Beginner little rookie , I hope it's like taking notes and recording what I've learned , Also hope to help the same entry-level people , I hope the big guys can help correct it ~ Tort made delete .

Catalog

One 、 Problem definition

Two 、 Solutions to problems

1、 Solution introduction

2、 Ideal solution

3、 Practical solutions

One 、 Problem definition

The number of training samples in different categories in classification tasks varies greatly .

Two 、 Solutions to problems

1、 Solution introduction

In linear classifier , We use it When classifying new samples , We predict with a threshold y Value comparison . Usually we choose 0.5 Threshold value （ Think that the possibility of true positive and negative examples is the same ）, namely ：

  when , The prediction is a positive example .

2、 Ideal solution

But when the number of positive and negative examples in the training set is different , The observation probability is , among Is the number of positive examples , Is the number of counterexamples . Usually, we assume that the training set is unbiased sampling of the real sample population , Then the observed probability represents the real probability . There are ：

when , The prediction is a positive example , The opposite is true .

namely ：, This is a basic strategy in category imbalance learning “ Zoom again ”（ Similar to cost sensitive learning Instead of , among The cost of being divided into positive classes by mistake into negative classes ）.

3、 Practical solutions

The assumption of the above solution is “ The training set is the unbiased sampling of the real sample population ”, But this assumption is not necessarily true , That is, it is difficult to deduce the real probability from the observed probability .

Therefore, there are three actual treatment schemes ： 

（1） Directly under sample the anti class samples in the training set , That is to remove some anti class samples , Make the number of positive and negative samples close to （ Multiple random undersampling can be performed , Training multiple classifiers , The test results are the most predicted results of these classifiers , In this way, there is not so much information lost ）

（2） Oversampling the positive samples in the training set , That is, add some positive samples , Make the number of positive and negative samples close to （ Oversampling —— Additional samples can be generated by interpolating positive samples , You cannot simply repeat the sample , It's easy to over fit ）

（3） Use the ideal solution above —— Threshold shift

You are welcome to criticize and correct in the comment area , Thank you very much! ~