Mercurial > hg > camir-aes2014

annotate toolboxes/FullBNT-1.0.7/bnt/learning/bic_score_family.m @ 0:e9a9cd732c1e tip

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
first hg version after svn
author wolffd
date Tue, 10 Feb 2015 15:05:51 +0000
parents
children
rev   line source
wolffd@0 1 function [S, LL] = bic_score(counts, CPT, ncases)
wolffd@0 2 % BIC_SCORE Bayesian Information Criterion score for a single family
wolffd@0 3 % [S, LL] = bic_score(counts, CPT, ncases)
wolffd@0 4 %
wolffd@0 5 % S is a large sample approximation to the log marginal likelihood,
wolffd@0 6 % which can be computed using dirichlet_score.
wolffd@0 7 %
wolffd@0 8 % S = \log [ prod_j _prod_k theta_ijk ^ N_ijk ] - 0.5*d*log(ncases)
wolffd@0 9 % where counts encode N_ijk, theta_ijk is the MLE comptued from counts,
wolffd@0 10 % and d is the num of free parameters.
wolffd@0 11
wolffd@0 12 %CPT = mk_stochastic(counts);
wolffd@0 13 tiny = exp(-700);
wolffd@0 14 LL = sum(log(CPT(:) + tiny) .* counts(:));
wolffd@0 15 % CPT(i) = 0 iff counts(i) = 0 so it is okay to add tiny
wolffd@0 16
wolffd@0 17 ns = mysize(counts);
wolffd@0 18 ns_ps = ns(1:end-1);
wolffd@0 19 ns_self = ns(end);
wolffd@0 20 nparams = prod([ns_ps (ns_self-1)]);
wolffd@0 21 % sum-to-1 constraint reduces the effective num. vals of the node by 1
wolffd@0 22
wolffd@0 23 S = LL - 0.5*nparams*log(ncases);