EXPERIMENTS WITH BRUTEDL

Segal et al [10] presented a comparison of the performance of BruteDL against C4 [7]. The results of the comparisons show that BruteDL's performance is almost as good as C4's performance on some of the data sets from the UCI repository [5], and better on the data set MONKS #1 from the MONK's competition [11]. A summary of the characteristics of the different data sets is given in Table 1. No results have been given regarding the performance of the default rule strategy itself. To answer how good the default rule strategy is and how big the potential improvement we can get by modifying the default strategy, we have run experiments with BruteDL. A description of the experimental design and results are given in this chapter.

  
Table 1: Summary of the Selected Data Sets

• Experimental Design
• Experimental Results
  • Prediction Accuracy
  • Default Rule Performance
• Discussion
  • Training Set Size Influences Default Rule Coverage
  • Training Set Size and Class Distribution Influence Default Rule Accuracy
• Chapter Summary