next up previous
Next: Discussion Up: Experimental Results Previous: Prediction Accuracy

Default Rule Performance

As indicated by Table 31 and 33, the default rule is not used in most cases no matter what definition is used for the closest matching rule. It is heavily used in the cases where only 10% data is used for training on the glass, hepatitis and iris databases since in those cases only the default rule appears in BruteDL's output. For the lenses domain, the default rule is always the only rule in the output since the training set is so small. For the small soybean data set, the coverage of the default rule decreases only when the training set size reaches 50%. When using 20% data for training on hepatitis data, the default rule is also frequently used because the system is not given adequate training examples and not many rules other than the default rule are generated. On the 10% and 20% sized training sets of monks #2 data, not many rules are learned either and thereby the default rule is heavily used. Comparing the results shown in Table 32 and 34, we found that the default rule performs roughly the same in both versions of the BruteDL with closest-matching-rule strategy.

  
Table 31: Mean Coverage (%) of the Default Rule of BruteDL using the Closest Matching Rule (Defined by the Number of Matching Conjuncts) on Different Data Sets

  
Table 32: Mean Prediction Accuracies (%) of the Default Rule of BruteDL Using the Closest Matching Rule (Defined by the Number of Matching conjuncts) on Different Data Sets

  
Table 33: Mean Coverage (%) of the Default Rule of BruteDL Using the Closest Matching Rule (Defined by the Number of Unmatched Conjuncts) on Different Data Sets

  
Table 34: Mean Prediction Accuracies (%) of the Default Rule of BruteDL Using the Closest Matching Rule (Defined by the Number of Unmatched conjuncts) on Different Data Sets



next up previous
Next: Discussion Up: Experimental Results Previous: Prediction Accuracy

Jing Lin
Mon Apr 1 19:35:53 CST 1996