Selecting the regularisation coefficient can be difficult. It would therefore be useful for a utility that retrains many models (with warm start) and logs the regularisation coefficients as well as which groups are chosen with that regularisation coefficient.
There are two ways to accomplish this, one naive, simply doing a linear or logarithmic search on the group regularisation coefficient. The other approach I can think of is to start with a high regularisation and do a bisecting search for regularisation coefficients that lead to new groups being added and removed from the chosen variables. This latter approach is much more complex but may be implemented if there are interest (or if I have time to spare).
Selecting the regularisation coefficient can be difficult. It would therefore be useful for a utility that retrains many models (with warm start) and logs the regularisation coefficients as well as which groups are chosen with that regularisation coefficient.
There are two ways to accomplish this, one naive, simply doing a linear or logarithmic search on the group regularisation coefficient. The other approach I can think of is to start with a high regularisation and do a bisecting search for regularisation coefficients that lead to new groups being added and removed from the chosen variables. This latter approach is much more complex but may be implemented if there are interest (or if I have time to spare).