I understand one should normalize the features in supervised learning.
Does it ever make sense to normalize the response variable?
$\begingroup$
$\endgroup$
Add a comment
|
$\begingroup$
$\endgroup$
The purpose of normalization is to prevent a subset of features from dominating the behavior of the model. So normalizing the output variable is not critical.
Perhaps if you're implementing the model using low-precision or fixed point math. For example, in an MCU for an embedded system. On a server, desktop or laptop using floats or doubles it's not critical to normalize the output.
