Information Technology Reference
In-Depth Information
6.5.2.1
Datasets
The experiments reported in [200] used several real-world datasets and also
the artificial one depicted in Fig. 6.38 called ArtificialF2. The real-world
datasets are all publicly available: the Breast Tissue, CTG, Diabetes and
Sonar datasets can be found in [13], PB12 in [110] and Olive in [75]. Ta-
ble 6.24 contains a summary of the characteristics of these datasets showing
the number of data instances, number of features and the number of classes
for each dataset.
Tabl e 6 . 24
The datasets used in the experiments reported in [200].
Dataset
# samples
# features
# classes
ArtificialF2
222
2
3
Breast Tissue
106
9
6
CTG
2126
22
10
Diabetes
768
8
2
Olive
572
8
9
PB12
608
2
2
Sonar
208
60
2
6.5.2.2
Results
The results we present are reported in detail in [200]. Table 6.25 presents
the parameters of each modular neural network for the results shown in Ta-
ble 6.26. For each type of MNN Table 6.25 shows the number of experts
and, for each of them, the number of hidden neurons and of output neurons
(the number of classes in each cluster). The MLP topologies are denoted
[
d
:
n
h
:
c
], as in Sect. 6.1.1, where
d
is the number of inputs,
n
h
the number
of hidden neurons, and
c
the number of classes.
The presented topologies correspond to the best results of a large number
of experiments with different combinations of the number of neurons in each
module and in the gate.
The results in Tables 6.26 and 6.27 are averages and standard deviations of
the test set error rates for 20 repetitions of each experiment [200]. Table 6.26
shows in bold the result for the winning algorithm; the one achieving the
lowest average error rate. For these datasets EC-MNN is the algorithm that
wins more often. Moreover, EC-MNN outperforms K-MNN for the dataset
ArtificialF2, Breast Tissue, CTG, and Olive in a statistically significant way
(with
t
-test
p
value of 0.037 for ArtificialF2 and
≈
0 for the other ones),
whereas K-MNN outperforms EC-MNN in only one case (Sonar with
t
-test
p
=0
.
018).
