This routine provides binary classifiers that satisfy a
predefined error rate on one type of error and that
simlutaneously minimize the other type of error. For
convenience some points on the ROC curve around the
predefined error rate are returned.
nplNPL performs Neyman-Pearson-Learning for classification.
Usage
nplSVM(x, y, ..., class = 1, constraint = 0.05,
constraint.factors = c(3, 4, 6, 9, 12)/6, do.select = TRUE)
Arguments
x
either a formula or the features
y
either the data or the labels corresponding to the features x.
It can be a character in which case the data is loaded using liquidData.
If it is of type liquidData then after training and selection
the model is tested using the testing data (y$test).
...
configuration parameters, see Configuration. Can be threads=2, display=1, gpus=1, etc.
class
is the normal class (the other class becomes the alarm class)
constraint
gives the false alarm rate which should be achieved
constraint.factors
specifies the factors around constraint
do.select
if TRUE also does the whole selection for this model
Value
an object of type svm. Depending on the usage this object
has also $train_errors, $select_errors, and $last_result
properties.
Details
Please look at the demo-vignette (vignette('demo')) for more examples.
The labels should only have value c(1,-1).
min_weight, max_weight, weight_steps: you might have to define
which weighted classification problems will be considered.
The choice is usually a bit tricky. Good luck ...