A score like log-likelihood ratio can be used to establish the sparsity pattern for an association matrix whose non-zero elements are a more probabilistically defensible measure of similar occurrence.
This matrix can be used for clustering using, inter alia, spectral methods, agglomerative clustering or coloring methods. All of these require relatively few passes through the association matrix. On Wed, Aug 20, 2008 at 1:09 AM, sej <[EMAIL PROTECTED]> wrote: > > I'm not quite sure I understand your suggestion. Co-occurrence modeling > would be limited to finding the most interesting pairs. If you have a > follow up link to elaborate on item sets that extend beyond pairs > (cardinality > 2), that would be helpful. >
