[10] | 1 | %KSVC_NU Kernel Support Vector Classifier: NU algorithm
|
---|
| 2 | %
|
---|
| 3 | % [W,J,C,REG] = KSVC_NU(K,NU,KC,R)
|
---|
| 4 | %
|
---|
| 5 | % INPUT
|
---|
| 6 | % K NxN Kernel dataset
|
---|
| 7 | % NU Regularization parameter (0 < NU < 1);
|
---|
| 8 | % expected fraction of support vectors (optional; default: 0.25)
|
---|
| 9 | % KC Kernel centering, 1/0, (optional; default: 1)
|
---|
| 10 | % R Parameter: -1,0,1,2.
|
---|
| 11 | % -1 or 'flip', Changes the kernel by flipping negative eigenvalues to
|
---|
| 12 | % positive
|
---|
| 13 | % 0 or '', Uses the kernel as it is
|
---|
| 14 | % 1 or 'reg', Checks positive definiteness and regularizes the kernel by
|
---|
| 15 | % adding the minimal constant in the form of 10^i to the diagonal
|
---|
| 16 | % 2 or 'lamreg', Checks positive definiteness and regularizes the kernel by
|
---|
| 17 | % adding the minimal constant to the diagonal (equal to the magnitude of
|
---|
| 18 | % the smallest negative eigenvalue)
|
---|
| 19 | % (optional; default: 0, do not change the kernel)
|
---|
| 20 | %
|
---|
| 21 | % OUTPUT
|
---|
| 22 | % W Mapping: Support Vector Classifier
|
---|
| 23 | % J Object identifiers of support objects
|
---|
| 24 | % C Equivalent C regularization parameter of the KSVC algorithm
|
---|
| 25 | % REG Regularization parameter added to the diagonal, if used (R=1,2); a vector
|
---|
| 26 | % of eigenvalues of K (R=-1), or -1 if not checked (R=0)
|
---|
| 27 | %
|
---|
| 28 | % DEFAULT
|
---|
| 29 | % NU = 1
|
---|
| 30 | % KC = 1
|
---|
| 31 | % R = 0 (use the kernel as it is)
|
---|
| 32 | %
|
---|
| 33 | % DESCRIPTION
|
---|
| 34 | % Determines a support vector machine (SVM) for the kernel K. Quadratic programming
|
---|
| 35 | % formulation is used to solve the problem. K can be positive definite or indefinite.
|
---|
| 36 | % indefinite. J is a list of the indices of the support objects from K. NU belongs
|
---|
| 37 | % to the interval (0,1). The larger NU, the larger class overlap. Default NU = 0.25.
|
---|
| 38 | %
|
---|
| 39 | % For positive semidefinite kernels, NU is bounded from above by NU_MAX, where
|
---|
| 40 | % NU_MAX = (1 - ABS(Lp-Lm)/(Lp+Lm)), where Lp (Lm) is the number of positive
|
---|
| 41 | % (negative) samples. If NU > NU_MAX is supplied to the routine it will be changed to
|
---|
| 42 | % the NU_MAX.
|
---|
| 43 | % If NU is less than some NU_MIN which depends on the overlap between the classes,
|
---|
| 44 | % the algorithm will typically take long time to converge (if at all). So, it is
|
---|
| 45 | % advisable to set NU larger than the expected overlap.
|
---|
| 46 | % The output is rescaled in such a way as if returned by KSVC with the parameter C.
|
---|
| 47 | %
|
---|
| 48 | % If R = 2, then K is regularized by adding the smallest constant possible to the
|
---|
| 49 | % diagonal to make it positive definite. If R = 1, then K is regularized by adding
|
---|
| 50 | % the smallest constant in the form of 10^i to the diagonal. If R = -1, then the
|
---|
| 51 | % eigendecomposition of K is found as K = Q*L*Q' and the negative eigenvalues are
|
---|
| 52 | % flipped, so the SVM is built on K = Q*abs(L)*Q'.
|
---|
| 53 | %
|
---|
| 54 | % IMPORTANT
|
---|
| 55 | % The classifier cannot always be constructed for an indefinite kernel. If the norm
|
---|
| 56 | % of the determined weight vector V in the Krein (pseudo-Euclidean) space induced
|
---|
| 57 | % by K is negative, i.e. if V'*K*V < 0, then the proper SVM cannot be built. This
|
---|
| 58 | % happens when the hyperplane in the Krein space lies 'more' in the negative subspace
|
---|
| 59 | % than in the positive one. This means that the kernel is strongly indefinite.
|
---|
| 60 | % Currently, a warning is given and a pseudo-Fisher classifier is trained, instead.
|
---|
| 61 | % The pseudo-Fisher classifier is also trained in any situation when the quadratic
|
---|
| 62 | % optimization fails to find a solution.
|
---|
| 63 | %
|
---|
| 64 | % REMARKS
|
---|
| 65 | % Note that if D is a symmetric distance/dissimilarity matrix, then K = -D is an
|
---|
| 66 | % (indefinite) kernel. If D.^2 is a square Euclidean distance matrix, then K = -D.^2
|
---|
| 67 | % is a proper (conditionally negative definite) kernel. So, a linear SVM on some
|
---|
| 68 | % data X, based on the kernel K = X*X' is the same classifier as the kernel-based
|
---|
| 69 | % SVM on K = -D(X,X).^2.
|
---|
| 70 |
|
---|
| 71 | % DESCRIPTION
|
---|
| 72 | % Determines a support vector machine (SVM) for the kernel K. Quadratic programming
|
---|
| 73 | % formulation is used
|
---|
| 74 | % The smaller C, e.g.
|
---|
| 75 | % C < 1, the larger class overlap imposed. Optimal C will be different for a regularized
|
---|
| 76 | % kernel and for an indefinite one.
|
---|
| 77 | %
|
---|
| 78 | % SEE ALSO
|
---|
| 79 | % KSVO_NU, KSVO, KSVC, MAPPINGS
|
---|
| 80 | %
|
---|
| 81 | % LITERATURE
|
---|
| 82 | % 1. B.Scholkopf, A. Smola, Learning with kernels, MIT press, 2001,
|
---|
| 83 | % http://www.learning-with-kernels.org/.
|
---|
| 84 | % 2. B. Haasdonk, Feature Space Interpretation of SVMs with Indefinite Kernels.
|
---|
| 85 | % IEEE Trans. on PAMI, 27(4):482-492, 2005.
|
---|
| 86 | % 3. E. Pekalska, P. Paclik, R.P.W. Duin, A Generalized Kernel Approach to
|
---|
| 87 | % Dissimilarity-based Classification, JMLR, vol.2, no.2, 175-211, 2002.
|
---|
| 88 |
|
---|
| 89 |
|
---|
| 90 | % Copyright: Elzbieta Pekalska, Robert P.W. Duin, ela.pekalska@googlemail.com
|
---|
| 91 | % Based on SVC.M by D.M.J. Tax, D. de Ridder and R.P.W. Duin and SVC_NU.M by S.Verzakov
|
---|
| 92 | % Faculty EWI, Delft University of Technology and
|
---|
| 93 | % School of Computer Science, University of Manchester
|
---|
| 94 |
|
---|
| 95 |
|
---|
| 96 | function [W,J,reg,err,K] = ksvc(K,nu,kc,r,no)
|
---|
| 97 | prtrace(mfilename);
|
---|
| 98 |
|
---|
| 99 | if nargin < 5,
|
---|
| 100 | % Multi-class problems are solved by one-vs-all by calling KSVC.
|
---|
| 101 | % no is the class number in such a situation; 0 is the standard 2-class case
|
---|
| 102 | no = 0;
|
---|
| 103 | end
|
---|
| 104 | if nargin < 4,
|
---|
| 105 | r = 0;
|
---|
| 106 | end
|
---|
| 107 | if nargin < 3,
|
---|
| 108 | kc = 1;
|
---|
| 109 | end
|
---|
| 110 | if nargin < 2 | isempty(nu),
|
---|
| 111 | nu = 0.25;
|
---|
| 112 | prwarning(3,'Regularization parameter NU set to 0.25\n');
|
---|
| 113 | end
|
---|
| 114 | if nargin < 1 | isempty(K),
|
---|
| 115 | W = mapping(mfilename,{nu,kc,r});
|
---|
| 116 | W = setname(W,'Kernel Support Vector Classifier (KSVC-NU)');
|
---|
| 117 | return;
|
---|
| 118 | end
|
---|
| 119 |
|
---|
| 120 |
|
---|
| 121 | if all(kc ~= [0 1]),
|
---|
| 122 | error('Wrong KC parameter.');
|
---|
| 123 | end
|
---|
| 124 |
|
---|
| 125 | switch lower(r)
|
---|
| 126 | case {'flip', -1}, r = -1;
|
---|
| 127 | case {'', 0}, r = 0;
|
---|
| 128 | case {'reg', 1}, r = 1;
|
---|
| 129 | case {'lamreg', 2}, r = 2;
|
---|
| 130 | otherwise
|
---|
| 131 | error('Wrong parameter R.');
|
---|
| 132 | end
|
---|
| 133 |
|
---|
| 134 |
|
---|
| 135 | % TRAIN THE CLASSIFIER
|
---|
| 136 | if ~isa(nu,'mapping')
|
---|
| 137 | islabtype(K,'crisp');
|
---|
| 138 | isvaldset(K,1,2); % Expect at least 1 object per class and 2 classes
|
---|
| 139 | [m,k,c] = getsize(K);
|
---|
| 140 | if m ~=k,
|
---|
| 141 | error('The kernel is not a square matrix.');
|
---|
| 142 | end
|
---|
| 143 | nlab = getnlab(K);
|
---|
| 144 |
|
---|
| 145 | % The SVC is basically a two-class classifier.
|
---|
| 146 | % Multi-class problems are trained one-versus-all.
|
---|
| 147 |
|
---|
| 148 | if c == 2 % Two-class classifier
|
---|
| 149 | % Compute the parameters for the optimization:
|
---|
| 150 | y = 3 - 2*nlab; % y = +1/-1
|
---|
| 151 |
|
---|
| 152 | prec = 1e-12;
|
---|
| 153 | if ~issym(K,prec),
|
---|
| 154 | prwarning(1, 'The kernel is not symmetric. The values are averaged out.')
|
---|
| 155 | K = (K+K')/2;
|
---|
| 156 | end
|
---|
| 157 |
|
---|
| 158 | if kc, % Center the data
|
---|
| 159 | me = mean(+K,2)'; % Store the mean value
|
---|
| 160 |
|
---|
| 161 | B = -repmat(1/m,m,m);
|
---|
| 162 | B(1:m+1:end) = B(1:m+1:end)+1; % B = eye(m) - ones(m,m)/m
|
---|
| 163 | K = B*K*B; % K is now centered
|
---|
| 164 | else
|
---|
| 165 | me = [];
|
---|
| 166 | end
|
---|
| 167 | K = (K+K')/2; % K is averaged out as QP solver is sensitive to small asymmetry
|
---|
| 168 |
|
---|
| 169 |
|
---|
| 170 | % Check feasibility of the kernel:
|
---|
| 171 | if r == 0 & dmeans(K,y) < 0,
|
---|
| 172 | if no > 0,
|
---|
| 173 | ss = [num2str(no) '-vs-all '];
|
---|
| 174 | else,
|
---|
| 175 | ss = '';
|
---|
| 176 | end
|
---|
| 177 | prwarning(1,['The kernel is badly indefinite. The ' ss ' SVM cannot be defined. Pseudo-Fisher is computed instead.']);
|
---|
| 178 | err = 1;
|
---|
| 179 | v = prpinv([K ones(m,1)])*y;
|
---|
| 180 | J = [1:m]';
|
---|
| 181 | T = [];
|
---|
| 182 | reg = [];
|
---|
| 183 | else
|
---|
| 184 | % Perform the optimization
|
---|
| 185 | [v,J,T,reg,err] = ksvo_nu(+K,y,nu,r);
|
---|
| 186 | end
|
---|
| 187 |
|
---|
| 188 | % Store the results
|
---|
| 189 | W = mapping(mfilename,'trained',{me,J,T,v},getlablist(K),k,2);
|
---|
| 190 | %W = cnormc(W,a);
|
---|
| 191 | W = setname(W,'Kernel Support Vector Classifier (KSVC-NU)');
|
---|
| 192 | W = setcost(W,K);
|
---|
| 193 | % J = K.ident(J);
|
---|
| 194 |
|
---|
| 195 | else % MULTI-CLASS CLASSIFIER: C > 2
|
---|
| 196 |
|
---|
| 197 | % MCLASSC cannot be used here as we have a kernel K
|
---|
| 198 | W = [];
|
---|
| 199 | J = zeros(m,1);
|
---|
| 200 | lablist = getlablist(K);
|
---|
| 201 |
|
---|
| 202 | for i=1:c
|
---|
| 203 | lab = 2 - (nlab == i); % lab = 1/2
|
---|
| 204 | KK = setlabels(K,lab);
|
---|
| 205 | KK = remclass(KK,0);
|
---|
| 206 | KK = setfeatlab(K,lab);
|
---|
| 207 | if ~isempty(K.prior)
|
---|
| 208 | KK = setprior(KK,[K.prior(i),1-K.prior(i)]');
|
---|
| 209 | end
|
---|
| 210 | [V,j,reg(i),err(i)]= ksvc_nu(KK,nu,kc,r,i);
|
---|
| 211 | W = [W,setlabels(V(:,1),lablist(i,:))];
|
---|
| 212 | J(j) = 1;
|
---|
| 213 | end
|
---|
| 214 | J = find(J);
|
---|
| 215 | end
|
---|
| 216 |
|
---|
| 217 | else
|
---|
| 218 |
|
---|
| 219 | % EXECUTE THE CLASSIFIER
|
---|
| 220 | % nu is an SVM classifier now
|
---|
| 221 | n = size(K,1);
|
---|
| 222 | w = +nu;
|
---|
| 223 |
|
---|
| 224 | % Get the parameters from the classifier
|
---|
| 225 | me = w{1};
|
---|
| 226 | J = w{2};
|
---|
| 227 |
|
---|
| 228 | % The first % The first parameter w{1} stores the mean of the kernel.
|
---|
| 229 | % When it is non-empty, data centering should also be applied
|
---|
| 230 | % to the test kernel.
|
---|
| 231 |
|
---|
| 232 | if ~isempty(me),
|
---|
| 233 | % Center the kernel
|
---|
| 234 | m = length(me);
|
---|
| 235 | B = -repmat(1/m,m,m);
|
---|
| 236 | B(1:m+1:end) = B(1:m+1:end) + 1; % B = eye(m) - ones(m,m)/m
|
---|
| 237 | K = (K - me(ones(n,1),:)) * B;
|
---|
| 238 | end
|
---|
| 239 |
|
---|
| 240 | if ~isempty(w{3}), % this is the transformation that reverses the negative eigenvalues
|
---|
| 241 | K = K*w{3};
|
---|
| 242 | end
|
---|
| 243 |
|
---|
| 244 | % The classifier is stored in w{4}
|
---|
| 245 | % Data is mapped by the kernel, now we just have a linear classifier w*x+b:
|
---|
| 246 | d = [+K(:,J) ones(n,1)] * w{4};
|
---|
| 247 | d = sigm([d -d]);
|
---|
| 248 | W = setdat(K,d,nu);
|
---|
| 249 | end
|
---|
| 250 |
|
---|
| 251 | return
|
---|
| 252 |
|
---|
| 253 |
|
---|
| 254 |
|
---|
| 255 |
|
---|
| 256 | function dm = dmeans(K,y)
|
---|
| 257 | % Computes the square pseudo-Euclidean distance between the
|
---|
| 258 | % means of the two classes. This can be done by using the kernel only.
|
---|
| 259 | % Negative value means that the contribution of the negative
|
---|
| 260 | % subspace in the pseudo-Euclidean sense is too large.
|
---|
| 261 |
|
---|
| 262 | yy = y;
|
---|
| 263 | Z = find (y == 1);
|
---|
| 264 | T = find (y == -1);
|
---|
| 265 | yy(Z) = 1/length(Z);
|
---|
| 266 | yy(T) = -1/length(T);
|
---|
| 267 | dm = yy'*(+K)*yy;
|
---|
| 268 | return;
|
---|