We described the current state of the art in formalising the concept of diversity for classification ensembles, illustrating the clearly that the problem is actually one of quantifying
some form of correlation between non-ordinal predictor outputs. In the process of this we
show a paradox in that though diversity is widely sought after, it is still ill-defined. We
then suggested that an analogue of the bias-variance-covariance decomposition that applies
for 0-1 loss functions may be a useful research direction in this area; this will be further
discussed in the Conclusions chapter.