source: distools/private/auclpm.m @ 95

function [w,returnA,r2] = auclpm(x, C, rtype, par, unitnorm, usematlab)
%AUCLPM Find linear mapping with optimized AUC
%
%    W = AUCLPM(X, C, RTYPE, PAR)
%
% Optimize the AUC on dataset X and reg. param. C. This is done by
% finding the weights W for which the ordering of the objects mapped
% onto the line defined by W, is optimal. That means that objects from
% class +1 is always mapped above objects from the -1 class. This
% results in a constraint for each (+1,-1) pair of objects. The number
% of constraints therefore become very large.  The AUC constraints can
% be subsampled in different ways:
%
%               RTYPE     PAR
%         'full',     -   use all constraints
%         'subs',     N   subsample just N constraints
%         'subk',     k   subsample just k*#trainobj. constraints
%         'knn'       k   use only the k nearest neighbors
%         'xval'      N   subsample just N constraints and use the rest to
%                 optimize C (this version can be very slow)
%    'xvalk'     k   subsample k*#trainobj and use remaining constraints
%                 to optimize C
%    'kmeans'    k   use k-means clustering with k=PAR
%    'randk'     subsample objects to get PAR*(Npos+Nneg) constraints
%
%    W = AUCLPM(X, C, RTYPE, PAR, UNITNORM)
%
% Finally, per default the difference vectors are normalized to unit
% length. If you don't like that, set UNITNORM to 0.
%
% Default: RTYPE='subk'
%          PAR  = 1.0;
prtrace(mfilename);

if (nargin < 6)
        usematlab = 0;
end
if (nargin < 5)
        unitnorm = 0;
end
if (nargin < 4)
        par = 1.00;
end
if (nargin < 3)
        rtype = 'subk';
end
if (nargin < 2)
        prwarning(3,'Lambda set to ten');
        C = 10; 
end

if (nargin < 1) | (isempty(x))
        % just a check...
        if ~isa(C,'double')
                error('Please check your input parameters: C should be a double.');
        end
        w = prmapping(mfilename,{C,rtype,par,unitnorm,usematlab});
        w = setname(w,defauclpmname(C,rtype,par,unitnorm));
        return
end

if ~ismapping(C)   % train the mapping

        % just a check...
        if ~isa(C,'double')
                error('Please check your input parameters: C should be a double.');
        end
        % Unpack the dataset.
        islabtype(x,'crisp');
        isvaldset(x,1,2); % at least 1 object per class, 2 classes
        [n,k,c] = getsize(x); 
        % Check some values:
        if par<=0
                error('Parameter ''par'' should be larger than zero');
        end

        if c == 2  % two-class classifier:

                labl = getlablist(x); dim = size(x,2);
                % first create the target values (+1 and -1):
                % make an exception for a one-class or mil dataset:
                tnr = strmatch('target',labl);
                if isempty(tnr) % no target class defined.
                        tnr = strmatch('positive',labl);
                        if isempty(tnr) % no positive class defined.
                                % we just take the first class as target class:
                                tnr = 1;
                        end
                end
                y = 2*(getnlab(x)==tnr) - 1;
                tlab = labl(tnr,:);

                % makes the mapping much faster:
                X = +x; clear x;

                %---create A for optauc
        rstate = rand('state');
                seed = 0;
                [A,Nxi,Aval] = createA(X,y,rtype,par,seed);
        rand('state',rstate);
                if unitnorm
                        % normalize the length of A:
                        lA = sqrt(sum(A.*A,2));
                        lenn0 = find(lA~=0);  % when labels are flipped, terrible
                                              % things can happen
                        A(lenn0,:) = A(lenn0,:)./repmat(lA(lenn0,:),1,size(A,2));
                        if ~isempty(Aval)
                                % also normalize the length of Aval:
                                lA = sqrt(sum(Aval.*Aval,2));
                                lenn0 = find(lA~=0);
                                Aval(lenn0,:) = Aval(lenn0,:)./repmat(lA(lenn0,:),1,size(Aval,2));
                        end
                end
orgA = A;
                % negative should be present for the constraints:
                A = [A -A];
                % take also care for the xi:
                A = [A -speye(Nxi)];
                %A = [A -eye(Nxi)];
                %---create f
                % NO, do this later, maybe we want to optimize it!
                %f = [ones(2*k,1); repmat(C,Nxi,1)];
                %--- generate b
                b = -ones(Nxi,1);   % no zeros, otherwise we get w=0
                    % the constraint is changed here to <=-1
                %---lower bound constraints
                lb = zeros(2*k+Nxi,1);

                % should we run over a range of Cs?
                if ~isempty(Aval)
                        M = 25;
                        xtr = zeros(M,1);
                        xval = zeros(M,1);
                        C = logspace(-3,3,M);
         % run over all the Cs
                        for i=1:length(C)
                                %---create f again:
                                f = [ones(2*k,1); repmat(C(i),Nxi,1)];
                                %---solve linear program
                                if (exist('glpk')>0) & ~usematlab
                                        [z,fmin,status]=glpk(f,A,b,lb,[],repmat('U',Nxi,1),...
                                                repmat('C',size(f,1),1),1);
                                elseif (exist('glpkmex')>0) & ~usematlab
                                        [z,fmin,status]=glpkmex(1,f,A,b,repmat('U',Nxi,1),...
                                                lb,[],repmat('C',size(f,1),1));
                                else
                                        opts = optimset('Display','off','LargeScale','on','Diagnostics','off');
                                        z = linprog(f,A,b,[],[],lb,[],[],opts);
                                end
                                constr = Aval*(z(1:k)-z(k+1:2*k));
                                % the number of satisfied constraints (=AUC:)
                                I = find(constr<-0);
                                if ~isempty(I)
                                        xval(i) = length(I)/size(constr,1);
                                end
                                % the number of satisfied constraints (=AUC:)
                                constr = orgA*(z(1:k)-z(k+1:2*k));
                                I = find(constr<-0);
                                if ~isempty(I)
                                        xtr(i) = length(I)/size(constr,1);
                                end
                        end
if nargout>1
        returnA = xval;
        if nargout>2
                r2 = xtr;
        end
end
                        [minxval,mini] = max(xval);
                        C = C(mini);
                end
                %---create f
                f = [ones(2*k,1); repmat(C,Nxi,1)];
                %---solve linear program
                if (exist('glpkmex')>0) & ~usematlab
                        prwarning(7,'Use glpkmex');
                        param.msglev=0;
                        [z,fmin,status,xtra]=glpkmex(1,f,A,b,repmat('U',Nxi,1),...
                                lb,[],repmat('C',size(f,1),1),param);
                        alpha = []; %xtra.lambda;
                else
                        [z,fmin,exitflag,outp,alpha] = linprog(f,A,b,[],[],lb);
                end

                %---extract parameters
                u = z(1:k); u = u(:);
                v = z(k+1:2*k); v = v(:);
                zeta = z(2*k+1:2*k+Nxi); zeta = zeta(:);
        else
                error('Only a two-class classifier is implemented');
        end
        % now find out how sparse the result is:
        rel = (abs(u-v)>1e-6);
        nr = sum(rel);
        if (nr==0)
                error('None of the features is selected. Please make the C a bit larger.');
        end
        
        % and store the results:
        W.u = u-v; %the ultimate weights
        W.alpha = alpha;
        W.zeta = zeta;
        W.nr = nr;
        W.rel = rel;
        W.C = C;
        w = prmapping(mfilename,'trained',W,tlab,dim,1);
        w = setname(w,defauclpmname(C,rtype,par,unitnorm));
        
else
        % Evaluate the classifier on new data:
        W = getdata(C);
        n = size(x,1);

        % linear classifier:
        out = x*W.u;

        % and put it nicely in a prtools dataset:
        % (I am not really sure what I should output, I decided to give a 1D
        % output:)
        w = setdat(x,out,C);

end
                
return

function cl_name = defauclpmname(C,rtype,par,unitnorm)
% define the correct name:
if unitnorm
        cl_name = sprintf('AUClpm (C=%s, %s, k=%s) 1norm',num2str(C),rtype,num2str(par));
else
        cl_name = sprintf('AUClpm (C=%s, %s, k=%s)',num2str(C),rtype,num2str(par));
end

return