comparison toolboxes/FullBNT-1.0.7/bnt/learning/learn_struct_pdag_pc_constrain.m @ 0:e9a9cd732c1e tip

first hg version after svn
author wolffd
date Tue, 10 Feb 2015 15:05:51 +0000
parents
children
comparison
equal deleted inserted replaced
-1:000000000000 0:e9a9cd732c1e
1 function [pdag, G] = dn_learn_struct_pdag_pc_constrain(adj, cond_indep, n, k, varargin)
2 % LEARN_STRUCT_PDAG_PC Learn a partially oriented DAG (pattern) using the PC algorithm
3 % Pdag = learn_struct_pdag_pc_constrain(adj, cond_indep, n, k, ...)
4 %
5 % adj = adjacency matrix learned from dependency network P(i,j) = 1 => i--j; 0 => i j
6 % n is the number of nodes.
7 % k is an optional upper bound on the fan-in (default: n)
8 % cond_indep is a boolean function that will be called as follows:
9 % feval(cond_indep, x, y, S, ...)
10 % where x and y are nodes, and S is a set of nodes (positive integers),
11 % and ... are any optional parameters passed to this function.
12 %
13 %Output
14 % pdag Partially directed graph
15 % G Resulting adjacency graph prior to setting direction arrows
16 %
17 % The output P is an adjacency matrix, in which
18 % P(i,j) = -1 if there is an i->j edge.
19 % P(i,j) = P(j,i) = 1 if there is an undirected edge i <-> j
20 %
21 % The PC algorithm does structure learning assuming all variables are observed.
22 % See Spirtes, Glymour and Scheines, "Causation, Prediction and Search", 1993, p117.
23 % This algorithm may take O(n^k) time if there are n variables and k is the max fan-in,
24 % but this is quicker than the Verma-Pearl IC algorithm, which is always O(n^n).
25 %
26 %% Example
27 %% Given data in a comma separated, filename starting with the variable labels, then the data in rows.
28 %% filename test.txt consists of:
29 %%
30 %% Earthquake,Burglar,Radio,Alarm,Call
31 %% 1,2,2,2,1
32 %% 1,1,2,1,2
33 %% . . .
34 %[CovMatrix, obs, varfields] = CovMat('test.txt',5);
35 %
36 %dn = zeros(5,5);
37 %dn(1,2) = 1; % This was the known Markov blanket of the system that generated test.txt
38 %dn(2,1) = 1;
39 %dn(2,4) = 1;
40 %dn(4,2) = 1;
41 %dn(1,3) = 1;
42 %dn(3,1) = 1;
43 %dn(1,4) = 1;
44 %dn(4,1) = 1;
45 %dn(4,5) = 1;
46 %dn(5,4) = 1;
47 %dn(3,5) = 1; %loop r->c
48 %dn(5,3) = 1; %loop c-r
49 %dn(3,4) = 1;
50 %dn(4,3) = 1;
51 %
52 %max_fan_in = 4;
53 %alpha = 0.05;
54 %
55 %[pdag G] = learn_struct_pdag_pc_constrain(dn,'cond_indep_fisher_z', 5, max_fan_in, CovMatrix, obs, alpha);
56 %%
57 %%
58 %% Gary Bradski, 7/2002 Modified this to take an adjacency matrix from a dependency network.
59
60
61 sep = cell(n,n);
62 ord = 0;
63 done = 0;
64 G = ones(n,n);
65 G=setdiag(G,0);
66
67 while ~done
68 done = 1;
69 [X,Y] = find(G);
70 for i=1:length(X)
71 x = X(i); y = Y(i);
72 % nbrs = mysetdiff(myunion(neighbors(G, x), neighbors(G,y)), [x y]);%parents, children, but not self
73 nbrs = mysetdiff(myunion(neighbors(adj, x), neighbors(adj,y)), [x y]);%parents, children, but not self
74
75 if length(nbrs) >= ord & G(x,y) ~= 0
76 done = 0;
77 SS = subsets(nbrs, ord, ord); % all subsets of size ord
78 for si=1:length(SS)
79 S = SS{si};
80 %if (feval(dsep,x,y,S,adj)) | (feval(cond_indep, x, y, S, varargin{:}))
81 if feval(cond_indep, x, y, S, varargin{:})
82 %if isempty(S)
83 % fprintf('%d indep of %d ', x, y);
84 %else
85 % fprintf('%d indep of %d given ', x, y); fprintf('%d ', S);
86 %end
87 %fprintf('\n');
88
89 % diagnostic
90 %[CI, r] = cond_indep_fisher_z(x, y, S, varargin{:});
91 %fprintf(': r = %6.4f\n', r);
92
93 G(x,y) = 0;
94 G(y,x) = 0;
95 adj(x,y) = 0; %make sure found cond. independencies are marked out
96 adj(y,x) = 0;
97 sep{x,y} = myunion(sep{x,y}, S);
98 sep{y,x} = myunion(sep{y,x}, S);
99 break; % no need to check any more subsets
100 end
101 end
102 end
103 end
104 ord = ord + 1;
105 end
106
107
108
109
110 % Create the minimal pattern,
111 % i.e., the only directed edges are V structures.
112
113 pdag = G;
114 [X, Y] = find(G);
115 % We want to generate all unique triples x,y,z
116 % This code generates x,y,z and z,y,x.
117 for i=1:length(X)
118 x = X(i);
119 y = Y(i);
120 Z = find(G(y,:));
121 Z = mysetdiff(Z, x);
122 for z=Z(:)'
123 if G(x,z)==0 & ~ismember(y, sep{x,z}) & ~ismember(y, sep{z,x})
124 %fprintf('%d -> %d <- %d\n', x, y, z);
125 pdag(x,y) = -1; pdag(y,x) = 0;
126 pdag(z,y) = -1; pdag(y,z) = 0;
127 end
128 end
129 end
130
131 % Convert the minimal pattern to a complete one,
132 % i.e., every directed edge in P is compelled
133 % (must be directed in all Markov equivalent models),
134 % and every undirected edge in P is reversible.
135 % We use the rules of Pearl (2000) p51 (derived in Meek (1995))
136
137 old_pdag = zeros(n);
138 iter = 0;
139 while ~isequal(pdag, old_pdag)
140 iter = iter + 1;
141 old_pdag = pdag;
142 % rule 1
143 [A,B] = find(pdag==-1); % a -> b
144 for i=1:length(A)
145 a = A(i); b = B(i);
146 C = find(pdag(b,:)==1 & G(a,:)==0); % all nodes adj to b but not a
147 if ~isempty(C)
148 pdag(b,C) = -1; pdag(C,b) = 0;
149 %fprintf('rule 1: a=%d->b=%d and b=%d-c=%d implies %d->%d\n', a, b, b, C, b, C);
150 end
151 end
152 % rule 2
153 [A,B] = find(pdag==1); % unoriented a-b edge
154 for i=1:length(A)
155 a = A(i); b = B(i);
156 if any( (pdag(a,:)==-1) & (pdag(:,b)==-1)' );
157 pdag(a,b) = -1; pdag(b,a) = 0;
158 %fprintf('rule 2: %d -> %d\n', a, b);
159 end
160 end
161 % rule 3
162 [A,B] = find(pdag==1); % a-b
163 for i=1:length(A)
164 a = A(i); b = B(i);
165 C = find( (G(a,:)==1) & (pdag(:,b)==-1)' );
166 % C contains nodes c s.t. a-c->ba
167 G2 = setdiag(G(C, C), 1);
168 if any(G2(:)==0) % there are 2 different non adjacent elements of C
169 pdag(a,b) = -1; pdag(b,a) = 0;
170 %fprintf('rule 3: %d -> %d\n', a, b);
171 end
172 end
173 end
174
175
176
177