2
$\begingroup$

I have a question regarding feature transformation for when some your features are skewed. I've seen online that depending on the skew (right or left) and how great the skew is, it is preferable to use different types of techniques. My question is: Can you scale features based on their skew?

For example, if I have a feature that is right skewed and has significant skew I would use log transform, but If I have a feature that is also right skewed and is moderately skewed I would use square root transform. I've seen online. Just want to find out if you can modify the distribution of some features while not having to touch the distribution of features that already normal. Your clarification would be greatly appreciated!

Improve this question
$\endgroup$
2
  • 2
    $\begingroup$ Distribution of predictors and the shape of the relationship between X and Y are two separate things. $\endgroup$
    – Frank Harrell
    Commented Aug 20, 2022 at 11:30
  • $\begingroup$ Thanks a lot for that! $\endgroup$
    – Igor9094
    Commented Aug 20, 2022 at 13:43

1 Answer 1

Reset to default
2
$\begingroup$

You can mix variable transformations as you wish. If log-transforming one variable and square-root-transforming another works for you, do it; don’t let the fact that they are different transformations stop you.

However, do note that feature normality is seldom a requirement of machine learning models. In fact, none of the ones that come to mind for me require it (though I think about GLM and GLM-oid models like neural networks a lot more than I think about most others).

Improve this answer
$\endgroup$
1
  • $\begingroup$ Thanks a lot for your clarification! $\endgroup$
    – Igor9094
    Commented Aug 20, 2022 at 13:43

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.