What is label encoding? How to distinguish and use one hot encoding and label encoding?

What is? Label encoding

label encoding It means encoding with tags , That is, the original eigenvalue is encoded into a customized digital tag to complete the quantization coding process .

give an example :
If there are three color characteristics ： red 、 yellow 、 blue . In the use of machine learning algorithms, it is generally necessary to carry out vectorization or digitization . Then you may want to make red =1, yellow =2, blue =3. So this actually implements tag coding , That is to label different categories .

characteristic

• advantage ： It solves the problem of classification and coding , Quantitative numbers can be defined freely . But it is also a disadvantage , Because the value itself has no meaning , It's just sorting . For example, the code of large, medium and small is 123, It can also be coded as 321, That is, the value is meaningless .
• shortcoming ： Poor interpretability . Such as the [dog,cat,dog,mouse,cat], We turn it into [1,2,1,3,2], There is a strange phenomenon ：dog and mouse The average value of is cat. therefore ,Label encoding Coding doesn't have a wide application scenario .

one-hot encoding ,label encoding How to distinguish and use the two codes ？

1. Characteristic data type

For categorical data , It is recommended to use one-hot encoding. Classification is pure classification , Don't order , There's no logic . For example, gender is divided into male and female , There is no logical relationship between men and women , We can't say that men are better than women , Or vice versa . also , The classification of provinces and cities in China can also use the unique hot coding , Similarly, there is no logical relationship between provinces , Use at this time one-hot encoding It's better to meet you . But pay attention to , Generally, one variable is left out , For example, the opposite of men must be women , So women are repeating information , So just keep one of the variables .

For ordered data , It is recommended to use label encoding. Ordering types are also classifications , But there's sort logic , Higher in rank than in class . such as , Education is divided into primary schools , Junior high school , high school , Undergraduate , Graduate student , There is a certain logic between the categories , Obviously graduate education is the highest , Primary school is the lowest . Use at this time Label encoding It would be more appropriate , Because the custom number order can not destroy the original logic , And corresponding to this logic .

2. The model used

Models that are sensitive to numerical size must use one-hotencoding. A typical example is LR and SVM. The loss function of the two is sensitive to the numerical value , And the numerical value between variables is significant . and Label encoding There is no meaning of numerical value in the digital code of , It's just a sort of ordering , So for all of these models one-hot encoding.

Models insensitive to numerical size （ Like a tree model ） Not recommended one-hotencoding. Generally, this kind of model is tree model . If there are too many categories , that one-hot encoding It splits out a lot of characteristic variables . Now , If we limit the depth of the tree model and can't split it down , Some characteristic variables may be abandoned because the model can not continue to split . therefore , In this case, we can consider using Label encoding.

The above two considerations need to be considered comprehensively , Instead of judging alone . That is to say, we need to choose the coding method according to the data type and model .

————————————————
Copyright notice ： This paper is about CSDN Blogger 「 Plain paper and breeze 」 The original article of , follow CC 4.0 BY-SA Copyright agreement , For reprint, please attach the original source link and this statement .
Link to the original text ：https://blog.csdn.net/weixin_45834085/article/details/102991983