I am trying to implement a neural network on a regression problem. I have scaled the independent variables since this is a crucial step for neural networks. However, I see online that some people recommend scaling the target variable also, and at the same time, there are other opinions that don't recommend doing feature scaling on target variables for neural networks. This is confusing as I didn't find any clear guidelines for the problem of scaling the target variables for neural networks. So Please if someone illustrates what should I do exactly regarding scaling the target variables? Should I always scale the target variables ? or it depends on the case? Or I should never scale the target variables for neural networks in regression problems? Moreover, if feature scaling the target variable is possible, do I still use MSE after I return back the target variable to its old scale, or feature scaling the target variable will affect MSE in some way?
$\begingroup$
$\endgroup$
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
Here is a good answer for why scaling might sometimes be absolutely necessary in order to obtain useful results:
Why should I normalize also the output data?
Another point to consider is the scale of your model's output relative to the data you're fitting onto. If your output is in [0, 1] but your data is outside of that range, you will either need to adapt your model or scale your data.
