When using class weights is bad?

1
  • I have a DB with 50 different classes.
  • One of the classes has x10 more data than the other classes.
  • Each class has ~20K samples and the ‘big’ class has ~200K samples
  • When training classification model without class weights - I’m getting slightly better results than training the model with class weights.
    image
    image1113×164 4.67 KB
  1. How can it be ?
  2. According to my results - it seems that if I have an important class which I don’t want to have many mistakes - it is better to get more samples and it will not affect badly on the other classes, am I right ?
2

The small difference in accuracy could probably be explained with the Accuracy Paradox.

3

I ran classification_report and checked the F1 for each class.
The results (F1) are very very close (<1% difference) (for each classes) between with or without class weights.
So, according to my results, is there any logic why do I need to use class weights ?

4

Class weights are used to add a penalty to the model training e.g. to counter overfitting a majority class. Based on your previous output the model is already overfitting to the majority class.