Initial selection of suitable trees for drawing rules
Given the number of produced trees, it is obvious that many rules will be drawn from these trees. In order to draw conclusions from the rules, it is necessary to separate good trees from other produced models. In doing so, several criteria were defined for choosing trees, as follows:
• trees with over 70% accuracy
• trees with one or more levels (meaning a rule has been drawn)
• trees that includes at least one of the individual variables or exam variables
• trees whose effective variables are amongst either in the individual variable or exam variable groups.
By applying the above criteria, 17 trees were selected in the initial step to go for exam step, amongst which 3 trees were selected. All three had performance evaluation as their target variable and the algorithm used for all was C5.0.
Examining produced trees
Experimental data was used to examine the 17 selected models for the purpose of creating rules. From the final data,