In this paper we analyze experimentally the effect of other discretization
methods for ECL, in particular local supervised multivariate discretization, and
test and compare the resulting variants of ECL.
We propose a discretization method which uses the intervals generated by
a given (global supervised univariate discretization) algorithm for initializing
the inequalities introduced in a rule, and refines these inequalities, during the
evolutionary process, by means of mutation operators which use specific cut
points for shifting inequality boundaries. More specifically, we consider the
two following possible initializations of inequalities: a fine grain initialization,
using intervals formed by two consecutive boundary points, where a boundary
point is the midpoint of two consecutive attribute values having different class
labels [18]; and a coarser grain initialization, using intervals obtained from the
Fayyad & Irani algorithm (outlined above). During the evolutionary process
the mutation operators use the boundary points for modifying the inequalities.
The resulting ECL variants are called ECL-LSDf and ECL-LSDc, respectively.
We compare experimentally four variants of ECL with discretization: ECL
with Global univariate Discretization (ECL-GSD) which uses Fayyad & Irani algorithm
prior to evolution, ECL with local multivariate unsupervised discretization
(ECL-LUD), and ECL with the two variants ( ECL-LSDf and ECL-LSDc)
of the proposed local supervised discretization method.
In this paper we analyze experimentally the effect of other discretizationmethods for ECL, in particular local supervised multivariate discretization, andtest and compare the resulting variants of ECL.We propose a discretization method which uses the intervals generated bya given (global supervised univariate discretization) algorithm for initializingthe inequalities introduced in a rule, and refines these inequalities, during theevolutionary process, by means of mutation operators which use specific cutpoints for shifting inequality boundaries. More specifically, we consider thetwo following possible initializations of inequalities: a fine grain initialization,using intervals formed by two consecutive boundary points, where a boundarypoint is the midpoint of two consecutive attribute values having different classlabels [18]; and a coarser grain initialization, using intervals obtained from theFayyad & Irani algorithm (outlined above). During the evolutionary processthe mutation operators use the boundary points for modifying the inequalities.The resulting ECL variants are called ECL-LSDf and ECL-LSDc, respectively.We compare experimentally four variants of ECL with discretization: ECLwith Global univariate Discretization (ECL-GSD) which uses Fayyad & Irani algorithmprior to evolution, ECL with local multivariate unsupervised discretization(ECL-LUD), and ECL with the two variants ( ECL-LSDf and ECL-LSDc)of the proposed local supervised discretization method.
การแปล กรุณารอสักครู่..