What is the debiasing slope and intercept coefficient when using SVEM forward selection on a Beta distribution?
I do not find statistical details in the documentation....
If someone have informations it would be great!
Thank you,
Florence
Hi Florence,
I think this is a feature of all SVEM models, not just for Beta distribution.
The reason for it is that SVEM models tend to have a bias: they predict too high for low Y, and too low for high Y. It's very clear when you see the actual vs predicted for a SVEM model that has not been bias corrected. I believe this is an example of the bias-variance tradeoff that you also see with penalized regression models.
Chris Gotwalt found a simple solution to correct for this bias: Fit a simple regression of Y vs PredY(SVEM). The debiasing slope and intercept parameters are the parameters of this simple regression.
I hope that helps. @Peter_Hersh and I talk a bit about this (at 14:08) in our Discovery Summit presentation from 2021:
SVEM: Machine Learning Applied to DOE (2021-US-45MP-847) - JMP User Community
