matthiasm@8: function score = score_family(j, ps, node_type, scoring_fn, ns, discrete, data, args)
matthiasm@8: % SCORE_FAMILY_COMPLETE Compute the score of a node and its parents given completely observed data
matthiasm@8: % score = score_family(j, ps, node_type, scoring_fn, ns, discrete, data, args)
matthiasm@8: %
matthiasm@8: % data(i,m) is the value of node i in case m (can be a cell array)
matthiasm@8: % args is a cell array containing optional arguments passed to the constructor,
matthiasm@8: % or is [] if none
matthiasm@8: %
matthiasm@8: % We create a whole Bayes net which only connects parents to node,
matthiasm@8: % where node has a CPD of the specified type (with default parameters).
matthiasm@8: % We then evaluate its score ('bic' or 'bayesian')
matthiasm@8: 
matthiasm@8: % We should use a cache to avoid unnecessary computation.
matthiasm@8: % In particular, log_marginal_prob_node for tabular CPDs calls gammaln
matthiasm@8: % and compute_counts, both of which are slow.
matthiasm@8: 
matthiasm@8: [n ncases] = size(data);
matthiasm@8: dag = zeros(n,n);
matthiasm@8: % SML added to sort ps b/c mk_bnet, learn_params use sorted ps to make
matthiasm@8: % CPTs
matthiasm@8: % Kevin had: if ~isempty(ps), dag(ps, j) = 1; end
matthiasm@8: if ~isempty(ps), dag(ps, j) = 1;, ps = sort(ps);, end
matthiasm@8: 
matthiasm@8: bnet = mk_bnet(dag, ns, 'discrete', discrete);
matthiasm@8: %bnet.CPD{j} = xxx_CPD(bnet, j);
matthiasm@8: %eval(sprintf('bnet.CPD{j} = %s_CPD(bnet, j);', node_type));
matthiasm@8: fname = sprintf('%s_CPD', node_type);
matthiasm@8: %fprintf('score CPD %d\n', j);
matthiasm@8: if isempty(args)
matthiasm@8:   bnet.CPD{j} = feval(fname, bnet, j);
matthiasm@8: else
matthiasm@8:   bnet.CPD{j} = feval(fname, bnet, j, args{:});
matthiasm@8: end
matthiasm@8: switch scoring_fn
matthiasm@8:  case 'bic',
matthiasm@8:   fam = [ps j];
matthiasm@8:   %score = BIC_score_CPD(bnet.CPD{j}, fam, data, ns, bnet.cnodes);
matthiasm@8:   %bnet.CPD{j} = learn_params(bnet.CPD{j}, fam, data, ns, bnet.cnodes);
matthiasm@8:   
matthiasm@8:   % SML 03/16/04 had to special case gaussian b/c generic_CPD/learn_params
matthiasm@8:   % no longer supported because of simple interface to learn_params
matthiasm@8:   % introduced by KPM for tabular nodes below:
matthiasm@8:   % KPM 9 June 04 - tabular nodes have changed back!
matthiasm@8:   if 1 % (isempty(find(j==discrete)))
matthiasm@8:      bnet.CPD{j} = learn_params(bnet.CPD{j},  fam, data, ns, bnet.cnodes);
matthiasm@8:   else 
matthiasm@8:   	bnet.CPD{j} = learn_params(bnet.CPD{j}, data(fam, :));
matthiasm@8:   end
matthiasm@8:   L = log_prob_node(bnet.CPD{j}, data(j,:), data(ps,:));
matthiasm@8:   S = struct(bnet.CPD{j}); % violate object privacy
matthiasm@8:   score = L - 0.5*S.nparams*log(ncases);
matthiasm@8:  case 'bayesian', 
matthiasm@8:   %score = bayesian_score_CPD(bnet.CPD{j}, data(fam, :));
matthiasm@8:   score = log_marg_prob_node(bnet.CPD{j}, data(j,:), data(ps,:));
matthiasm@8:  otherwise,
matthiasm@8:   error(['unrecognized scoring fn ' scoring_fn]);
matthiasm@8: end