Machine learning needs normalization algorithm has SVM, logistic regression, neural network, KNN, linear regression,
Tree without the need of normalization , because they do not care about the value of the variable, but about a conditional probability distribution between the variables and variables, such as decision trees, random forests, for the tree structure, the tree model by Looking for the best split point constituted, the value of sample points zoom position does not affect the split point, the tree structure of the model is not affected,
and the tree can not be the model gradient descent, because the tree model is the step, the step is not derivable and derivation does not make sense, does not require normalization
2, the model for those who need normalization, mainly because the difference between the feature value is large, the use of gradient descent, loss contour is oval, require multiple iterations to achieve the best advantage, if the normalized, then the contour is rounded , to promote SGD origin iteration, resulting in fewer iterations required.
https://blog.csdn.net/qq_34872215/article/details/88363504
