Centroid Algorithm

function ypred = predict(Xtrain, ytrain, Xtest)

docs = size(Xtrain,1);
terms = size(Xtrain,2);
classes = size(ytrain,2);

%% Находим классическую TF*IDF матрицу

fprintf('computing TF*IDF... '); tic;

tf = Xtrain;
for i=1:docs
    tf(i,:) = tf(i,:)/sum(tf(i,:));
    %fprintf('%d out of %d tf lines computed\n', i, docs);
end

idf = zeros(1,terms);
for i=1:terms
    ni = sum(Xtrain(:,i)>0);
    if (ni==0)
        idf(i) = 0;
    else
        idf(i) = log(docs/ni);
    end
end

TFIDF = tf;
for i=1:docs
    TFIDF(i,:) = tf(i,:).*idf(1,:);
    %fprintf('%d of %d TFIDF lines computed\n',i,docs);
end

clear tf i idf ni;

toc;

%% Находим центры классов

fprintf('computing centroids... '); tic;

centroids = zeros(classes,terms);
for k=1:classes
    klass = ytrain(:,k)>0.5;
    centroids(k,:) = sum(TFIDF(klass,:))/sum(klass);
    %fprintf('%d out of %d centroids made\n', k, classes);
end
centroids = sparse(centroids);

clear k klass;
clear TFIDF;

toc;

%% Относим вектора к близким по косинусной мере ценрам классов

threshold = 0.825;

fprintf('classifying... '); tic;

ypred = zeros(size(Xtest,1),classes);
dists = zeros(classes,1);
for i=1:size(Xtest,1)
    for k=1:classes
        dists(k) = cossim(Xtest(i,:)',centroids(k,:)');
    end
    ypred(i,dists<threshold)=1;
    [~,I] = min(dists);
    ypred(i,I)=1; % как минимум к одному точно
    %fprintf('%d out of %d processed (%d)\n', i, size(Xtest,1), sum(ypred(i,:)));
end

clear threshold dists i k I;
clear centroids;

toc;

%% Прибираем за собой (имело смысл в скрипте до выделения функции)
clear docs terms classes;