Limitations
ChiMerge cannot be used to discretize data for un-
supervised learning (clustering) tasks (i.e., where the
examples are not divided into classes), and there does
not appear to be any reasonable way to extend it to do
so. Also, ChiMerge is only attempting to discover first-
order (single attribute) correlations, thus might not
perform correctly when there is a second-order corre-
lation without a corresponding first-order correlation,
which might happen if an attribute only correlates in
the presence of some other condition.