wolffd@0: function score = score_family(j, ps, node_type, scoring_fn, ns, discrete, data, args)
wolffd@0: % SCORE_FAMILY_COMPLETE Compute the score of a node and its parents given completely observed data
wolffd@0: % score = score_family(j, ps, node_type, scoring_fn, ns, discrete, data, args)
wolffd@0: %
wolffd@0: % data(i,m) is the value of node i in case m (can be a cell array)
wolffd@0: % args is a cell array containing optional arguments passed to the constructor,
wolffd@0: % or is [] if none
wolffd@0: %
wolffd@0: % We create a whole Bayes net which only connects parents to node,
wolffd@0: % where node has a CPD of the specified type (with default parameters).
wolffd@0: % We then evaluate its score ('bic' or 'bayesian')
wolffd@0: 
wolffd@0: % We should use a cache to avoid unnecessary computation.
wolffd@0: % In particular, log_marginal_prob_node for tabular CPDs calls gammaln
wolffd@0: % and compute_counts, both of which are slow.
wolffd@0: 
wolffd@0: [n ncases] = size(data);
wolffd@0: dag = zeros(n,n);
wolffd@0: % SML added to sort ps b/c mk_bnet, learn_params use sorted ps to make
wolffd@0: % CPTs
wolffd@0: % Kevin had: if ~isempty(ps), dag(ps, j) = 1; end
wolffd@0: if ~isempty(ps), dag(ps, j) = 1;, ps = sort(ps);, end
wolffd@0: 
wolffd@0: bnet = mk_bnet(dag, ns, 'discrete', discrete);
wolffd@0: %bnet.CPD{j} = xxx_CPD(bnet, j);
wolffd@0: %eval(sprintf('bnet.CPD{j} = %s_CPD(bnet, j);', node_type));
wolffd@0: fname = sprintf('%s_CPD', node_type);
wolffd@0: %fprintf('score CPD %d\n', j);
wolffd@0: if isempty(args)
wolffd@0:   bnet.CPD{j} = feval(fname, bnet, j);
wolffd@0: else
wolffd@0:   bnet.CPD{j} = feval(fname, bnet, j, args{:});
wolffd@0: end
wolffd@0: switch scoring_fn
wolffd@0:  case 'bic',
wolffd@0:   fam = [ps j];
wolffd@0:   %score = BIC_score_CPD(bnet.CPD{j}, fam, data, ns, bnet.cnodes);
wolffd@0:   %bnet.CPD{j} = learn_params(bnet.CPD{j}, fam, data, ns, bnet.cnodes);
wolffd@0:   
wolffd@0:   % SML 03/16/04 had to special case gaussian b/c generic_CPD/learn_params
wolffd@0:   % no longer supported because of simple interface to learn_params
wolffd@0:   % introduced by KPM for tabular nodes below:
wolffd@0:   % KPM 9 June 04 - tabular nodes have changed back!
wolffd@0:   if 1 % (isempty(find(j==discrete)))
wolffd@0:      bnet.CPD{j} = learn_params(bnet.CPD{j},  fam, data, ns, bnet.cnodes);
wolffd@0:   else 
wolffd@0:   	bnet.CPD{j} = learn_params(bnet.CPD{j}, data(fam, :));
wolffd@0:   end
wolffd@0:   L = log_prob_node(bnet.CPD{j}, data(j,:), data(ps,:));
wolffd@0:   S = struct(bnet.CPD{j}); % violate object privacy
wolffd@0:   score = L - 0.5*S.nparams*log(ncases);
wolffd@0:  case 'bayesian', 
wolffd@0:   %score = bayesian_score_CPD(bnet.CPD{j}, data(fam, :));
wolffd@0:   score = log_marg_prob_node(bnet.CPD{j}, data(j,:), data(ps,:));
wolffd@0:  otherwise,
wolffd@0:   error(['unrecognized scoring fn ' scoring_fn]);
wolffd@0: end