|
wolffd@0
|
1 function mcmc_post = mcmc_sample_to_hist(sampled_graphs, dags)
|
|
wolffd@0
|
2 % MCMC_SAMPLE_TO_HIST Convert a set of sampled dags into a histogram over dags
|
|
wolffd@0
|
3 % hist = mcmc_sample_to_hist(sampled_graphs, dags)
|
|
wolffd@0
|
4 %
|
|
wolffd@0
|
5 % sampled_graphs{m} is the m'th sampled dag
|
|
wolffd@0
|
6 % dags{i} is the i'th dag in the hypothesis space
|
|
wolffd@0
|
7 % hist(i) = Pr(model i | data)
|
|
wolffd@0
|
8
|
|
wolffd@0
|
9 ndags = length(dags);
|
|
wolffd@0
|
10 nsamples = length(sampled_graphs);
|
|
wolffd@0
|
11 nnodes = length(dags{1});
|
|
wolffd@0
|
12 % sampled_bitv(m, :) is the m'th sampled graph represented as a vector of n^2 bits, computed
|
|
wolffd@0
|
13 % by stacking the columns of the adjacency matrix vertically.
|
|
wolffd@0
|
14 sampled_bitvs = zeros(nsamples, nnodes*nnodes);
|
|
wolffd@0
|
15 for m=1:nsamples
|
|
wolffd@0
|
16 sampled_bitvs(m, :) = sampled_graphs{m}(:)';
|
|
wolffd@0
|
17 end
|
|
wolffd@0
|
18
|
|
wolffd@0
|
19 [ugraphs, I, J] = unique(sampled_bitvs, 'rows'); % each row of ugraphs is a unique bit vector
|
|
wolffd@0
|
20 sampled_indices = subv2ind(2*ones(1,nnodes*nnodes), ugraphs+1);
|
|
wolffd@0
|
21 counts = hist(J, 1:size(ugraphs,1)); % counts(i) = number of times graphs(i,:) occurs in the sample
|
|
wolffd@0
|
22
|
|
wolffd@0
|
23 mcmc_post = zeros(1, ndags);
|
|
wolffd@0
|
24 for i=1:ndags
|
|
wolffd@0
|
25 bitv = dags{i}(:)';
|
|
wolffd@0
|
26 % Find the samples that corresponds to this graph by converting the graphs to bitvectors and
|
|
wolffd@0
|
27 % then to integers.
|
|
wolffd@0
|
28 ndx = subv2ind(2*ones(1,nnodes*nnodes), bitv+1);
|
|
wolffd@0
|
29 locn = find(ndx == sampled_indices);
|
|
wolffd@0
|
30 if ~isempty(locn)
|
|
wolffd@0
|
31 mcmc_post(i) = counts(locn);
|
|
wolffd@0
|
32 end
|
|
wolffd@0
|
33 end
|
|
wolffd@0
|
34 mcmc_post = normalise(mcmc_post);
|
|
wolffd@0
|
35
|
|
wolffd@0
|
36
|
