Feature Evaluation by Filter, Wrapper, and Embedded Approaches - Feature Selection for Data and Pattern Recognition - page 33

3.4.2 Machine Learning Techniques Used in Research
In the research performed two distinctively different approaches to classification
were used, namely artificial neural networks simulated by software in multilayer
perceptron (MLP) topology, and decision algorithms induced within Dominance-
based Rough Set Approach (DRSA).
3.4.2.1 ANN Classifier
MLP is a feed-forward, unidirectional network, which at the training phase often
employs backpropagation algorithm [ 11 ]. Firstly for all connections some random
weights are assigned, then they are modified in such way that results in minimising
the difference between the value generated on the network output and the one that is
expected, for all outputs and all training samples. Its popularity MLP owes to good
generalisation properties—once a network learns characteristics of the training set,
it can correctly classify also unknown instances.
Within the first steps of ANN classifier construction that encompasses establishing
network parameters the number of input nodes was set to the number of considered
variables, the number of outputs as corresponding to two recognised classes, and
in the internal structure the number of hidden layers was set to two, with the total
number of neurons in themequal to the number of inputs. Tominimise the influence of
random interconnection weights on the training phase the multi-starting procedure
was employed, with repetitive learning and calculations of median, minimal, and
maximal performance.
3.4.2.2 DRSA Classifier
In rough set approach the objects of the universe are perceived through granules of
knowledge [ 28 ]. In classical version, invented by Pawlak [ 29 ], these granules are
equivalence classes of instances that cannot be discerned basing on values of con-
sidered conditional attributes. Classical Rough Set Approach (CRSA) enables only
nominal classification, which can be insufficient for multicriteria decision making
[ 10 ]. Replacing indiscernibility relation with dominance and observing weak pref-
erence orderings in value sets of attributes allows for ordinal classification and gives
Dominance-based Rough Set Approach [ 12 ].
DRSA approximates dominance cones—upward and downward unions of deci-
sion classes and induces decision rules that form rule-based classifiers. The advantage
of this approach is enhanced understanding of approximated information, as rules
explicitly specify the conditions that need to be met for some object to be classified
to the specific class (to be precise, in this case to the union of classes). There are
many algorithms for rule induction and depending on them the sets of generated
rules can greatly vary [ 5 , 32 ], not only with respect to their cardinalities, but, more