get_groups fails to tell a Random classifier from real classifiers

[andreplima]

Ideally, any comparison framework should be able to discern a Random classifier from real ones by their accuracy results, even if the data does not quite satisfy the requirements pointed out in Demsar, 2006 (N > 10, k > 5).

The data:

classifier_name   BRDT   BRRF     DT  Linear    MLP  Polygrid     RF  Random  Ridge
dataset_name
cancer (mc)      0.908  0.924  0.909   0.930  0.941     0.950  0.841   0.490  0.934
iris (mc)        0.936  0.899  0.943   0.799  0.710     0.981  0.767   0.367  0.846
penguins (mc)    0.932  0.825  0.922   0.988  0.822     0.991  0.778   0.323  0.983
wine (mc)        0.833  0.977  0.892   0.973  0.962     0.994  0.918   0.333  0.990

produces the correct ranking,

Average rank per model, according to accuracy
   Polygrid     1.00
   Ridge        3.25
   Linear       4.00
   BRRF         4.50
   DT           5.00
   BRDT         5.50
   MLP          5.50
   RF           7.25
   Random       9.00

but all classifiers are included in a single group:

Groups of models with statistically indistinguishible performance:
  Group 1: ['BRDT', 'BRRF', 'DT', 'Linear', 'MLP', 'Polygrid', 'RF', 'Random', 'Ridge']