Table 16 and Table 17 present the prediction accuracies and the standard deviation of the accuracies, respectively, of BruteDL that stores all the homogeneous rules regardless of the goodness of the rules for any training examples. Table 18 displays the result of comparing the prediction accuracy of Storing-all-rules BruteDL against the original BruteDL.
As can be seen in Table 18, the mean prediction accuracy given by BruteDL with storing-all-rules is generally lower than that by the original BruteDL on all the data sets. (Although BruteDL with storing-all-rules provides higher prediction accuracy on most of the training sets on hepatitis domain, the accuracies on the other domains are generally lower.) The difference is significant on some training sets such as the 70% sized small soybean data. Theoretically, storing all rules should give more informative predictions for the testing observations than storing only the ``best'' rules and using the default rule when no existing rule matches an observation. It was unexpected that the performance of the system degrades when storing all rules.
Table 16: Mean Prediction Accuracies (%) of BruteDL with Storing-All-Rules
on Different Data Sets
Table 17: Standard Deviation of Prediction Accuracies of BruteDL with
Storing-All-Rules on Different Data Sets
Table 18: Comparison of the Prediction Accuracy of BruteDL with
Storing-all-rules against the Original BruteDL
