1 | %FSELLPC LP (linear programming) classifier |
---|
2 | % |
---|
3 | % [W1,W2,W3] = fsellpc(A,BIAS,TYPE,PARAM) |
---|
4 | % |
---|
5 | % INPUT |
---|
6 | % A Dataset |
---|
7 | % BIAS YES or NO (optional; default: 1 (YES)) |
---|
8 | % TYPE Type of a classifier |
---|
9 | % 'SIMPLE' - the most simple formulation; all features used; PARAM = []; |
---|
10 | % 'STANDARD' - minimization of the training misclassification errors; |
---|
11 | % all features used; PARAM = []; |
---|
12 | % 'C-SPARSE' - feature selection (sparse solution); a formulation similar |
---|
13 | % to the LP_1 SVM; PARAM is a tradeoff parameter. |
---|
14 | % (optional; DEFAULT: 1). |
---|
15 | % 'MU-SPARSE' - feature selection (sparse solution); a formulation similar |
---|
16 | % to the LP_1 SVM, based on the paper of Graepel, Herbrich, Smola etc |
---|
17 | % 'Classification on proximity data with LP-machines'. |
---|
18 | % PARAM is a tradeoff parameter, usually PARAM = 0.05 or 0.1. |
---|
19 | % It is an upper bound on the misclassfied training objects. |
---|
20 | % So, for well separable problems, PARAM = 0.01 or PARAM = 0.02. |
---|
21 | % (optional; DEFAULT: the LOO 1-NN error * 1.3). |
---|
22 | % PARAM Parameter connected to the TYPE, as above |
---|
23 | % |
---|
24 | % OUTPUT |
---|
25 | % W1 LP Classifier in the complete space |
---|
26 | % W2 LP Classifier in a reduced space |
---|
27 | % W3 Feature selection mapping; the indices of chosen features are in +W3. |
---|
28 | % |
---|
29 | % DEFAULTS |
---|
30 | % BIAS = 1 |
---|
31 | % TYPE = 'STANDARD' |
---|
32 | % PARAM = [] |
---|
33 | % |
---|
34 | % DESCRIPTION |
---|
35 | % Classification problem on a N x M data A with LP-machines. Assume a two-class |
---|
36 | % problem. Let DLPC select J features. Then: |
---|
37 | % W1 is an M x 2 classifier in the original space, W2 is an J x 2 classifier |
---|
38 | % in the feature space defined by the J chosen features and W3 is an M x R feature |
---|
39 | % selection such that W1 = W3 * W2. Note that the indices of the selected features |
---|
40 | % can be retrieved by +W3. |
---|
41 | % |
---|
42 | % A linear classifier is built on A: |
---|
43 | % |
---|
44 | % f(A(x,F)) = A(x,F) * w + w0, |
---|
45 | % |
---|
46 | % where F are the features and w are the weights. If BIAS is 1, then w0 is |
---|
47 | % also sought, otherwise it equals 0. This means that the hyperplane is |
---|
48 | % forced to go through the origin. |
---|
49 | % |
---|
50 | % For C-class problems, C classifiers are trained, one against all others. |
---|
51 | % In such a case, only W1 is returned and W3 in now NOT a feature selection, |
---|
52 | % but directly the indices of the selected features. |
---|
53 | % |
---|
54 | % DEFAULT: |
---|
55 | % BIAS = 1 |
---|
56 | % TYPE = 'STANDARD' |
---|
57 | % PARAM = 1 |
---|
58 | % |
---|
59 | |
---|
60 | % Elzbieta Pekalska, Robert P.W. Duin, e.pekalska@tudelft.nl |
---|
61 | % Faculty of Electrical Engineering, Mathematics and Computer Science, |
---|
62 | % Delft University of Technology, The Netherlands. |
---|
63 | |
---|
64 | |
---|
65 | |
---|
66 | |
---|
67 | function [W1,W2,W3] = fsellpc(a,is_w0,type,par,usematlab,prec) |
---|
68 | |
---|
69 | if nargin < 6, prec = 1e-7; end |
---|
70 | if nargin < 5, usematlab = 0; end |
---|
71 | if nargin < 3 | isempty(type), type = 'standard'; end |
---|
72 | if nargin < 4 | isempty(par), |
---|
73 | switch upper(type) |
---|
74 | case {'MU-SPARSE'} |
---|
75 | par = max(0.01,1.3*testk(a,1)); % upperbound error: 1.3 * loo 1-nn error |
---|
76 | case {'C-SPARSE'} |
---|
77 | par = 1; |
---|
78 | case {'SIMPLE','STANDARD'}, |
---|
79 | par = []; |
---|
80 | otherwise |
---|
81 | disp(type) |
---|
82 | error('Wrong type.') |
---|
83 | end |
---|
84 | end |
---|
85 | if nargin < 2 | isempty(is_w0), is_w0 = 1; end |
---|
86 | if nargin < 1 | isempty(a) |
---|
87 | W1 = mapping(mfilename,{is_w0,type,par,usematlab}); |
---|
88 | W1 = setname(W1,'FSELLPC'); |
---|
89 | W2 = []; |
---|
90 | W3 = []; |
---|
91 | return |
---|
92 | end |
---|
93 | |
---|
94 | |
---|
95 | |
---|
96 | if ~isdataset(a), |
---|
97 | error('The first parameter should be a dataset.') |
---|
98 | end |
---|
99 | if ~isnumeric(is_w0) | (is_w0 ~= 0 & is_w0 ~= 1), |
---|
100 | error('The second parameter should be 0 or 1.'); |
---|
101 | end |
---|
102 | |
---|
103 | |
---|
104 | lab = getnlab(a); |
---|
105 | lablist = getlablist(a); |
---|
106 | [m,k,C] = getsize(a); |
---|
107 | |
---|
108 | |
---|
109 | z = (is_w0 > 0); % is the bias used or not? |
---|
110 | |
---|
111 | % This is the status of the optimization procedure. |
---|
112 | % For GLPK, this is the exit code; see GLPKMEX for details. |
---|
113 | % For Matlab LINPROG, if negative then no solution is found. |
---|
114 | |
---|
115 | status = 1; |
---|
116 | |
---|
117 | |
---|
118 | % If more than 2 classes, train the classifier one-against-all. |
---|
119 | if C > 2, |
---|
120 | |
---|
121 | % W1 = mclassc(a,mapping(mfilename,{is_w0,type,par,usematlab})); |
---|
122 | |
---|
123 | W1 = []; |
---|
124 | W2 = []; |
---|
125 | W3 = []; |
---|
126 | N = []; |
---|
127 | for i=1:C |
---|
128 | mlab = 2 - (lab == i); |
---|
129 | aa = dataset(+a,mlab); |
---|
130 | [v1,v2,v3]= fsellpc(aa,is_w0,type,par,usematlab); |
---|
131 | j = +v3; |
---|
132 | if isempty(v1), |
---|
133 | W1 = []; |
---|
134 | W2 = []; |
---|
135 | W3 = []; |
---|
136 | prwarning(1,'No solution found.'); |
---|
137 | return; |
---|
138 | end |
---|
139 | W1 = [W1,setlabels(v1(:,1),lablist(i,:))]; |
---|
140 | W2 = [W2;setlabels(v2(:,1),lablist(i,:))]; |
---|
141 | W3(j) = ones(length(j),1); |
---|
142 | N = [N j]; |
---|
143 | end |
---|
144 | [N1,N2,N3] = unique(N); |
---|
145 | W3 = featsel(k,N1); |
---|
146 | W2 = featsel(length(N1),N3)*W2; |
---|
147 | return |
---|
148 | |
---|
149 | else |
---|
150 | |
---|
151 | Y1 = 3 - 2 * lab; % labels +/-1 |
---|
152 | Y = ones(k,1); |
---|
153 | |
---|
154 | alpha(1:k+1,1) = 0; |
---|
155 | |
---|
156 | aa = +a; |
---|
157 | switch upper(type) |
---|
158 | case {'SIMPLE'} |
---|
159 | f = zeros(k+z,1); |
---|
160 | b = -ones(m,1); |
---|
161 | if is_w0, |
---|
162 | A = -[(Y1*Y').* aa Y1]; |
---|
163 | else |
---|
164 | A = -[(Y1*Y').* aa]; |
---|
165 | end |
---|
166 | [al,fval,status] = linprog(f,A,b); |
---|
167 | alpha(1:k+z) = al; |
---|
168 | |
---|
169 | |
---|
170 | |
---|
171 | case {'STANDARD'} |
---|
172 | L = ones(k,1)/k; |
---|
173 | |
---|
174 | f = [zeros(k+z,1); L]; |
---|
175 | lb = [-Inf .*ones(k+z,1); zeros(k,1)]; |
---|
176 | ub = Inf .* ones(2*k+z,1); |
---|
177 | b = -ones(m,1); |
---|
178 | if is_w0, |
---|
179 | A = -[(Y1*Y').* aa Y1 eye(m,k)]; |
---|
180 | else |
---|
181 | A = -[(Y1*Y').* aa eye(m,k)]; |
---|
182 | end |
---|
183 | [al,fval,ststus] = linprog(f,A,b,[],[],lb,ub); |
---|
184 | alpha(1:k+z) = al(1:k+z); |
---|
185 | |
---|
186 | |
---|
187 | |
---|
188 | case {'C-SPARSE'} |
---|
189 | L = ones(k,1); |
---|
190 | ub = Inf .* ones(3*k+z,1); |
---|
191 | lb = [zeros(2*k,1); -Inf.*ones(z,1); zeros(k,1)]; |
---|
192 | b = -ones(m,1); |
---|
193 | ay = (Y1*Y').* aa; |
---|
194 | if is_w0, |
---|
195 | f = [ones(2*k,1); 0; par*L]; |
---|
196 | A = -[ay -ay Y1 eye(m,k)]; |
---|
197 | else |
---|
198 | f = [ones(2*k,1); par*L]; |
---|
199 | A = -[ay -ay eye(m,k)]; |
---|
200 | end |
---|
201 | if (exist('glpkmex')>0) & (usematlab==0) |
---|
202 | smin = 1; % solve minimum |
---|
203 | ctype = char(ones(m,1)*abs('U')); % sign of inequalities |
---|
204 | vartype = char(ones(3*k+z,1)*abs('C')); % continous variables |
---|
205 | % lpsolver = 1; % Revised Simlex Method |
---|
206 | lpsolver = 2; % Interior Point Method |
---|
207 | params.msglev = 0; % no outputs |
---|
208 | [sss,hostname] = unix('hostname'); |
---|
209 | hostname = hostname(1:end-1); |
---|
210 | if strcmp(hostname,'saturnus') | strcmp(hostname,'polaris') | strcmp(hostname,'neptunus') |
---|
211 | [al,fval,status] = glpkmex_redhat(smin,f,A,b,ctype,lb,ub,vartype,params,lpsolver); |
---|
212 | else |
---|
213 | [al,fval,status] = glpkmex(smin,f,A,b,ctype,lb,ub,vartype,params,lpsolver); |
---|
214 | end |
---|
215 | else |
---|
216 | [al,fval,status] = linprog (f,A,b,[],[],lb,ub); |
---|
217 | end |
---|
218 | alpha(1:k) = al(1:k) - al(k+1:2*k); |
---|
219 | if is_w0, |
---|
220 | alpha(k+1) = al(2*k+1); |
---|
221 | end |
---|
222 | |
---|
223 | |
---|
224 | case {'MU-SPARSE'} |
---|
225 | L = ones(k,1)/k; |
---|
226 | f = [zeros(2*k+z,1); L; -par]; |
---|
227 | ub = Inf .* ones(3*k+1+z,1); |
---|
228 | lb = [zeros(2*k,1); -Inf.*ones(z,1); zeros(k+1,1)]; |
---|
229 | Aeq = [ones(2*k,1); zeros(k+1+z,1)]'; |
---|
230 | beq = 1; |
---|
231 | b = zeros(m,1); |
---|
232 | ay = (Y1*Y').* aa; |
---|
233 | |
---|
234 | if is_w0, |
---|
235 | A = -[ay -ay Y1 eye(m,k) -ones(m,1)]; |
---|
236 | else |
---|
237 | A = -[ay -ay eye(m,k) -ones(m,1)]; |
---|
238 | end |
---|
239 | |
---|
240 | if (exist('glpkmex')>0) & (usematlab==0) |
---|
241 | smin = 1; % solve minimum |
---|
242 | ctype = char([ones(m,1)*abs('U'); 'S']); % sign of inequalities |
---|
243 | vartype = char(ones(3*k+1+z,1)*abs('C')); % continous variables |
---|
244 | % lpsolver = 1; % Revised Simlex Method |
---|
245 | lpsolver = 2; % Interior Point Method |
---|
246 | params.msglev = 0; % no outputs |
---|
247 | [sss,hostname] = unix('hostname'); |
---|
248 | hostname = hostname(1:end-1); |
---|
249 | if strcmp(hostname,'saturnus') | strcmp(hostname,'polaris') | strcmp(hostname,'neptunus') |
---|
250 | [al,fval,status] = glpkmex_redhat(smin,f,[A; Aeq],[b; beq],ctype,lb,ub,vartype,params,lpsolver); |
---|
251 | else |
---|
252 | [al,fval,status] = glpkmex(smin,f,[A; Aeq],[b; beq],ctype,lb,ub,vartype,params,lpsolver); |
---|
253 | end |
---|
254 | else |
---|
255 | [al,fval,status] = linprog(f,A,b,Aeq,beq,lb,ub); |
---|
256 | end |
---|
257 | alpha(1:k) = al(1:k) - al(k+1:2*k); |
---|
258 | if is_w0, |
---|
259 | alpha(k+1) = al(2*k+1); |
---|
260 | end |
---|
261 | |
---|
262 | otherwise |
---|
263 | disp(type) |
---|
264 | error ('Wrong type.'); |
---|
265 | end |
---|
266 | end |
---|
267 | |
---|
268 | % Choose features |
---|
269 | ss = sum(abs(alpha(1:k))); |
---|
270 | J = find(abs(alpha(1:k)) > ss*prec); |
---|
271 | |
---|
272 | if isempty(J) | (status <= 0) | (status > 181 | status == 150), |
---|
273 | prwarning(1,'Fisher classifier is trained.'); |
---|
274 | W1 = fisherc(a); |
---|
275 | W2 = W1; |
---|
276 | W3 = featsel(k,[1:k]); |
---|
277 | else |
---|
278 | W3 = featsel(k,J); |
---|
279 | w = [Y; 1] .* alpha(1:k+1); |
---|
280 | W2 = affine(w(J),w(k+1),a(:,J),lablist,k,2); |
---|
281 | W2 = cnormc(W2,a(:,J)); |
---|
282 | W1 = W3*W2; |
---|
283 | W1 = setname(W1,'FSELLPC'); |
---|
284 | W2 = setname(W2,'FSELLPC'); |
---|
285 | end |
---|
286 | return; |
---|