Table 2 Effects of abstaining, stabilization and discretization on prediction accuracy. The table shows different experimental setups to test the effects of abstaining, stabilization, and the new discretization technique. All the experiments involve 10-fold cross-validation on the ESR data set. Column 1 represents the number of target genes used. Column 2 indicates the number of iterations. Column 3 specifies the discretization technique used. Columns 4 and 5 specify whether abstaining and stabilization were used. Column 6 reports the average test loss. The first and fifth experiments are the original GeneClass algorithm. The fourth and eighth experiments are the Robust GeneClass algorithm with all three algorithmic improvements. Comparing rows 2, 3 and 6, 7 we see that the use of new discretization technique alone improves test loss. Comparing rows 3, 4 and 7, 8 we see that stabilization alone does not have any significant effect on test loss. Comparing rows 1, 2 and 5, 6 we observe that abstaining increases test loss. The combined effect of the three algorithmic improvements results in a small increase in test loss due to the somewhat poorer accuracy of abstaining weak rules; however, abstaining improves the interpretability of the model and significantly speeds up training time.

From: A classification-based framework for predicting and analyzing gene regulatory response