Figure 4

Results of the second experiment on the synthetic data. Heat-maps of mean testing accuracy (Acc, the first row), area under the ROC curve (AUC, the second row) and Mathews correlation coefficient (MCC, the third row) of Random forest classifiers trained with either standard bootstrapping (the first column), stratified bootstrapping (the second column) or hierarchical sampling (the third column). Panels capture relation between given performance metrics and bin size/label noise level combinations. Note that values of Mathews correlation coefficient can be as low as -1, which is the reason why upper parts of corresponding plots have uniform coloring (i.e. all values in this region are smaller than zero).