Information gain split to center to 0 numerical attributes for linear models?

Consider scenarios where attributes have a clear correlation with the label (binominal). Couldn't it be beneficial to center on 0 based on where an information gain split would be? Instead of for example centering based on standard deviation or simply scaling from 0 to 1.

It would seem the classes would become more separable, maybe useful for when accuracy is needed. But maybe it wouldn't change much for AUC. Furthermore I have no idea how this scales beyond a simple linear model.

Did anybody try this? I don't think there's a way to do it in RapidMiner right now, is there?

Just a thought, maybe someone with more knowledge can easily twist it around and make something of it.