinMaxScaler()
>>> X_train_minmax
min_max_scaler.fit_transform(X_train)
>>> X_train_minmax
0.5
ottom:auto!important; float:none!important; height:auto!important; left:auto!important; line-height:1.8em!important; outline:0px!important; overflow:visible!important; position:static!important; right:auto!important; top:auto!important; vertical-align:baseline!important; width:auto!important; font-family:Consolas,
0.33333333
ottom:auto!important; float:none!important; height:auto!important; left:auto!important; line-height:1.8em!important; outline:0px!important; overflow:visible!important; position:static!important; right:auto!important; top:auto!important; vertical-align:baseline!important; width:auto!important; font-family:Consolas,
]])
>>>
#将相同的缩放应用到测试集数据中
>>> X_test
np.array([[
3.
4.
]])
>>> X_test_minmax
min_max_scaler.transform(X_test)
>>> X_test_minmax
1.5
1.66666667
]])
#缩放因子等属性
>>> min_max_scaler.scale_
...])
>>> min_max_scaler.min_
...])
