Five minutes of machine learning every day: why do we need to normalize the characteristics of numerical types?

This paper mainly

Gradient descent algorithm in the case of multiple features , Use feature scaling techniques , It can make the gradient descent algorithm faster , The number of cycles required for gradient descent is smaller .

If there is a machine learning problem with multiple eigenvalues , We need to ensure that these features have similar scales , This will help the gradient descent algorithm converge faster .

For example

If there are two characteristics in the housing price problem , features x1 Indicates the area value （0_{1000）, features x2 Indicates the number of bedrooms （1}5）, We can see the characteristics x1 The value of is far greater than x2 Of , At this time, we draw the outline of the machine learning problem, which is like this ：

We can see that the image of the loss function will appear very flat , The gradient descent algorithm needs many iterations to converge . If you want to solve this problem , One technique we need to use is feature scaling .

Feature scaling

Feature scaling is trying to scale all features to -1 To 1 Between （ Of course -1.5 To 0.5 It's OK , This can be done in a small range , Of course, if it is 0.00000001 To 0.00001 This is not a very good feature scaling ）. The specific solution is to let us define x1= Building area /2000、x2= Number of bedrooms /5