Figure 8. Representation of the multiple comparison of the calibration error for all algorithms using Bergmann-Hommel posthoc correction after a non-parametric Friedman test.

Dashed lines show the average rank of the corresponding algorithms over all datasets. The bold line shows where the null hypothesis of equal distribution means could not be rejected. Two groups are identified: one excluding SVM-lin and one excluding Naive Bayes since these two had significantly different ranks.