mostFreqClasses = str2double(argv(){4});
numEx = 1;
scenario = 'off';
solver = argv(){3};

% load data
trainD = dlmread(argv(){1}, ',', 1, 0);
testD = dlmread(argv(){2}, ',', 1, 0);
D = [trainD; testD];
X = D(:, 6 : end);
y = D(:, 2);
session = D(:, 2);
z = D(:, 5);

% remove missing values and outliers
active = all(X ~= -1, 2);
active = active & (z > 2) & (z < 3);

% remove least frequent classes
tab = tabulate(y);
[nil, delc] = sort(tab(:, 2), 'descend');
delc = delc(mostFreqClasses + 1 : end);
for c = delc'
  active(y == c) = false;
end

% update data
X = X(active, :);
y = y(active);
session = session(active);
z = z(active);
numClasses = max(y);
N = size(X, 1);
K = size(X, 2);

% normalization
X = zscore(X);

train = 1 : sum(active(1 : size(trainD, 1)));
test = train(end) + 1 : N;

prec = zeros(numEx, 0);
recall = zeros(numEx, 0);
for ex = 1 : numEx
  % NB classifier with multivariate Gaussians
  Py = zeros(numClasses, 1);
  Pxy = zeros(numClasses, K);
  Sigma = zeros(K, K);
  for c = 1 : numClasses
    sub = train(y(train) == c);
    if (~isempty(sub))
      Py(c) = length(sub) / length(train);
      Pxy(c, :) = mean(X(sub, :), 1);
      Sigma = Sigma + Py(c) * cov(X(sub, :));
    end
  end
  
  switch (solver)
    case 'NB'
      % NB inference
      logp = repmat(log(Py)', N, 1);
      for c = 1 : numClasses
        if (Py(c) > 0)
          logp(:, c) = log(Py(c)) + log(mvnpdf(X, Pxy(c, :), Sigma));
        end
      end

    case 'SHT'
      % sequential hypothesis testing
      logp = zeros(N, numClasses);
      for c = 1 : numClasses
        if (Py(c) > 0)
          logp(:, c) = log(mvnpdf(X, Pxy(c, :), Sigma));
        end
      end

      nhyp = zeros(N, 1);
      for i = 1 : N
        if ((i == 1) || (session(i - 1) ~= session(i)))
          logp(i, :) = logp(i, :) + log(Py');
          nhyp(i) = 2;
        else
          logp(i, :) = logp(i, :) + logp(i - 1, :);
          nhyp(i) = nhyp(i - 1) + 1;
        end
      end
  end
  
  % prediction
  [conf, yp] = max(logp, [], 2);
  
  % sum up all but the highest probability
  norm1 = logp - repmat(conf, 1, numClasses);
  norm1((1 : N) + (yp' - 1) * N) = -Inf;
  norm1 = log(sum(exp(norm1), 2));
  
  % evaluation
  for i = 1 : 1000
    th = 3 - i;
    sub = test(norm1(test) < th);
    prec(ex, i) = mean(y(sub) == yp(sub));
    recall(ex, i) = length(sub) / length(test);
  end
end
prec(isnan(prec)) = 1;

hold on;
plot(100 * recall, 100 * prec, '-', ...
  'LineWidth', 1, 'MarkerSize', 4, 'MarkerFaceColor', 'w');
xlabel('Recall [%]');
ylabel('Precision [%]');
hold off;
pause
pr = [recall',prec'];
save pr.mat pr;

% figure;
% A = X - Pxy(y, :);
% for k = 1 : 9
%   subplot(2, 5, k);
%   hist(A(:, k), -5 : 0.1 : 5);
%   h = findobj(gca, 'Type', 'patch');
%   set(h, 'FaceColor', [0.5, 1, 0.5], 'LineStyle', 'none')
%   axis([-3, 3, 0, Inf]);
%   xlabel(sprintf('x_%i - E[x_%i | y]', k, k));
%   set(gca, 'XTick', []);
%   ylabel(sprintf('P(x_%i - E[x_%i | y])', k, k));
%   set(gca, 'YTick', []);
% end