Table 3.
Statistical comparisons between C4.5 and the two BRL algorithms using BACC
| Comparison | Average | Diff. | t-test | Wilcoxon |
|---|---|---|---|---|
| (t-score) | (Z-score) | |||
| BRL1versus | 63.66 versus | 7.02 | 0.015 | 0.011 |
| C4.5 | 56.64 | (2.624) | (2.555) | |
| BRL1000 versus | 67.65 versus | 11.01 | 0.001 | 0.001 |
| C4.5 | 56.64 | (3.988) | (3.254) | |
| BRL1000 versus | 67.65 versus | 3.99 | 0.050 | 0.029 |
| BRL1 | 63.66 | (2.071) | (2.190) |
We do not compare Ripper and Conjunctive Rule Learner because C4.5 and the two BRL algorithms completely dominate on both performance measures. BRL1 stands for BRL with beam size 1 and BRL1000 represents BRL with beam size 1000. We use both two-sided t-test and two-sided Wilcoxon signed rank test. Those P-values that are significant (≤0.05) are in bold and scores with a positive value favor the first method in the comparison.