0
$\begingroup$

I have a few question about log transformation and standardization.

First: Should I standardize my features after doing log transformation?

Second: I still do not understand, because when doing log transformation it change the shape distribution to normal distribution. But when do standarization (standardscaler) it will not change the distribution shape. So which one should I use if I have skewed features?

Third: I have seen in many Kaggle competitions that people tend to use standardscaler over log transformation event though the features are skewed. What is the reason? If it is okay to use standard scaler, does that mean that machine learning models can also work well even having skewed features?

$\endgroup$

1 Answer 1

1
$\begingroup$

The logarithm is a non-linear transformation: It's normal that the result distribution is quite different from the raw one.

Then the use of log or standard scaler depends on the algorithm. There is no universal rule to set whether log is worse or better than standard scaler. The best you know how the algorithm works, the better you know which preprocessing formula would work best.

$\endgroup$

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.