|
wolffd@0
|
1 function L = log_marg_lik_complete(bnet, cases, clamped)
|
|
wolffd@0
|
2 % LOG_MARG_LIK_COMPLETE Compute sum_m sum_i log P(x(i,m)| x(pi_i,m)) for a completely observed data set
|
|
wolffd@0
|
3 % L = log_marg_lik_complete(bnet, cases, clamped)
|
|
wolffd@0
|
4 %
|
|
wolffd@0
|
5 % This differs from log_lik_complete because we integrate out the parameters.
|
|
wolffd@0
|
6 % If there is a missing data, you must use an inference engine.
|
|
wolffd@0
|
7 % cases(i,m) is the value assigned to node i in case m.
|
|
wolffd@0
|
8 % (If there are vector-valued nodes, cases should be a cell array.)
|
|
wolffd@0
|
9 % clamped(i,m) = 1 if node i was set by intervention in case m (default: clamped = zeros)
|
|
wolffd@0
|
10 % Clamped nodes contribute a factor of 1.0 to the likelihood.
|
|
wolffd@0
|
11 %
|
|
wolffd@0
|
12 % If there is a single case, clamped is a list of the clamped nodes, not a bit vector.
|
|
wolffd@0
|
13
|
|
wolffd@0
|
14 if iscell(cases), usecell = 1; else usecell = 0; end
|
|
wolffd@0
|
15
|
|
wolffd@0
|
16 n = length(bnet.dag);
|
|
wolffd@0
|
17 ncases = size(cases, 2);
|
|
wolffd@0
|
18 if n ~= size(cases, 1)
|
|
wolffd@0
|
19 error('data should be of size nnodes * ncases');
|
|
wolffd@0
|
20 end
|
|
wolffd@0
|
21
|
|
wolffd@0
|
22 if ncases == 1
|
|
wolffd@0
|
23 if nargin < 3, clamped = []; end
|
|
wolffd@0
|
24 clamp_set = clamped;
|
|
wolffd@0
|
25 clamped = zeros(n,1);
|
|
wolffd@0
|
26 clamped(clamp_set) = 1;
|
|
wolffd@0
|
27 else
|
|
wolffd@0
|
28 if nargin < 3, clamped = zeros(n,ncases); end
|
|
wolffd@0
|
29 end
|
|
wolffd@0
|
30
|
|
wolffd@0
|
31 L = 0;
|
|
wolffd@0
|
32 for i=1:n
|
|
wolffd@0
|
33 ps = parents(bnet.dag, i);
|
|
wolffd@0
|
34 e = bnet.equiv_class(i);
|
|
wolffd@0
|
35 u = find(clamped(i,:)==0);
|
|
wolffd@0
|
36 L = L + log_marg_prob_node(bnet.CPD{e}, cases(i,u), cases(ps,u));
|
|
wolffd@0
|
37 end
|
|
wolffd@0
|
38
|
|
wolffd@0
|
39
|
|
wolffd@0
|
40
|
