The current implementation of SLIPPER only handles
two-class classification problems. The output of
SLIPPER is a weighted ruleset, in which each rule R
is associated with a confidence CR. To classify an instance
x, one computes the sum of the confidences of
all rules that cover x, then predicts according to the
sign of this sum: if the sum is greater than zero, one
predicts the positive class. In order to make the ruleset
more comprehensible, we further constrain SLIPPER
to generate only rules that are associated with
a positive confidence rating--that is, all rules predict
membership in the positive class. The only rule with
a negative confidence rating (i.e., that predicts membership
in the negative class) is a single default rule.
This representation is a generalization of propositional
DNF, and is similar to that used by many other rule
learners: for most rule learners the classifier is a set of
rules, often with some associated numerical confidence
measure, and often with some sort of voting scheme for
resolving possible conflicts in the predictions.
Below, we describe the SLIPPER algorithm in detail.