[10] | 1 | %KSVO Kernel Support Vector Optimizer
|
---|
| 2 | %
|
---|
| 3 | % [V,J,T,REG] = KSVO(K,Y,C,R)
|
---|
| 4 | %
|
---|
| 5 | % INPUT
|
---|
| 6 | % K Kernel or a square similarity matrix
|
---|
| 7 | % Y Label list consisting of -1/+1
|
---|
| 8 | % C Scalar for weighting the errors (optional; default: 1)
|
---|
| 9 | % R Parameter: -1,0,1,2
|
---|
| 10 | % -1 or 'flip', Changes the kernel by flipping negative eigenvalues to
|
---|
| 11 | % positive
|
---|
| 12 | % 0 or '', Uses the kernel as it is
|
---|
| 13 | % 1 or 'reg', Checks positive definiteness and regularizes the kernel by
|
---|
| 14 | % adding the minimal constant in the form of 10^i to the diagonal
|
---|
| 15 | % 2 or 'lamreg', Checks positive definiteness and regularizes the kernel by
|
---|
| 16 | % adding the minimal constant to the diagonal (equal to the magnitude of
|
---|
| 17 | % the smallest negative eigenvalue)
|
---|
| 18 | % (optional; default: 0, do not change the kernel)
|
---|
| 19 | %
|
---|
| 20 | % OUTPUT
|
---|
| 21 | % V Vector of weights for the support vectors
|
---|
| 22 | % J Index vector pointing to the support vectors
|
---|
| 23 | % T Transformation matrix for the test kernel; to be used in testing
|
---|
| 24 | % REG Regularization parameter added to the diagonal, if used (R=1,2); a vector
|
---|
| 25 | % of eigenvalues of K (R=-1), or -1 if not checked (R=0)
|
---|
| 26 | %
|
---|
| 27 | % DESCRIPTION
|
---|
| 28 | % A low level routine that optimizes the set of support vectors for a two-class
|
---|
| 29 | % classification problem based on the kernel or similarity matrix K. KSVO is called
|
---|
| 30 | % directly from KSVC. The labels Y should indicate the two classes by +1 and -1.
|
---|
| 31 | % Optimization is done by a quadratic programming. If available, the QLD function
|
---|
| 32 | % is used for a positive definite kernel, otherwise an appropriate Matlab routine.
|
---|
| 33 | %
|
---|
| 34 | % SEE ALSO
|
---|
| 35 | % KSVC, SVC, SVO
|
---|
| 36 |
|
---|
| 37 | % Copyright: Elzbieta Pekalska, Robert P.W. Duin, ela.pekalska@googlemail.com
|
---|
| 38 | % Based on SVC.M by D.M.J. Tax, D. de Ridder and R.P.W. Duin
|
---|
| 39 | % Faculty EWI, Delft University of Technology and
|
---|
| 40 | % School of Computer Science, University of Manchester
|
---|
| 41 |
|
---|
| 42 |
|
---|
| 43 | function [v,J,T,reg,err] = ksvo(K,y,C,r)
|
---|
| 44 | prtrace(mfilename);
|
---|
| 45 |
|
---|
| 46 | if (nargin < 4),
|
---|
| 47 | r = 0; % Default: the kernel is used as provided
|
---|
| 48 | end
|
---|
| 49 | if (nargin < 3)
|
---|
| 50 | prwarning(3,'Third parameter not specified, assuming 1.');
|
---|
| 51 | C = 1;
|
---|
| 52 | end
|
---|
| 53 |
|
---|
| 54 | if all(r ~= [-1,0,1,2]),
|
---|
| 55 | error('Wrong parameter R.');
|
---|
| 56 | end
|
---|
| 57 |
|
---|
| 58 | err = 0;
|
---|
| 59 | v = []; % Weights of the SVM
|
---|
| 60 | J = []; % Index of support vectors
|
---|
| 61 | reg = -1; % Regularization on the diagonal or a list of eigenvalues, if used; -1, if not checked
|
---|
| 62 | vmin = 1e-9; % Accuracy to determine when an object becomes the support object
|
---|
| 63 | iacc = -14; % Accuracy of 10^i to determine whether the kernel is pd or not
|
---|
| 64 | T = []; % Matrix to transform the test kernel if r = -1
|
---|
| 65 |
|
---|
| 66 | % Set up variables for the optimization.
|
---|
| 67 | n = size(K,1);
|
---|
| 68 | Ky = (y*y').*K;
|
---|
| 69 | f = -ones(1,n);
|
---|
| 70 | A = y';
|
---|
| 71 | b = 0;
|
---|
| 72 | lb = zeros(n,1);
|
---|
| 73 | ub = repmat(C,n,1);
|
---|
| 74 | p = rand(n,1);
|
---|
| 75 |
|
---|
| 76 |
|
---|
| 77 | if r ~= 0,
|
---|
| 78 | % Find whether the kernel is positive definite.
|
---|
| 79 | i = -20;
|
---|
| 80 | while (pd_check (Ky + (10.0^i) * eye(n)) == 0)
|
---|
| 81 | i = i + 1;
|
---|
| 82 | end
|
---|
| 83 |
|
---|
| 84 | if r == 1,
|
---|
| 85 | if i > iacc, % if i is smaller, then the regularization is within numerical accuracy
|
---|
| 86 | prwarning(1,'K is not positive definite. The diagonal is regularized by 10.0^(%d)',i);
|
---|
| 87 | end
|
---|
| 88 |
|
---|
| 89 | % Make the kernel positive definite by adding a constant to the diagonal.
|
---|
| 90 | reg = 10.0^i;
|
---|
| 91 | Ky = Ky + (10.0^(i)) * eye(n);
|
---|
| 92 |
|
---|
| 93 | elseif r == 2,
|
---|
| 94 | L = preig(Ky);
|
---|
| 95 | L = diag(real(L));
|
---|
| 96 | I = find(L < 0);
|
---|
| 97 | if ~isempty(I),
|
---|
| 98 | % Make the kernel positive definite by adding a constant to the diagonal.
|
---|
| 99 | reg = -1.001*min(L(I));
|
---|
| 100 | Ky = Ky + reg * eye(n);
|
---|
| 101 | pow = round(log10(reg));
|
---|
| 102 | if pow >= 0
|
---|
| 103 | prwarning(1,['K is not positive definite. The diagonal is regularized by %' num2str(pow) '.2f'],reg);
|
---|
| 104 | elseif pow > iacc,
|
---|
| 105 | prwarning(1,['K is not positive definite. The diagonal is regularized by %0.3g'], reg);
|
---|
| 106 | else
|
---|
| 107 | ;
|
---|
| 108 | end
|
---|
| 109 | else
|
---|
| 110 | reg = 0;
|
---|
| 111 | end
|
---|
| 112 |
|
---|
| 113 | else % r = -1,
|
---|
| 114 | % reverse the negative eigenvalues of the kernel K
|
---|
| 115 | if i >= iacc,
|
---|
| 116 | [Q,L] = preig(K);
|
---|
| 117 | Q = real(Q);
|
---|
| 118 | L = diag(real(L));
|
---|
| 119 | reg = L;
|
---|
| 120 | if all(L >= 0),
|
---|
| 121 | T = [];
|
---|
| 122 | else
|
---|
| 123 | K = Q*diag(abs(L))*Q';
|
---|
| 124 | Ky = (y*y').*K;
|
---|
| 125 | T = Q* diag(sign(L)) *Q'; % transformation matrix for the test kernel
|
---|
| 126 | end
|
---|
| 127 | else
|
---|
| 128 | % if i < iacc, then the regularization is within numerical accuracy, so apply it
|
---|
| 129 | Ky = Ky + (10.0^(i)) * eye(n);
|
---|
| 130 | end
|
---|
| 131 | end
|
---|
| 132 | end
|
---|
| 133 |
|
---|
| 134 |
|
---|
| 135 | % Minimization procedure
|
---|
| 136 | % QP minimizes: 0.5 x'*Ky*x + f'x
|
---|
| 137 | % subject to: Ax <= b
|
---|
| 138 | %
|
---|
| 139 |
|
---|
| 140 | done = 0;
|
---|
| 141 | if r ~= 0 & (exist('qld') == 3)
|
---|
| 142 | prwarning(5,'QLD routine is used.')
|
---|
| 143 | v = qld (Ky, f, -A, b, lb, ub, p, 2);
|
---|
| 144 | done = 1;
|
---|
| 145 | end
|
---|
| 146 |
|
---|
| 147 | if r == 0 | ~done,
|
---|
| 148 | if (exist('quadprog') == 2)
|
---|
| 149 | prwarning(5,'Matlab QUADPROG is used.')
|
---|
| 150 | opt = optimset;
|
---|
| 151 | opt.Diagnostics = 'off';
|
---|
| 152 | opt.LargeScale = 'off';
|
---|
| 153 | opt.Display = 'off';
|
---|
| 154 | opt.MaxIter = 500;
|
---|
| 155 | v = quadprog(Ky, f, [], [], A, b, lb, ub,[],opt);
|
---|
| 156 | else
|
---|
| 157 | error(5,'KSVC requires Matlab 6.x')
|
---|
| 158 | end
|
---|
| 159 | end
|
---|
| 160 |
|
---|
| 161 |
|
---|
| 162 |
|
---|
| 163 | % Compute the square pseudo-norm of the weights in the corresponding
|
---|
| 164 | % Hilbert/Krein space. If it is negative (which may happen for a highly
|
---|
| 165 | % indefinite kernel), then the SVM is not proper.
|
---|
| 166 |
|
---|
| 167 | vnorm = v'*Ky*v;
|
---|
| 168 | if vnorm < 0 | abs(vnorm) < 10^iacc,
|
---|
| 169 | prwarning(1,'SVM: ||w||^2_PE < 0. The SVM is not properly defined. Pseudo-Fisher is computed instead.');
|
---|
| 170 | err = 1;
|
---|
| 171 | v = prpinv([K ones(n,1)])*y;
|
---|
| 172 | J = [1:n]';
|
---|
| 173 | return;
|
---|
| 174 | end
|
---|
| 175 |
|
---|
| 176 |
|
---|
| 177 | % Find all support vectors SVs.
|
---|
| 178 | J = find(v > vmin);
|
---|
| 179 |
|
---|
| 180 | % First find the SVs on the boundary
|
---|
| 181 | I = find((v > vmin) & (v < C-vmin));
|
---|
| 182 | Ip = find(y(I) == 1);
|
---|
| 183 | Im = find(y(I) == -1);
|
---|
| 184 |
|
---|
| 185 | if (isempty(v) | isempty(Ip) | isempty(Im))
|
---|
| 186 | prwarning(1,'Quadratic Optimization failed. Pseudo-Fisher is computed instead.');
|
---|
| 187 | err = 1;
|
---|
| 188 | v = prpinv([K ones(n,1)])*y;
|
---|
| 189 | J = [1:n]';
|
---|
| 190 | return;
|
---|
| 191 | end
|
---|
| 192 |
|
---|
| 193 |
|
---|
| 194 | v = y.*v;
|
---|
| 195 |
|
---|
| 196 | % Find the output for the SVs:
|
---|
| 197 | out = y(I)-(K(I,J)*v(J));
|
---|
| 198 | b = mean(out);
|
---|
| 199 | v = [v(J); b];
|
---|
| 200 | return;
|
---|