機器學習中爲何要作歸一化normalization

咱們處理feature的時候每每先要normalize encoding，使用python能夠很容易作：

from sklearn import preprocessing
from scipy.stats import rankdata

x = [[1], [3], [34], [21], [10], [12]]
std_x = preprocessing.StandardScaler().fit_transform(x)
norm_x= preprocessing.MinMaxScaler().fit_transform(x)
norm_x2= preprocessing.LabelEncoder().fit_transform(x)

print('std_x=\n', std_x)
print('norm_x=\n', norm_x)
print('norm_2=\n', norm_x2)

print('oringial order =', rankdata(x))
print('stand order    =', rankdata(std_x))
print('normalize order=', rankdata(norm_x))

其中preprocessing.LabelEncoder().fit_transform(x)就是作normalize encoding，上面的程序輸入以下：

std_x=
 [[-1.1124854 ]
 [-0.93448773]
 [ 1.82447605]
 [ 0.66749124]
 [-0.31149591]
 [-0.13349825]]
norm_x=
 [[0.        ]
 [0.06060606]
 [1.        ]
 [0.60606061]
 [0.27272727]
 [0.33333333]]
norm_2=
 [0 1 5 4 2 3]
oringial order = [1. 2. 6. 5. 3. 4.]
stand order    = [1. 2. 6. 5. 3. 4.]
normalize order= [1. 2. 6. 5. 3. 4.]

能夠看到normailize以後的結果是 [0 1 5 4 2 3]。這樣作的好處是什麼呢？

下面圖片轉自知乎（https://www.zhihu.com/questio...）