Results of the first experiment on the synthetic data. Evolution of testing accuracy, area under the ROC curve and Mathews correlation coefficient of Random forest classifiers trained with three sampling schemes and under the increasing size of partitions of the grain-defining feature (Panel A). Vertical bars indicate empirical 95% confidence intervals. Panel B displays corresponding change in Random forest feature importance metrics for all features of the synthetic data set under the standard bootstrapping. This metric captures an increase in out-of-bag classification error when the values of given feature are shuffled.