hdda_find_seuil.m

高维数据判别分析程序！数据挖掘及图像分析利器！matlab版！

function prms = hdda_find_seuil(Xl,varargin);
% High Dimensionality Discriminant Analysis - This function allows to learn
% the dimensionality parameter from the learning dataset. 
%
% Usage: s = hdda_find_seuil(X,'model','AiBiQiDi','n_it',10);
%
% Author: C. Bouveyron <[e-mail]> - 2004-2006
% 
% Reference: C. Bouveyron, S. Girard and C. Schmid, "High Dimensional Discriminant Analysis",
%            Communications in Statistics, Theory and methods, in press, 2007.

% Global parameters
model = 'AiBiQiDi'; display = 0;
[N,p] = size(Xl.data);
n_it = 25;
n = floor(9 * size(Xl.data,1) / 10);
seuils = [1e-3,5e-3,1e-2:1e-2:19e-2,2e-1:5e-2:4e-1];
dim_max = p-1;
common_d = 0;
    
% PARAMETERS MANAGEMENT
varrem={};
for i=1:2:length(varargin)
    if ~isstr(varargin{i}) | ~exist(varargin{i},'var')
        varrem = varargin(i:end);
    end
    eval([varargin{i} '= varargin{i+1};']);
end

if ~isempty(strmatch(model,strvcat('AijBiQiD', 'AijBQiD', 'AiBiQiD', 'ABiQiD', 'AiBQiD', 'ABQiD','AjBQD', 'ABQD'),'exact'))
    common_d = 1;
    %fprintf('--> the dimensions will be common between classes!\n')
end


% Finding the optimal "seuil" !
fprintf('--> Learning: please wait ');
if ~common_d
    for j=1:length(seuils)
        fprintf('.');
        s = seuils(j);
        for i=1:n_it
            [Xtrn.data,Xtrn.cls,Xtst.data,Xtst.cls] = halfsampling(Xl.data,Xl.cls,n);
            prms = hdda_learn(Xtrn,'model',model,'seuil',s,'display',0);
            res = hdda_classif(prms,Xtst.data);
            taux(j,i) = sum(res == Xtst.cls) / size(Xtst.cls,1);
        end  
    end
    fprintf('\n');
    
    % Display results
    tx = mean(taux'); st = std(taux');[val,ind] = max(tx);
    s_opt = seuils(ind);
    fprintf('--> Optimal threshold: %g\n',s_opt);
    
    % Draw results
    if display, 
        figure, plot(seuils,tx,'*'), hold on,
        plot(seuils(ind),tx(ind),'ro'),
        axis([0 seuils(end) min(tx)-(1-max(tx)) 1])
        for i=1:length(seuils), plot([seuils(i),seuils(i)],[tx(i)-st(i)/2,tx(i)+st(i)/2],':+'), end
    end
    
    % Learn the classifier
    prms = hdda_learn(Xl,'model',model,'seuil',s_opt,'display',0)
    
% Finding the optimal dimension !       
else 
    dims = [1:dim_max];
    for d=1:dim_max
        fprintf('.');
        for i=1:n_it
            [Xtrn.data,Xtrn.cls,Xtst.data,Xtst.cls] = halfsampling(Xl.data,Xl.cls,n);
            prms = hdda_learn(Xtrn,'model',model,'dim',d,'display',0);
            res = hdda_classif(prms,Xtst.data);
            taux(d,i) = sum(res == Xtst.cls) / size(Xtst.cls,1);
        end
    end
    fprintf('\n');
    
    % Display results
    tx = mean(taux'); st = std(taux');[val,ind] = max(tx);
    d_opt = dims(ind);
    fprintf('--> Optimal dimension: %g\n',d_opt);
    
    % Draw results
    if display, 
        figure,
        plot(tx,'-*'), hold on, plot(dims(ind),tx(ind),'ro'),
        axis([0 dim_max+1 min(tx)-(1-max(tx)) 1])
        for i=1:dim_max, plot([i,i],[tx(i)-st(i)/2,tx(i)+st(i)/2],':+'), end
    end
    
    % Learn the classifier
    prms = hdda_learn(Xl,'model',model,'dim',dim_opt,'display',0)
end

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
function [L,cls_l,T,cls_t] = halfsampling(X,cls,n);
% [L,cls_l,T,cls_t] = halfsampling(X,cls,n);
[N,p] = size(X);
ind = randperm(N);
L = X(ind(1:n),:); cls_l = cls(ind(1:n));
T = X(ind(n+1:end),:); cls_t = cls(ind(n+1:end));