hssvrtrain.m

Matlab codes for Hidden Sapce Support vector machines

function [alpha,bias, svi, nsv] = hsvrtrain(samples,labels,kernel,kernelparam,C,lossparam,threshold)


  if (nargin ~= 7) % check correct number of arguments
    help svc
  else

    fprintf('Hidden Space Support Vector Classification\n')
    fprintf('__________________________________________\n')
    n = size(samples,1);


    % tolerance for Support Vector Detection
    tol = 1e-5;
   
    % Construct the Kernel matrix
    fprintf('Constructing ...\n');
    K = zeros(n,n);  
    st = cputime;
%     for i=1:n
%           K(:,i) = bsvkernel(samples,samples(i*ones(n,1),:),kernel,kernelparam);
%     end
    K = evalkernel(samples,samples,kernel,kernelparam);

    % Set up the parameters for the Optimisation problem
    
    H = K'*K;
    H = [H -H; -H H];
    c = [(lossparam(1)*ones(n,1) - labels); (lossparam(1)*ones(n,1) + labels)];  
    vlb = zeros(n,1);      % Set the bounds: alphas >= 0
    vub = C*ones(n,1);     %                 alphas <= C
    x0 = zeros(n,1);       % The starting point is [0 0 0   0]
   
    if threshold
        A = [ones(1,n) -ones(1,n)];, b = 0;     % Set the constraint Ax = b
    else
        A = [];, b = []; 
    end

    % Solve the Optimisation Problem
    
    fprintf('Optimising ...\n');
    
    %[alpha lambda how] = qp(H, c, A, b, vlb, vub, x0, threshold);
    [beta,y,z,s,w,flag] = quadsolve(H,c,A,b,C);
%     [beta lambda how] = quadprog(H, c, A, b,[],[], vlb, vub, x0);
%    alpha = smo(H,c,x0,C);
    fprintf('Execution time: %4.1f seconds\n',cputime - st);  
    

            
    % Compute the number of Support Vectors
    %svi = find( beta > tol);
    %nsv = length(svi);
    beta =  beta(1:n) - beta(n+1:2*n);
    alpha = zeros(n,1);
   
    for i = 1 : n
        for j = 1 : n
            alpha(i) = alpha(i) + beta(j)*K(i,j);
        end
    end
    svi = find( abs(alpha) > tol);
    nsv = length(svi);
    fprintf('Support Vectors : %d (%3.1f%%)\n',nsv,100*nsv/n);
    
        % Implicit bias, bias
    bias = 0;

    % Explicit bias, bias 
    if threshold
      % find bias from average of support vectors on margin
      % SVs on margin have alphas: 0 < alpha < C
      svii = find( beta > tol & beta < (C - tol));
      if length(svii) > 0
        bias =  (1/length(svii))*sum(labels(svii) - K(svii,:)*alpha);
      else 
        fprintf('No support vectors on margin - cannot compute bias.\n');
      end
    end
    
  end