Figure 5

Adjusting the voting threshold for an unbalanced logP data set. The vertical dotted lines indicate the applicable voting thresholds. (A) Training pool distribution of predictions (blue lines) and errors (red lines) for Model logP3-1a, for which the naïve majority rules threshold of 16.5 was used. Observed distributions are represented by the solid lines and fitted beta binomials by dashed lines. (B) Observed and predicted error rate distributions for Model logP3-1a. The red line represents the observed values and the dashed black line represents the predicted profile calculated from the beta binomials shown in panel A. (C) Training pool distribution of predictions (blue) and errors (green) for Model logP3-1b, for which the voting threshold was shifted to 24.5. Observed distributions are represented by solid lines and fitted beta binomials by dashed lines. (D) Observed and predicted training set error rate distributions for Model logP3-1b. The green line represents the observed values and the dashed black line represents the predicted profile calculated from the fitted beta binomials in panel C.