LAP works for unlabeled graph classification

Author: JoVo

.DS_Store

@@ -2,39 +2,55 @@
 
 constants = get_constants(adjacency_matrices,class_labels);     % get constants to ease classification code
 
-% get indices for test data
-tst_ind = [2:constants.s0, constants.s0+2:constants.s];
-
 ytrn = [0 1];
-ytst = class_labels(tst_ind);
 
-Atrn(:,:,1) = adjacency_matrices(:,:,1);
-Atrn(:,:,2) = adjacency_matrices(:,:,constants.s0+1);
-Atst        = zeros(constants.n,constants.n,constants.s-2); % pre-allocate memory for Atst
+P       = params;
+P.lnE0  = log(P.E0);
+P.ln1E0 = log(1-P.E0);
+P.lnE1  = log(P.E1);
+P.ln1E1 = log(1-P.E1);
 
+yhat    = zeros(constants.s/2,1);
+correct = zeros(constants.s/2,1);
 
 %% do LAP
 
 if strcmp(alg.names(1),'LAP')
-    k=0;
-    for j=tst_ind
-        k=k+1;
-        if j < constants.s0,
-            B=Atrn(:,:,1);
-        else
-            B=Atrn(:,:,2);
-        end
-        A=adjacency_matrices(:,:,j);
-        [myp,f]=munkres(-B*A');     %Note that sum(sum(-A(myp,:).*B))==f
-        Atst(:,:,k)=A(myp,:);
-    end
 
-    Gtrn = get_constants(Atrn,ytrn);
-    Gtst = get_constants(Atst,ytst);
+    trn0=randperm(constants.s0);
+    trn1=randperm(constants.s1)+constants.s0;
+    tst =randperm(constants.s);
 
-    LLAP = graph_classify_ind_edge(Atrn,Gtrn,alg,Atst,Gtst,params);
-    
-    Lhats.LAP = LLAP.nb;
+    for j=1:constants.s/2
+        
+        Atrn(:,:,1)=adjacency_matrices(:,:,trn0(j));
+        Atrn(:,:,2)=adjacency_matrices(:,:,trn1(j));
+        Atst=adjacency_matrices(:,:,tst(j));
+        ytst=class_labels(tst(j));
+        
+        B=Atst;
+        [myp0,f]=munkres(-Atrn(:,:,1)*B');     %Note that sum(sum(-A(myp,:).*B))==f
+        %         Atst1=B(myp1,:);
+        
+        worked0(j)=norm(Atrn(:,:,1)-B(myp0,:)) < norm(Atrn(:,:,1)-B);
+        
+        B=Atst;
+        [myp1,f]=munkres(-Atrn(:,:,2)*B');     %Note that sum(sum(-A(myp,:).*B))==f
+        %         Atst2=B(myp2,:);
+
+        worked1(1)=norm(Atrn(:,:,2)-B(myp1,:)) < norm(Atrn(:,:,2)-B);
+
+        lik0=sum(sum(Atst(myp0,:).*P.lnE0+(1-Atst(myp0,:)).*P.ln1E0));
+        lik1=sum(sum(Atst(myp1,:).*P.lnE1+(1-Atst(myp1,:)).*P.ln1E1));
+        
+        [~, bar] = sort([lik0, lik1]);
+        yhat(j)=bar(2)-1;
+        correct(j)=(yhat(j)==class_labels(tst(j)));
+                
+    end
+   
+    Lhats.LAP = 1-mean(correct);
+ 
 end
 
 %% do QAP
@@ -46,30 +62,32 @@
 
 inds{alg.QAP_max_iters} = [];
 
-for ii=1:alg.QAP_max_iters
-    
-    k=0;
-    for j=tst_ind
-        k=k+1;
-        if j < constants.s0,
-            B=Atrn(:,:,1);
-        else
-            B=Atrn(:,:,2);
-        end
-        A=adjacency_matrices(:,:,j);
-        %FW approximate solution of QAP
-        [f,myp,x,iter]=sfw(A,-B,ii); % Note that sum(sum(-A.*B(myp,myp)))==f
-        Atst(myp,myp,k)=A;
-        num_iters(ii,j)=iter;
-        % myp=1:constants.n;
-    end    
+
+k=0;
+for j=tst_ind
+    k=k+1;
+    if j < constants.s0,
+        B=Atrn(:,:,1);
+    else
+        B=Atrn(:,:,2);
+    end
+    A=adjacency_matrices(:,:,j);
+    %FW approximate solution of QAP
+    [f,myp,x,iter,fs,myps]=sfw(A,-B,alg.QAP_max_iters,A)
+    %         [f,myp,x,iter]=sfw(A,-B,ii); % Note that sum(sum(-A.*B(myp,myp)))==f
+    Atst(myp,myp,k)=A;
+    num_iters(ii,j)=iter;
+    % myp=1:constants.n;
 
     Gtrn = get_constants(Atrn,ytrn);
     Gtst = get_constants(Atst,ytst);
 
     LQAP = graph_classify_ind_edge(Atrn,Gtrn,alg,Atst,Gtst,params);
 
     Lhats.QAP(ii)   = LQAP.nb;
+    
+    
 end
 
+
 if alg.save, save([alg.datadir alg.fname '_results'],'Lhats','inds'); end

MRI/.DS_Store

@@ -8,7 +8,9 @@
 n   = siz(1);         % # vertices
 constants.s0  = length(constants.y0);   % # of class 0 examples
 constants.s1  = length(constants.y1);   % # of class 1 examples
-constants.s   = siz(3);         % # training samples
+if numel(siz)==2, constants.s=1;
+else constants.s   = siz(3);         % # training samples
+end
 
 dummy=ones(n);
 if nargin==4,

algorithms/.DS_Store

@@ -0,0 +1,116 @@
+
+
+
+clear; clc
+
+alg.datadir = '~/Research/data/graph_sims/unlabeled/';
+alg.figdir  = '~/Research/figs/graph_sims/unlabeled/';
+alg.fname   = 'models';
+alg.save    = 1;
+
+n   = 10;   % # of vertices
+p   = 0.5;  % prob of connection for kidney
+q0  = 0.25; % prob of connection for egg
+q1  = 0.75; % prob of connection for egg
+egg = 1:3;  % vertices in egg    
+
+E0=p*ones(n);   % params in class 0
+E0(egg,egg)=q0; % egg params in class 0
+
+E1=p*ones(n);   % params in class 1
+E1(egg,egg)=q1; % egg params in class 1
+
+params1.E0=E0; params1.E1=E1;
+
+
+n   = 10;   % # of vertices
+
+E0=rand(n);   % params in class 0
+E1=rand(n);   % params in class 1
+
+
+params2.E0=E0; params2.E1=E1;
+
+
+%% plot params
+
+figure(1), clf
+fs=6;
+w=0.3;
+h=0.3;
+s=(1-3*w)/15;
+
+
+emax=max(max([params1.E0(:) params1.E1(:) abs(params1.E0(:)-params1.E1(:))]));
+
+% subplot('Position',[s .1 w h])
+subplot(231)
+image(60*params1.E0/emax)
+colormap('gray')
+title('class 0 mean','fontsize',fs)
+% xlabel('vertices','fontsize',fs)
+% ylabel('vertices','fontsize',fs)
+set(gca,'fontsize',fs,'DataAspectRatio',[1 1 1])
+
+% subplot('Position',[2*s+w .1 w h])
+subplot(232)
+image(60*params1.E1/emax)
+title('class 1 mean','fontsize',fs)
+set(gca,'fontsize',fs)
+set(gca,'fontsize',fs,'DataAspectRatio',[1 1 1])
+
+% subplot('Position',[4*s+2*w .1 w h])
+subplot(233)
+image(60*abs(params1.E0-params1.E1)/emax)
+colormap('gray')
+title('difference','fontsize',fs)
+set(gca,'fontsize',fs)
+set(gca,'fontsize',fs,'DataAspectRatio',[1 1 1])
+
+Position=get(gca,'Position');
+pos(1)=Position(1)+Position(3)+0.01;
+pos(2)=Position(2)+.05;
+pos(3)=0.01;
+pos(4)=Position(4)*.75;
+colorbar('Position',pos,'YTick',[0 15 30 45 60],'YTickLabel',[0 25 50 75 100],'fontsize',fs)
+
+% subplot('Position',[s .1 w h])
+subplot(234)
+image(60*params2.E0/emax)
+colormap('gray')
+title('class 0 mean','fontsize',fs)
+xlabel('vertices','fontsize',fs)
+ylabel('vertices','fontsize',fs)
+set(gca,'fontsize',fs,'DataAspectRatio',[1 1 1])
+
+% subplot('Position',[2*s+w .1 w h])
+subplot(235)
+image(60*params2.E1/emax)
+title('class 1 mean','fontsize',fs)
+set(gca,'fontsize',fs)
+set(gca,'fontsize',fs,'DataAspectRatio',[1 1 1])
+
+% subplot('Position',[4*s+2*w .1 w h])
+subplot(236)
+image(60*abs(params2.E0-params2.E1)/emax)
+colormap('gray')
+title('difference','fontsize',fs)
+set(gca,'fontsize',fs)
+set(gca,'fontsize',fs,'DataAspectRatio',[1 1 1])
+
+Position=get(gca,'Position');
+pos(1)=Position(1)+Position(3)+0.01;
+pos(2)=Position(2)*1.5;
+pos(3)=0.01;
+pos(4)=Position(4)*.75;
+colorbar('Position',pos,'YTick',[0 15 30 45 60],'YTickLabel',[0 25 50 75 100],'fontsize',fs)
+
+if alg.save
+    wh=[3 2];   %width and height
+    set(gcf,'PaperSize',wh,'PaperPosition',[0 0 wh],'Color','w');
+    figname=[alg.figdir alg.fname '_params'];
+    print('-dpdf',figname)
+    print('-deps',figname)
+    saveas(gcf,figname)
+end
+

algorithms/lmnn/.DS_Store

@@ -0,0 +1,32 @@
+clear; clc
+
+alg.datadir = '~/Research/data/graph_sims/unlabeled/';
+alg.figdir  = '~/Research/figs/graph_sims/unlabeled/';
+alg.fname   = 'models';
+alg.save    = 1;
+
+n   = 10;   % # of vertices
+p   = 0.5;  % prob of connection for kidney
+q0  = 0.25; % prob of connection for egg
+q1  = 0.75; % prob of connection for egg
+egg = 1:3;  % vertices in egg    
+
+E0=p*ones(n);   % params in class 0
+E0(egg,egg)=q0; % egg params in class 0
+
+E1=p*ones(n);   % params in class 1
+E1(egg,egg)=q1; % egg params in class 1
+
+params1.E0=E0; params1.E1=E1;
+
+
+n   = 10;   % # of vertices
+
+E0=rand(n);   % params in class 0
+E1=rand(n);   % params in class 1
+
+
+params2.E0=E0; params2.E1=E1;
+
+
+plot_models(params1,params2,alg)

algorithms/lmnn/applypca.m

@@ -11,7 +11,7 @@
 q0  = 0.25; % prob of connection for egg
 q1  = 0.75; % prob of connection for egg
 egg = 1:3;  % vertices in egg    
-S   = 102;  % # of samples
+S   = 20000;  % # of samples
 
 E0=p*ones(n);   % params in class 0
 E0(egg,egg)=q0; % egg params in class 0
@@ -37,7 +37,7 @@
 alg.names = [{'LAP'}; {'QAP'}];
 % alg.names = [{'LAP'}];
 
-alg.QAP_max_iters = 10;
+alg.QAP_max_iters = 15;
 
 % provide indices to classifiers
 alg.nb_ind = 1:n^2;                 % use all edges
@@ -92,6 +92,8 @@
 Lhats.LAP = Lhat.LAP;
 Lhats.QAP = Lhat.QAP;
 
+save([alg.datadir alg.fname],'adjacency_matrices','class_labels','params','alg','Lhat','num_iters')
+
 %% make plots