ClusterTools Contents

ClusterTools User Guide



Classify cluster objects on the basis of (some) class labels

including confidences by combining clustering levels.


 LABC Index array, size [M,K], indices of cluster prototypes for M objects in K clusterings.
 LABT Vector of M elements with true object labels. Preferably  numbers between 1 and C, if C is the number of classes.  Unknowns should be given by 0.

 LABOUT Class labels assigned to all objects based on the maximum  confidence.
 CONF Array, size MxC, with class confidences for all objects.


Initially all labeled objects are given a confidence, one for the  particular class and zero for all other classes. Next, from the highest  cluster level (smallest number of clusters) downwards, all class  confidences of all objects are updated by adding the average class  confidences of the cluster to which they belong, similar to the sum rule  of classifier combiners.

The numeric class labels LABT can be obtained from a partially labeled  PRTools dataset A by LABT = GETNLAB(A);

For good results the clustering levels of LABC should be diverse. In  particular they should not be nested.

Do not call this routine as a mapping in case LABC is a single level  clustering. It generates a hard to understand error.


 randreset                    % define random initialisation
 a = gendatclust(3000);       % generate 3000 2D objects in 10 clusters
 labn = getnlab(a);           % its numeric labels
 x = +a;                      % the unlabeled dataset
 labc = clustk(x,[20 25 35 50 100 200]);  % cluster by multilevel kmeans
 L = unique(labc(:,1));       % The prototypes in the first cluster level
 labt = zeros(size(x,1),1);   % construct labt, unlabeled
 labt(L) = labn(L);           % labeled part (20 objects)
 labout = clustc(labc,labt);  % run clustc
 e1 = mean(labout~=labn);     % compare output with true labels
 e2 = a*knnc(a(L,:),1)*testc; % result of using prototypes in 1NN
 fprintf('\n160 objects of which 20 cluster prototypes are labeled\n')
 fprintf('%5.3f error using clustering based confidences\n',e1)
 fprintf('%5.3f error using prototypes only\n',e2)

See also

datasets, mappings, dclusth, cluste, clustf, clusth, clustk, clustm, dcluste, dclusth, dclustk, dclustm, clusteval, clustcerr, clustnum, clusthc, plotdg,

ClusterTools Contents

ClusterTools User Guide

This file has been automatically generated. If badly readable, use the help-command in Matlab.