CLUSTCERR

Cluster classification error on indiviual clusterings

    [E,N} = CLUSTCERR(LABC,LABT)
    [E,N} = CLUSTCERR(LABC,A)
    [E,N] = LABC*CLUSTCERR(LABT)

Input
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.  Default: true labels used in some other routine.
A	PRTools labeled dataset used for the clustering.

Output
E	Vector with classification errors of the clusterings in LABC.
N	Vector with number of clusters per clustering.

Description

This routine evaluates a clustering of a dataset A by comparing the true  labels LABT of A with cluster labels derived from a cluster prototype  object. This can be understood as evaluating the clusterings by active  labeling.

A is a labeled dataset with M objects. LABT is a vertical vector  containing the true numeric labels of A. LABT = GETNLAB(A). LABC is a set  of K clustering results with indices pointing for every object to a  cluster prototype in A.

E is the fraction of misclassified objects in A if every cluster is  assigned to the class of the object indicated by LABC. This is based on a  training set of N prototypes. The classification error E is based on all  objects, including the prototypes.

In case LABC is an MxK result of a multilevel clustering, E and N are  vectors with K elements. Use MCLUSTCERR for a classification result  that combines clusterings.

Example(s)

 randreset; % take care of reproducability
 A = gendat(mnist8,25000);
 randreset; labc1 = A*clustm(false);   % no nesting
 randreset; labc2 = A*clustm;          % nesting
 [e1,n1] = clustcerr(labc1,A);
 [e2,n2] = clustcerr(labc2,A);
 figure;
 semilogx(n1,e1); hold on
 semilogx(n2,e2);
 title(['Active learning curve: ' getname(A)])
 xlabel('Training set size - number of clusters')
 ylabel('Classification error')
 legend('Original','Nested')
 fontsize(16)

See also

datasets, mappings, knnc, cluste, clusth, clustk, clustkh, clustm, clustf, clustr, dcluste, dclustf, dclusth, dclustk, dclustm, dclustr, clusteval, clustc, clustnum, clustlcurve,