Knowledge Discovery and Data Mining in Medicine - Advanced Techniques in Knowledge Discovery and Data Mining - page 185

Fig.7.6. Modified network structure after structure-level adaptation.
medical experts. However, it is difficult to translate all of their knowledge into
rule sets and prepare a perfect network structure. The newly added neuron may
compensate such a deficit in rule sets.
In the next section, we extract fuzzy rules from the network to give a clear
explanation of the relation between input and output signals.
7.2.4 Extracting Rules from KBANN
Extracting explicit rules from neural networks has attracted a lot of attention
because it has potential for applications to knowledge acquisition from databases.
However, distributed representation of knowledge in the hidden layers makes it
difficult to extract explicit rules from trained networks.
In the case of KBANN, the initial network structure is transformed from
If-Then rules given by experts. Extracting rules from KBANN may be easier than
from usual neural networks. After the structure-level adaptation of a NN, the
initial network structure and the trained one differ only by the number of neurons
in the hidden layers. We can acquire new knowledge by investigating the neurons
that are newly added to the hidden layers.
First, we should investigate the weights connected to the new neurons. If a
connection weight value is smaller than a predetermined threshold value, the
connection to the hidden neuron does not represent new knowledge. On the
contrary, if the connection weight value is larger than the predetermined threshold
value, an If-Then rule can be extracted as new knowledge from the strength of its
weight. The connections between the input neurons and a hidden neuron indicate
the antecedent part of an If-Then rule. The connection between a hidden neuron
and an output neuron indicates the consequent part of an If-Then rule. The strength
of its weight represents the agreement between the antecedent part and the
consequent part represented by the fuzzy membership functions. The fuzzy
membership function uses the fuzzy linguistic values {“very true,” “true,” “rather