Current implementation just one-hot-encodes categorical variables, giving them in principle more weight for the computation of the correlation difference.
Correlation differences from columns created by one-hot-encoding should be downsclaed based on the number of categories for that column.