You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
GAM or partially additive models can be fitted by components using backfitting.
This is very similar to the more recent group coordinate descend algorithms.
We can fit any GAM with specified basis functions as one large single optimization problem, however for some parts like choosing penalization weights or for plots it might be useful to have a "marginal" model that just has one component and all other terms are in the offset.
This will be a mixture between backfitting/group coordinate descend and IRLS or standard optimization.
The text was updated successfully, but these errors were encountered:
related issues: #5689 GAM with backfitting in sandbox (This would now be simpler with constant removed from spline basis, but constant handling/removal would still be needed with kernel regression.) #2710 sparse matrix support in models
I also have a prototype somewhere with partial linear model that absorbs or backfits linear part
elastic net fit_regularized uses coordinate descent in a "backfitting" loop
It would be good to get a prototype for the optimization part, i.e. the fit method, the surroundings would be as in a standard models
mainly an idea for some implementation details
GAM or partially additive models can be fitted by components using backfitting.
This is very similar to the more recent group coordinate descend algorithms.
We can fit any GAM with specified basis functions as one large single optimization problem, however for some parts like choosing penalization weights or for plots it might be useful to have a "marginal" model that just has one component and all other terms are in the offset.
This will be a mixture between backfitting/group coordinate descend and IRLS or standard optimization.
The text was updated successfully, but these errors were encountered: