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