3
$\begingroup$

Suppose I have 2 classes. One class has 16 samples and the other class has 435 samples. Is it justified to oversample the class with 16 sample to have a 435 number of samples? Or is it better to undersample the class with 435 samples? If so, what should be the number of samples after the undersampling is carried out?

$\endgroup$

1 Answer 1

2
$\begingroup$

This depends on the nature of your data. If you can effectively simulate 435 samples using any given oversampling methods such as SMOTE or ADASYN for instance, then I would say oversampling would be better. Because it would provide data for various scenarios. But if exact replication is not possible and if replicated there might be an inherent problem with the model then you should choose to undersample.

A good example for undersampling would be that for example if you're conducting a scientific experiment and the data you have are limited to a few scenarios and the other data is yet to be fully observed, you would choose to undersample.

Oversampling would be the reverse case, where you can effectively simulate data for various classes based on some parameter and the generated data can mimic actual data and scenarios, then you should oversample.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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