Skip to main content
Figure 6 | Journal of Cheminformatics

Figure 6

From: Using beta binomials to estimate classification uncertainty for ensemble models

Figure 6

Effect of model refinement on error distributions and uncertainty for the validation set. (A) The error distribution for logP3-1a (for which the naive 50% voting threshold was used) is shown in red and the error distribution for logP3-1b (which was refined by shifting the voting threshold to 24.5) is shown in green. The corresponding thresholds are shown as dotted lines and the fitted beta binomials are shown as dashed lines. The distribution of predictions for both models and its fitted beta binomial are represented by the solid and dotted blue line, respectively. (B) The calculated uncertainties and observed error rates for models logP3-1a and logP3-1b are shown in red and green, respectively. Observed and predicted error rates are represented by solid lines and dashed lines. The black dashed line represents the composite uncertainty profile that was obtained by using the refined uncertainty profile below the threshold and the naïve uncertainty profile above it.

Back to article page