Hello !!

I read your article and it says:
"First step is to assign a linspace to the sorted features from -1…1, then apply the inverse of error function ErfInv to shape them like gaussians, then I substract the mean. Binary features are not touched with this trafo (eg. 1-hot ones). This works usually much better than standard mean/std scaler or min/max."
http://fastml.com/preparing-continuous-features-for-neural-networks-with-rankgauss/

But you mentioned Jahrer interview and he says something different:
"normalization
Input normalization for gradient-based models such as neural nets is critical. For lightgbm/xgb it does not matter. The best what I found during the past and works straight of the box is "RankGauss". Its based on rank transformation. First step is to assign a linspace to the sorted features from 0..1, then apply the inverse of error function ErfInv to shape them like gaussians, then I substract the mean. Binary features are not touched with this trafo (eg. 1-hot ones). This works usually much better than standard mean/std scaler or min/max."
https://www.kaggle.com/c/porto-seguro-safe-driver-prediction/discussion/44629

So should it be sort from -1 to 1 or sort from 0 to 1 ?

Thanks

Hello, if I want to use my algorithm online with gaussrank, is it possible to implement a version with separated fit and predict function please ? While using quantiles/percentiles instead of sorted ranks for example ?