Should only be used by experts! This selects for every task and cell the best hyper-parameter based on the validation errors in the folds. This is saved and will afterwards be used in the evaluation of the decision functions.
selectSVMs(model, command.args = NULL, ..., d = NULL,
warn.suboptimal = getOption("liquidSVM.warn.suboptimal", TRUE))
the svm
-model
further arguments aranged in a list, corresponding to the arguments
of the command line interface to svm-select
, e.g. list(d=2,R=0)
is equivalent to svm-select -d 2 -R 0
.
See command-args for details.
parameters passed to selection phase e.g. retrain_method="select_on_entire_train_set"
level of display information
if TRUE this will issue a warning
if the boundary of the hyper-parameter grid was hit too many times.
The default can be changed by setting options(liquidSVM.warn.suboptimal=FALSE)
.
a table giving training and validation errors and more internal statistic
for all the SVMs that were selected.
This is also recorded in model$select_errors
.
The following parameters can be used as well:
h=[<level>]
Displays all help messages.
Meaning of specific values: <level> = 0 => short help messages <level> = 1 => detailed help messages
Allowed values: <level>: 0 or 1
Default values: <level> = 0
N=c(<class>,<constraint>)
Replaces the best validation error in the search for the best hyper-parameter combination by an NPL criterion, in which the best detection rate is searched for given the false alarm constraint <constraint> on class <class>.
Allowed values: <class>: -1 or 1 <constraint>: float between 0.0 and 1.0
Default values: Option is deactivated.
R=<method>
Selects the method that produces decision functions from the different folds.
Meaning of specific values: <method> = 0 => select for best average validation error <method> = 1 => on each fold select for best validation error
Allowed values: <method>: integer between 0 and 1
Default values: <method> = 1
W=<number>
Restrict the search for the best hyper-parameters to weights with the number <number>.
Meaning of specific values: <number> = 0 => all weights are considered.
Default values: <number> = 0
Some learning scenarios have to perform several selection runs:
for instance in quantile regression for every quantile.
This is done by specifying weight_number
ranging from 1 to the number of quantiles.
See command-args for details.
command-args, svm
, init.liquidSVM
, selectSVMs
, predict.liquidSVM
, test.liquidSVM
and clean.liquidSVM