161 records were randomly selected as experimental data to be used for this step. In other words, all 161 selected records were used for training the tree and making the model.
In this step, each of the 17 selected trees in the former step was examined by inputting the experimental data and measuring their accuracy. Because the accuracy of three trees, whose target variable was ‘performance evaluation’ had significantly declined, those trees as well as the ‘performance evaluation’ variable were eliminated before entering the rule creation step. The accuracy levels for the eliminated trees are as show in Table 5.
Therefore the number of selected models for creating rules was reduced to 14. Considering that the models and the rules generated by CART (Ordered) and CART (Towing) algorithms were identical, only one of these was chosen. The test results of the algorithms are presented in Table 6.
It can be concluded that the ‘promotion score’ variable and CART algorithm have resulted in the highest accuracy. This occurred whilst the ‘performance evaluation’ variable was not observed in any of the final models and the models built based on that had resulted in a high error level.
Create rules
After the final selection of trees, rules and trees need to be created. The purpose of creating rules is to evaluate each