Introduction
Boosting (Schapire 1990; Freund 1995; Freund
Schapire 1997) is usually used to create ensemble classifters.
It is popular because it is simple, easy to
implement, well-understood formally, and effective at
improving accuracy. One disadvantage of boosting
is that improvements in accuracy are often obtained
at the expense of comprehensibility. If comprehensibility
is important, it is more appropriate to use
some learner that produces a compact, understandable
hypothesis--for instance, a rule learning system like
CN2 (Clark & Niblett 1989), RIPPER (Cohen 1995),
or C4.5rules (Quinlan 1994). However, the rule learning
systems that perform best experimentally have the
disadvantage of being complex, hard to implement, and
not well-understood formally.
Here, we describe a new rule learning algorithm
called SLIPPER.1 SLIPPER generates rulesets by repeatedly
boosting a simple, greedy, rule-builder. SLIPPER’s
rule-builder is much like the inner loops of RIPPER
(Cohen 1995) and IREP (Fiirnkranz & Widmer
1994). However, SLIPPER does not employ the "setcovering"
process used by conventional rule learners--
rather than removing examples covered by a new rule,
SLIPPER uses boosting to reduce the weight of these
examples.