Changeset 30 for distools

Timestamp:

05/23/12 13:52:16 (12 years ago)

Author:

bduin

Message:

unsupervised criteria included in protselfd

File:

• distools/protselfd.m (modified) (6 diffs)

distools/protselfd.m

@@ -5 +5 @@
 %
 % INPUT
-%   D     Dataset, square dissimilarity matrix
+%   D     Dataset, dissimilarity matrix
 %   K     Integer, desired number of prototypes
-%   PAR  'LOO' - leave-one-out option. This should be used if
-%          the objects are related to themselves. If D is not square,
-%          it is assumed that the first sets of objects in columns and
-%          rows match.
-%        'ALL' - use all objects (default).
+%   PAR   'SUPER' supervised selection using 1NN error on prototypes.
+%         'LOO' - supervised selection using leave-one-out error estimation.
+%         'MAXDIST' - unsupervised selection minimizing the maximum
+%                     distance to the nearest prototype.
+%         'MEANDIST' - unsupervised selection minimizing the average
+%                      distance to the nearest prototype.
 %
 % OUTPUT
 %   W     Selection mapping ('feature selection')
 %   E     Error stimate as a function of number of selected prototypes
-%         (only reliable for prototype sizes >= class size)
-%   KOPT  Estimate for best size in avoiding peaking
+%         (for supervised selection only reliable for prototype sizes >= class size)
+%   KOPT  Estimate for best size in avoiding peaking 
+%         (supervised selection only)
 %
 % DESCRIPTION
-% This procedure for optimizing the representation set of a 
-% dissimilarity matrix is based on a greedy, forward selection of
-% prototypes using the leave-one-out error estimate of the 1NN rule
-% as a criterion. As this is computed on the given distances in
-% D, the procedure is based on sorting and counting only and is
-% thereby fast. In case K=1 just a single prototype has to be returned,
-% but computing the 1NN error is not possible as all objects are assigned
-% to the same class. In that case the centre object of the largest class
-% will be returned.
-%
-% Note that the search continues untill K prototypes are found.
-% This might be larger than desired due to peaking (curse of
-% dimensionality, overtraining). Therefor an estimate for the
-% optimal number of prototype is returned in KOPT. 
-%
-% The prototype selection may be applied by C = B*W(:,1:KSEL),
-% in which B is a dissimilarity matrix based on the same 
-% representation set as A (e.g. A itself) and C is a resulting 
-% dissimilarity matrix in which the KSEL (e.g. KOPT) best prototypes
-% are selected.
+% This procedure for optimizing the representation set of a dissimilarity 
+% matrix is based on a greedy, forward selection of prototypes. 
+%
+% In case of supervised selection D should be a labeled dataset with 
+% prototype labels stored as feature labels. The 1NN error to the nearest
+% prototype is used as a criterion. In case of leave-one-out error 
+% estimation it is assumed that the first objects in D correspond with the
+% prototypes.
+% 
+% In case K=1 just a single prototype has to be returned, but computing the
+% 1NN error is not possible as all objects are assigned to the same class.
+% In that case the centre object of the largest class will be returned.
+%
+% Note that the search continues untill K prototypes are found. This might
+% be larger than desired due to peaking (overtraining). Therefor an
+% estimate for the optimal number of prototype is returned in KOPT. 
+%
+% The prototype selection may be applied by C = B*W(:,1:KSEL), in which B
+% is a dissimilarity matrix based on the same representation set as A (e.g.
+% A itself) and C is a resulting dissimilarity matrix in which the KSEL
+% (e.g. KOPT) best prototypes are selected.
+%
+% In case of unsupervised selection the maximum or the mean distances to
+% the nearest prototype are minimized. These criteria are the same as used
+% in the KCENTRE and KMEDIOD cluster procedures.
 %
 % REFERENCE
@@ -55 +62 @@
 % 
 
-function [R,e,D,J,nlab,clab] = protselfd(D,ksel,par,J,e,nlab,clab)
-
-if nargin < 2, ksel = []; end
-if nargin < 3 | isempty(par), par = 'all'; end
-
-if nargin < 4 % user call
-  
-  if nargin < 1 | isempty(D)  % allow for D*protselfd([],pars)
-    R = mapping(mfilename,'untrained',{ksel,par});
+function [R,e,D] = protselfd(D,ksel,type)
+
+  if nargin < 2, ksel = []; end
+  if nargin < 3, type = []; end
+
+  if nargin < 1 || isempty(D)  % allow for D*protselfd([],pars)
+    R = mapping(mfilename,'untrained',{ksel,type});
     R = setname(R,'Forward Prototype Sel');
     return
   end
   
+  switch lower(type)
+    case {'loo','LOO','super','SUPER','',''}
+      [R,e,D,J,nlab,clab] = protselfd(D,ksel,type);
+    case {'maxdist','meandist'}
+      R = protselfd_unsuper(D,ksel,type);
+    otherwise
+      error('Unknown selection type')
+  end
+      
+return  
+
+
+function [R,e,D,J,nlab,clab] = protselfd_super_init(D,ksel,par)
+% this routine takes care of the initialisation of supervised selection
+
+  isdataset(D);
   [m,k,c] = getsize(D);
   if isempty(ksel), ksel = k; end
@@ -99 +120 @@
                 % this will be a deep recursive call !!!
                 prwaitbar(ksel,'Forward prototype selection')
-                [R,e,D,J,nlab,clab] = protselfd(D,ksel,R,J,e,nlab,clab);
+                [R,e,D,J,nlab,clab] = protselfd_super(D,ksel,R,J,e,nlab,clab);
                 prwaitbar(0);
         end
@@ -110 +131 @@
   D = floor((Jopt(end)+Jopt(1))/2);
   
-  % done!
-  
-else  % internal call, parameters may have another meaning!
-  
-  R = par;  % prototypes sofar
+return
+  
+function [R,e,D,J,nlab,clab] = protselfd_super(D,ksel,R,J,e,nlab,clab)
+
   [m,k,c] = getsize(D);
   d = +D;
@@ -135 +155 @@
     de = sum(ds);
                 % if better, use it
-    if ee < emin | ((ee == emin) & (de < dmin))
+    if ee < emin || ((ee == emin) && (de < dmin))
       emin = ee;
       jmin = j;
@@ -145 +165 @@
   end
 
-  if emin <= e(r) | 1 % we even continue if emin increases due to peaking
+  if emin <= e(r) || 1 % we even continue if emin increases due to peaking
     e(r+1) = emin;
     R = Rmin;
     if (r+1) < ksel
-                        [R,e,D,J,nlab,clab] = protselfd(D,ksel,R,Jmin,e,nlab,clab);
+                        [R,e,D,J,nlab,clab] = protselfd_super(D,ksel,R,Jmin,e,nlab,clab);
     end
   end
   
+return
+
+%PROTSELFD_UNSUPER Forward prototype selection
+%
+%               N = PROTSELFD_UNSUPER(D,P,CRIT)
+%
+% INPUT
+%   D     Square dissimilarity matrix, zeros on diagonal
+%   P     Number of prototypes to be selected
+%   CRIT  'dist' or 'centre'
+%
+% OUTPUT
+%   N     Indices of selected prototypes
+%
+% DESCRIPTION
+% Sort objects given by square dissim matrix D using a greedy approach
+% such that the maximum NN distance from all objects (prototypes)
+% to the first K: max(min(D(:,N(1:K),[],2)) is minimized.
+%
+% This routines tries to sample the objects such that they are evenly
+% spaced judged from their dissimilarities. This may be used as
+% initialisation in KCENTRES. It works reasonably, but not very good.
+%
+% SEE ALSO
+% KCENTRES
+
+% Copyright: R.P.W. Duin, r.p.w.duin@prtools.org
+% Faculty EWI, Delft University of Technology
+% P.O. Box 5031, 2600 GA Delft, The Netherlands
+
+function N = protselfd_unsuper(d,p,crit)
+
+d = +d;
+[m,k] = size(d);
+if isempty(crit), crit = 'max'; end
+if nargin < 2 || isempty(p), p = k; end
+L = 1:k;
+N = zeros(1,p);
+switch crit
+  case 'maxdist'
+    [~,n] = min(max(d));    % this is the first (central) prototype
+  case 'meandist'
+    [~,n] = min(mean(d));   % this is the first (central) prototype
 end
+e = d(:,n);                 % store here the distances to the nearest prototype (dNNP)
+f = min(d,repmat(e,1,k));   % replace distances that are larger than dNNP by dNNP
+N(1) = n;                   % ranking of selected prototypes
+L(n) = [];                  % candidate prototypes (all not yet selected objects)
+
+for j=2:p                   % extend prototype set
+  switch crit               % select the next prototype out of candidates in L
+    case 'maxdist'
+      [~,n] = min(max(f(:,L))); 
+    case 'meandist'
+      [~,n] = min(mean(f(:,L))); 
+  end
+  e = min([d(:,L(n)) e],[],2);   % update dNNP
+  f = min(d,repmat(e,1,k));      % update replacement of distances that are larger
+                                 % than dNNP by dNNP
+  N(j) = L(n);                   % update list of selected prototypes
+  L(n) = [];                     % update list of candidate prototypes
+end
+