TADW.m

clear all
data = 'citeseer';  % 'cora' or 'citeseer' or 'wiki'
%% Parameters
k=80;  % dimensions of matrix W and H
lambda = 0.2;   % regularization parameter
textRank = 200; % dimension of text feature, used when preprocessing
train_ratio = 0.1;  % training ratio for SVM classifier
C=5;    % parameter for SVM, you may get a slightly better result if you tune it carefully
if strcmp(data,'wiki')==1
    C=15;
end

loss = 10;  % square loss, don't change
load([data,'/graph.txt']);

%% Reduce the dimension of text feature from TFIDF matrix
display ('Preprocessing text feature...');
if strcmp(data,'wiki')==0   % compute TFIDF matrix for cora and citeseer datasets
    load([data,'/feature.txt']);
    numOfNode = size(feature,1);
    for i=1:size(feature,2)
        if (nnz(feature(:,i)) > 0)
            feature(:,i) = feature(:,i)*log(numOfNode/nnz(feature(:,i)));
        end
    end
    [U,S,V] = svds(feature, textRank);
    text_feature = U * S;
    clear U S V
else
    load([data,'/tfidf.txt']);
    tfidf(:,1) = tfidf(:,1) + 1;
    tfidf(:,2) = tfidf(:,2) + 1;
    tfidf = sparse(tfidf(:,1),tfidf(:,2),tfidf(:,3),max(tfidf(:,1)),max(tfidf(:,2)));
    [U,S,V] = svds(tfidf, textRank);
    text_feature = U * S;
    clear U S V
end
    

numOfNode = size(text_feature,1);
%% Build matrix M=(A+A*A)/2 
display ('Computing matrix M...');
graph(:,1) = graph(:,1) + ones(size(graph(:,1)));
graph(:,2) = graph(:,2) + ones(size(graph(:,2)));
graph = [graph;graph(:,2) graph(:,1)];
graph = sparse(graph(:,1),graph(:,2),ones(size(graph(:,1))),numOfNode,numOfNode);

Features = speye(numOfNode);
ColFeatures = text_feature;
for i=1:size(graph,1)
  if (norm(graph(i,:))>0)
      graph(i,:) = graph(i,:)/nnz(graph(i,:)) ;
  end
end
g2 = graph * graph;
graph = graph + g2;
graph = graph ./ 2 ;

for i=1:size(ColFeatures,2)
    if (norm(ColFeatures(:,i))>0)
        ColFeatures(:,i) = ColFeatures(:,i)/norm(ColFeatures(:,i));
    end
end

display ('Learning Parameters...');
[W,H,time] = train_mf(sparse(graph), sparse(Features), sparse(ColFeatures), [' -l ' num2str(lambda) ' -k ' num2str(k) ' -t 10' ' -s ' num2str(loss)]); 
acc=svmTest(W,H,text_feature,data,train_ratio,C);
acc
% representation learned by TADW = [W' text_feature*H'];