When dealing with unbalanced class, which is better,

  1. oversampling/undersampling of the classes or
  2. randomly selecting equal number of positive samples and negative samples from the training dataset and combining as training samples, transforming the imbalanced classification problem and replacing by multiple balanced data classification problems?

Does one of them have an advantage over the other? If so, which one? I am asking for a generalized point of view. If you had an unbalanced dataset, which option would you choose, 1 or 2?

  • $\begingroup$ Can you give more insights about your data. In general case your question is meaningless $\endgroup$ – Daniel Chepenko Jul 20 '18 at 8:24
  • 1
    $\begingroup$ Would need to know the data and use case. Many algorithms actually work fine, or can be adapted, to imbalanced classes, so without any further information I would go for secret option 3. Use an algorithm which can deal with the data I have. $\endgroup$ – Ken Syme Jul 20 '18 at 14:18

There can be two things for an imbalanced dataset:

. For eg, in a cancer detection dataset, the number of positive(cancer present) samples, will be quite less than the negative( cancer absent) samples. In this case, you should not apply sampling technqiues. You should rather focus on minimizing False Negatives , as cancer should not be missed while diagnosing a patient at the expense of a few false positives.

.Second case, for eg, The popular Iris Datset where the classes should be balanced. But suppose you get an imbalanced version of the dataset, then you should apply sampling technqiues to the dataset. Because, the class proportion does not match the actual proportion , which was not the case in the Cancer Example.

| improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.