scalem.m

The pattern recognition matlab toolbox

%SCALEM Compute scaling map
% 
%   W = SCALEM(A,T)
%
% INPUT
%   A  Dataset
%   T  Type of s	caling (optional; default: the class priors weighted mean of A is shifted to the origin)
%
% OUTPUT
%   W  Scaling mapping
%
% DESCRIPTION
% Computes a scaling map W, whose type depends on the parameter T:
%
% [], 'c-mean'  - The mean of A is shifted to the origin. Class priors are taken 
%                 into account.
% 'mean'        - The mean of A is shifted to the origin. This is computed for 
%                 the data as given in A, neglecting class priors.
% 'c-variance'  - The mean of A  is shifted to the origin and the average class 
%                 variances (within-scatter) are normalized.  Class priors are 
%                 taken into account.
% 'variance'    - The mean of A is shifted to the origin and the total variances 
%                 of all features are scaled to 1. This is computed for the
%                 data as given in A, neglecting class priors.
% '2-sigma'     - For each feature f, the 2-sigma interval, [f-2*std(f),f+2*std(f)] 
%                 is rescaled to [0,1]. Values outside this domain are clipped.
%                 Class priors are taken into account.
% 'domain'      - The domain for all features in the dataset A is set to [0,1].
%                 This is computed for the data as given in A, neglecting class priors.
%
% The map W may be applied to a new dataset B by B*W.
% 
% SEE ALSO
% MAPPING, DATASET

% Copyright: R.P.W. Duin, [e-mail]
% Faculty EWI, Delft University of Technology
% P.O. Box 5031, 2600 GA Delft, The Netherlands

% $Id: scalem.m,v 1.7 2007/04/21 23:03:40 duin Exp $

function W = scalem(a,t)
	prtrace(mfilename);

	if nargin < 2
		prwarning(4,'No mapping type specified. The class priors weighted mean of A is shifted to the origin.');
		t = ''; 
  elseif isempty(t)
    t = '';
  end
	
  %DXD Make the naming even better:
  dname = 'Scaling mapping';
  switch t
  case 'mean'
	  dname = 'zero-mean';
  case {'','c-mean'}
	  dname = 'zero-mean (+class priors)';
  case 'variance'
	  dname = 'unit-variance';
  case 'c-variance'
	  dname = 'unit-variance (+class priors)';
  case '2-sigma'
	  dname = '0-1 scaling (+-2 standard dev.)';
  case 'domain'
	  dname = '0-1 scaling';
  end
  % No data, return an untrained mapping.
	if (nargin < 1) | isempty(a)
		W = mapping('scalem',{t});
		W = setname(W,dname);
		return;
	end

	[a,c,lablist] = cdats(a);
	[m,k] = size(a);

	switch t
	case {'','c-mean'}
    u = meancov(a);
		u =getprior(a)*(+u);
		s = ones(1,k);
		clip = 0;

	case 'mean'
		u = mean(a,1);
		s = ones(1,k);
		clip = 0;

	case 'c-variance'
		p = getprior(a);
		r = classsizes(a);
		U = meancov(a);
		u = p*(+U);
		s = zeros(1,k);
		for j=1:c
			s = s + p(j)*var(seldat(a,j)*cmapm(+U(j,:),'shift'));
		end
		s = sqrt(s);
		clip = 0;

	case 'variance'
		%[u,G] = meancov(+a);
		%s = sqrt(diag(G))';
  	u = mean(a,1);
		s = std(a,0,1);
		clip = 0;

	case '2-sigma'
		p = getprior(a);
		U = meancov(a);
		u = p*(+U);
		s = zeros(1,k);
		for j=1:c
			s = s + p(j)*var(seldat(a,j));
		end
		s = sqrt(s);
		s = 4*s;
		u = u-0.5*s;
		clip = 1;

	case 'domain'
		mx = max(a,[],1)+eps;
		mn = min(a,[],1)-eps;
		u = mn;
		s = (mx - mn)*(1+eps);
		clip = 0;

	otherwise
		error('Unknown option')
	end

	J = find(s==0);
	s(J) = ones(size(J)); 				% No scaling for zero variances.

	W = cmapm({u,s,clip},'scale');	
	W.size_out = a.featsize;
	W = setlabels(W,getfeatlab(a));
	
return;