Strong rules for BASIL is different from the one proposed for glmnet.
At every basil iteration, we must do a KKT check where we compute the gradient for all variables. Current formulation of BASIL uses a simple strong rule where we add delta number of variables with the highest (absolute) gradient. The paper mentions one extension is to use the usual strong rules as in glmnet to add variables in a data-dependent fashion. The usual strong rule rarely discards more than necessary, so this seems like a better rule.
Strong rules for BASIL is different from the one proposed for
glmnet
.At every basil iteration, we must do a KKT check where we compute the gradient for all variables. Current formulation of BASIL uses a simple strong rule where we add
delta
number of variables with the highest (absolute) gradient. The paper mentions one extension is to use the usual strong rules as inglmnet
to add variables in a data-dependent fashion. The usual strong rule rarely discards more than necessary, so this seems like a better rule.