Consolidated Trees: An Analysis of Structural Convergence - Data Mining: Theory, Methodology, Techniques, and Applications - page 43

proportions appearing in the figure are representative of the proportions happening in
each step of the CT's building process, X c will be “sex” in the first step with branches
“m” and “f”, and “age” in the second one with 30 selected as cut value.
In the last step of the example in Fig. 1, the proposal is not to split the node in
two of the partitions and in another one the proposal is to split it using “color”
variable. If this proportion is maintained, the final decision will be to consolidate
the node as a leaf.

3 Experimental Methodology

Twenty databases of real applications have been used for the experimentation. Most
of them belong to the well known UCI Repository benchmark [2]. The Segment
domain has been used for experimentation in two different ways: taking into account
the whole set of data ( segment2310 ) and conserving the training/test division of the
original data set ( Segment210) . The Faithful database is a real data application from
our environment, centred in the electrical appliance's sector. Table 1 shows the wide
range of characteristics of the used domains: the number of patterns ( N. of patterns ) goes
from 148 to 24,507, the number of features ( N. of features ) from 4 to 57 and the
number of classes of the dependent variable ( N. of classes ) from 2 to 15.

Table 1. Description of experimental domains

Domain N. of patterns N. of features N. of classes

Breast-W

699 10 2

Heart-C

303 13 2

Hypo

3163 25 2

Lymph

148 18 4

Credit-G

1000 20 2

Segment210

210 19 7

Iris

150 4 3

Glass

214 9 7

Voting

435 16 2

Hepatitis

155 19 2

Soybean-L

290 35 15

Sick-E

3163 25 2

Liver

345 6 2

Credit-A

690 14 2

Vehicle

846 18 4

Breast-Y

286 9 2

Heart-H

294 13 2

Segment2310

2310 19 7

Spam

4601 57 2

Faithful

24507 49 2

The CTC methodology has been compared to the C4.5 tree building algorithm
Release 8 of Quinlan, using the default parameter settings. Both kinds of trees have
been pruned, using the pruning algorithm of the C4.5 R8 software, to situate both