The basic idea is to make a version of leave-one-out
cross validation in MCA. This involves writing efficient
algorithms. and thinking about sensible criteria, then
use these criteria for "model-selection".
One possibility is to leave-one-out, then fit in the
object point passively, then predict which categories
it "should" belong to (the closest ones), then see
how well this fits.
This can be used to choose dimensionality, to
choose coding of categories, and so on.
A related question is how well MCA does in recovering
the qualitative information in the indicator matrices.
