%% ct_intro: Introduction of clustering by ClusterTools
% 
% <http://37steps.com/prtools/ PRTools> and <http://37steps.com/clustertools/ 
% ClusterTools> should be in the path
% 
% <http://37steps.com/clustertools/clustertools-examples/ A review of all
% Clustertools examples>.
% 
% <http://37steps.com/exam_ct/ct_intro/ct_intro.m Download 
% the m-file from here>.

%% Prepare environment
prtime(10)                   % restrict iterative optimisation to 10s
delfigs                      % delete existing figures
randreset;                   % takes care of reproducability
prwarning(2)                 % show warnings

%% Define a dataset
% we use some standard routines to create 8 two-dimensional clusters. After 
% creation the label information is removed.

m = 1000;
a = gendatclust2(m);
x = +a; % remove labels
scattern(x); axis equal

%%
% This dataset consist of:
% 
% * two straight, noisy lines
% * two noisy circles inside each other
% * two spherical normal distributions
% * two elongated non-normal distributions.

%% Find 8 clusters by five routines
% * Run K-Means, Single Linkage, Complete Linkage, Mode Seeking, Exemplar
% * Find 8 clusters.
% * Compare with ground truth.
% * Display active learning error, i.e. the error if the entire clusters 
% are assigned to the prototype class. 
%
% First a separate coding per cluster routine is shown. 
% In the bottom experiment a more compact coding is used.

figure;
labt = getnlab(a);
subplot(3,2,1); scatn(labt,x,'True Labels'); 
markcols(1);
axis equal; drawnow

labk = clustk(x,8); e = clusteval(labk,a,'actl');
title = ['K-Means ' num2str(e,'%5.2f')];
subplot(3,2,2); scatn(labk,x,title); 
markcols(1);
axis equal; drawnow

labhs = clusth(x,8,'s'); e = clusteval(labhs,a,'actl');
title = ['Single Linkage ' num2str(e,'%5.2f')];
subplot(3,2,3); scatn(labhs,x,title); 
markcols(1);
axis equal; drawnow

labhc = clusth(x,8,'c'); e = clusteval(labhc,a,'actl');
title = ['Complete Linkage ' num2str(e,'%5.2f')];
subplot(3,2,4); scatn(labhc,x,title); 
markcols(1);
axis equal; drawnow

labm = clustm(x,8); e = clusteval(labm,a,'actl');
title = ['Mode Seeking ' num2str(e,'%5.2f')];
subplot(3,2,5); scatn(labm,x,title); 
markcols(1);
axis equal; drawnow

labe = cluste(x,8); e = clusteval(labe,a,'actl');
title = ['Exemplar ' num2str(e,'%5.2f')];
subplot(3,2,6); scatn(labe,x,title);
markcols(1);
axis equal; drawnow
clear title
%%
% K-Means, Complete Linkage and Mode Seeking result in similar clusters,
% that are about spherical and of the same size. Single Linkage shows an
% interesting result as it finds the two bars as well as  the two circles
% as single clusters, but it merges the four Gaussian blobs. In addition
% there are three single object clusters. Exemplar performs badly. Note
% that it generated a warning as its optimisation needed too much time.

%% Find 20 clusters by five routines
% * Run K-Means, Single Linkage, Complete Linkage, Mode Seeking, Exemplar
% * Find 20 clusters.
% * Compare with ground truth.
% * Display active learning error, i.e. the error if the entire clusters 
% are assigned to the prototype class. 
%
% Coding here is more comapact than above. It uses the fact that the 
% cluster routines are programmed as PRTools mappings.

figure; 
k = 20;
w = {clustk(k) clusth(k,'s') clusth(k,'c') clustm(k) cluste(k)};
names = {'True Labels','K-Means','Single Linkage','Complete Linkage', ...
         'Mode Seeking','Exemplar'};
for i=1:6
  % for i==1 a trick is used to get ClusterTools labels from dataset
  if i == 1, lab = getnlab(a)*clust2proto(a); e = 0;
  else lab = x*w{i-1}; e = clusteval(lab,a,'actl'); end
  title = [names{i} ' ' num2str(e,'%5.2f')];
  subplot(3,2,i); scatn(lab,x,title); markcols(1);
  axis equal; drawnow
end

%%
% As there are more clusters than in the first experiment (20 instead of 
% 8) for more objects the true labels are used. These  are assigned to all
% objects in the corresponding cluster, resulting in a better perforemance.
% Exemplar now performs well. Frequently it is good (but very time 
% consuming) for larger datasets and more clusters.