search

stephenslab / susieR

R package for "sum of single effects" regression.
https://stephenslab.github.io/susieR
Other
176 stars 45 forks source link

Add option to initialize variational parameters #6

Closed jhmarcus closed 6 years ago

jhmarcus commented 6 years ago

Add an argument in susie() and / or susie_auto() to take alpha_init as an argument. The use case is to initialize with the alpha using fast a best subset selection procedure i.e. implemented in https://github.com/hazimehh/L0Learn

stephens999 commented 6 years ago

currently susie can take a previous susie fit object as init. so would it work to write a function that creates a suitable susie fit for a given alpha that could then be used as s_init ?

stephens999 commented 6 years ago

also do you really want to pass in alpha, or just a list of L "candidate" covariates to initialize at? (corresponding to each row of alpha has one 1 and all others 0s, like 0000010000 )

jhmarcus commented 6 years ago

ah yes just looked a bit more closely at that.

jhmarcus commented 6 years ago

Ok so the previous "susie fit" is s_init. Ya I think a list of candidate covariates makes sense thanks!

gaow commented 6 years ago

@KaiqianZhang I think we should formalize & generalize your L0Learn initialization function -- I'm more talking about an interface. For example:

susie_set_initialization(pos, value=NULL, ...)

where pos is the index vector of non-zero effects to initialize alpha, and value is optional the effect size to initialize mu. We need to think of default behaviors for other ... parameters.