Accuracy Performance Clause Samples

Accuracy Performance. Table 1 presents the mean accuracy and the standard deviation over five runs of 10 fold cross-validation using C4.5 algorithm as the base classifier. The shaded boxes represent cases where the difference between CAP-Best-First-Kappa and the corresponding method is statistically significant with 95% confidence using t-test. A win-loss-tie summarization based on mean value and t test is attached at the bottom of the table. Generally Kappa measure slightly outperforms symmetrical uncertainty and GA search outperforms Best First search. Using adjusted ▇▇▇▇▇▇▇▇ test the null hypothesis that all pruning methods perform the same over multiple data sets and the observed differences are merely random has been reject with FF (8, 232) =10.05, p < 0.001. We proceed with a post-hoc Bonferroni-▇▇▇▇ test using CAP-Best-First-Kappa as the controlled method. We concluded that all variations of CAP method perform almost the same. Still CAP-Best-First-Kappa significantly outperforms CAP-Best-First-Symmetrical-Uncertainty with z= 2.26, p<0.05. The accuracy of the proposed pruned ensemble is similar to the accuracy of the original ensemble (no pruning). CAP-Best-First-Kappa significantly outperforms Kappa Ranking with z=4.14, p<0.001. Moreover CAP-GA-Kappa significantly outperforms GASEN-b with z=2.32, p<0.01. This indicates that the using collective merit measure is more accurate than using the wrapper approach when GA search strategy is used. This conclusion is not expected, because wrapper approach is generally considered to be slow but accurate mean to direct the search process. Table 2 presents the mean accuracy and the standard deviation over five runs of 10 fold cross-validation using Decision ▇▇▇▇▇ algorithm as the base classifier. The shaded boxes represent cases where the difference between CAP-Best-First-Kappa and the corresponding method is statistically significant with 95% confidence using t-test. A win-loss-tie summarization based on mean value and t test is attached at the bottom of the table. All pruning methods slightly reduce the accuracy performance when compared to the No-Pruning results. Nevertheless CAP-Best-First-Kappa significantly outperforms No-Pruning in the Wine dataset. Generally Kappa measure slightly outperforms symmetrical uncertainty. Using adjusted ▇▇▇▇▇▇▇▇ test the null hypothesis that all pruning methods perform the same over multiple data sets and the observed differences are merely random has been reject with FF (8, 232) = 7.168, p < 0.00...