Knowledge Discovery and Data Mining in Medicine - Advanced Techniques in Knowledge Discovery and Data Mining - page 178

structure-level adaptation of the NN [7], makes a suitable configuration of a
network on the basis of the observation of change in connection weights and
output activities of hidden neurons during tanning. Using the structure-level
adaptation of the NN, we can give a clear explanation of the relation between
input and output signals. To verify the effectiveness of this method, we develop a
model of the occurrence of hypertension [8].

7.2.1 KBANN

Towell and Shavlik proposed the rule-to-network algorithm [9] shown in Table
7.1. This algorithm is abstracted from the seven-step rules-to-network translation
algorithm, which consists of rewriting, mapping, numbering, adding hidden
neurons, adding input neurons, adding links, and perturbing.

1. Rewriting. The first step of the algorithm transforms the set of approximately
correct rules into a format that clarifies its hierarchical structure and makes it
possible to directly translate the rules into a neural network. If there is more than
one rule for a consequence, then every rule for this consequence is rewritten as
two rules, as shown in Fig. 7.1. B' and B'' are newly created antecedents terms.
Rules do not map from inputs to outputs directly. Sometimes the rules provide
intermediate conclusions, which may be used by other rules to reach the final
conclusion or other intermediate conclusions. The rule sets form the hierarchical
Therefor
Table 7.1. The rule-to-network algorithm.
1. Rewrite rules so that disjunctions are expressed as a rule set that has only one

antecedent.

2. Directly map the rule structure into a neural network.

3. Label neurons in the network according to their “level.”

4. Add hidden neurons to the network at user-specified levels.

5. Add neurons to known input features that are not referred to in the rules.

6. Add links that are not specified by translation between all neurons in topologically

contiguous levels.

7. Perturb the network by adding near-zero random numbers to each of the link

weights.

Fig.7.1. Rewriting rules.