机器学习中为什么要做归一化normalization

我们处理feature的时候往往先要normalize encoding&＃xff0c;使用python可以很容易做&＃xff1a;

from sklearn import preprocessing
from scipy.stats import rankdata
x &＃61; [[1], [3], [34], [21], [10], [12]]
std_x &＃61; preprocessing.StandardScaler().fit_transform(x)
norm_x&＃61; preprocessing.MinMaxScaler().fit_transform(x)
norm_x2&＃61; preprocessing.LabelEncoder().fit_transform(x)
print(&＃39;std_x&＃61;\n&＃39;, std_x)
print(&＃39;norm_x&＃61;\n&＃39;, norm_x)
print(&＃39;norm_2&＃61;\n&＃39;, norm_x2)
print(&＃39;oringial order &＃61;&＃39;, rankdata(x))
print(&＃39;stand order &＃61;&＃39;, rankdata(std_x))
print(&＃39;normalize order&＃61;&＃39;, rankdata(norm_x))


其中preprocessing.LabelEncoder().fit_transform(x)就是做normalize encoding&＃xff0c;上面的程序输入如下&＃xff1a;

std_x&＃61;
[[-1.1124854 ]
[-0.93448773]
[ 1.82447605]
[ 0.66749124]
[-0.31149591]
[-0.13349825]]
norm_x&＃61;
[[0. ]
[0.06060606]
[1. ]
[0.60606061]
[0.27272727]
[0.33333333]]
norm_2&＃61;
[0 1 5 4 2 3]
oringial order &＃61; [1. 2. 6. 5. 3. 4.]
stand order &＃61; [1. 2. 6. 5. 3. 4.]
normalize order&＃61; [1. 2. 6. 5. 3. 4.]


可以看到normailize之后的结果是 [0 1 5 4 2 3]。这样做的好处是什么呢&＃xff1f;

下面图片转自知乎&＃xff08;https://www.zhihu.com/questio...&＃xff09;