Figure 4

Comparison of unit cost, balancing cost, undersampling and oversampling. ROC curves for alleles A1101 (left panel) and B0702 (right panel) comparing the results of trees constructed with the oversampled training set (black curve), the undersampled training set (red curve), and the full training set without training costs, that is, λ₁ = λ₂ = 1 (green curve) and with the balancing training cost, that is, λ₁ = 1 and λ₂ = (1 - p)/p (blue curve). The ROC curves were constructed by varying the threshold used to label a node from 0 to 1 and evaluating its sensitivity and specificity at each threshold.