Discretization - Data Preprocessing in Data Mining - page 261

Table 9.3 Parameters of the
discretizers and classifiers

Method Parameters

C4.5 Pruned tree, confidence = 0.25, 2 examples per

leaf
DataSqueezer Pruning and generalization threshold = 0.05
KNN K = 3, HVDM distance

PUBLIC 25 nodes between prune

Ripper k = 2, grow set = 0.66

1R 6 examples of the same class per interval

CADD Confidence threshold = 0.01

Chi2 Inconsistency threshold = 0.02

ChiMerge Confidence threshold = 0.05

FDD Frequency size = 30

FUSINTER α = 0 . 975, λ = 1

HDD Coefficient = 0.8

IDD Neighborhood = 3, windows size = 3,

nominal distance

MODL Optimized process type

UCPD Intervals [3, 6], KNN map type,

neighborhood

=

6,
Minimum support = 25, merged
threshold = 0.5,
Scaling factor

=

= 0.5, use discrete

The data sets considered are partitioned using the 10-FCV procedure. The para-
meters of the discretizers and classifiers are those recommended by their respective
authors. They are specified in Table 9.3 for those methods which require them. We
assume that the choice of the values of parameters is optimally chosen by their own
authors. Nevertheless, in discretizers that require the input of the number of intervals
as a parameter, we use a rule of thumb which is dependent on the number of instances
in the data set. It consists in dividing the number of instances by 100 and taking the
maximum value between this result and the number of classes. All discretizers and
classifiers are run one time in each partition because they are non-stochastic.
Two performance measures are widely used because of their simplicity and suc-
cessful application when multi-class classification problems are dealt with. We refer
to accuracy and Cohen's kappa [ 31 ] measures, which will be adopted to measure the
efficacy discretizers in terms of the generalization classification rate. The explanation
of Cohen's kappa was given in Chap. 2 .
The empirical study involves 30 discretization methods from those listed in
Table 9.1 . We want to outline that the implementations are only based on the descrip-
tions and specifications given by the respective authors in their papers.