Regularization and continuous annotations about polyfun

[omerwe]

Hi Gabe,

That's a very good question. There's no right answer, i.e., there's a scale for each annotation that will yield optimal prediction accuracy, but there's no simple way to know this scale in advance. However, I suspect that the effect will be negligible, unless these are very high impact annotations for the trait you're studying. Hence, I would normalize continuous annotations to have a unit variance of 1.0.

BTW one could ask the same question about binary annotations. i.e., encoding a binary annotation as 1/0 is different than encoding it as 2/0. These are good questions but we haven't studied them in our paper so I don't have very good answers, except that I suspect that it probably wouldn't make a big difference in most cases.

Best,

Omer