In the two other folds, the same attribute appears in the same formula with
0.65 and 0.695652 resp. as value on the right side. It seems that the proportion
of nominal attributes plays a role on the performance between Stacking and
Grading: in case there about 2
3
or less of the attributes are nominal, Stacking
works significantly better than Grading.
A tentative explanation for this model may be that a smaller proportion of
nominal attributes makes learning harder for the base-learners, since most of
them are better equipped to handle nominal data. Stacking seems to be able to
compensate for this, since its meta-level data is independent of the base-level
data13 and is processed by MLR which is well equipped to handle numeric data.
However, Grading seems to be unable to compensate for this since its meta-level
data contains just the base-level attributes. Thus its meta learner IBk can be
expected to be susceptible to a smaller proportion of nominal attributes in the
same way as the base learners.