Dear Raz Malka ,

my suggestion is to perform class balancing only when necessary.

Most classification algorithms are unaffected by small unbalanced classes.

Moreover, to perform class balancing you should either:

- undersample the most represented class, thus losing information

- oversample the less represented class (either directly or using some algorithm), thus promoting overfitting

You may slightly change the cost function, assigning a weight to each error, which is inversely proportional to the class size.

Hi Raz Malka,

I think this is not a necessary step; the data in real life by its nature is unbalanced. Deal with the data you have it as it is!!

Dear Raz Malka

These article might be useful, have a look

1. https://towardsdatascience.com/balancing-is-unbalancing-5f517936f626

2. https://www.r-bloggers.com/2020/06/why-balancing-your-data-set-is-important/

Kind Regards

Qamar Ul Islam

Dear Raz Malka ,

When the imbalance ratio is nearly 1:1, i.e., 55:45 or 60:40 majority and minority class ratio, You may not need to balance the dataset either using oversampling or undersampling. This much imbalance is very trivial in real-life datasets.

But, if the minority class examples are very important for the correct predictions (e.g- disease datasets), and you don't want any minority class data to be left out due to an imbalanced dataset as it may cause severe effect, you can perform undersampling or oversampling to have better and accurate results.

Regards

Sayan Surya Shaw