CLUSTEVAL

Evaluate clusterings by various performance measures

      E = CLUSTEVAL(LABC,LABT,TYPE)
      E = CLUSTEVAL(LABC,A,TYPE)
      E = LABC*CLUSTEVAL(LABT,TYPE)
      E = LABC*CLUSTEVAL(A,TYPE)

Input
LABC	Index array, size [M,N], indices of cluster prototypes for M objects in N clusterings.
LABT	Double array, size [M,1] with true object labels.
A	PRTools dataset used for obtaining LABC by some clustering  procedure.
TYPE	String with desired performance measure, see below.

Output
E	Evaluation result using the performance measure give by TYPE.  See below for possibilities.

Description

Computation of a set of cluster performance measures between estimated  cluster labels LABC and true object labels LABT. In case LABC is a  multilevel clustering (N>1) the result E is a structure ready to be  plotted by PLOTE. If E is omitted (no output) the result is directly  plotted. Performance measures that do not generate a curve (see below)  are plotted on the screen.

In case LABC is a single clustering (N==1) just scalar results are  returned (except for TYPE is 'roc', which generates two values, see  CLUSTROC). In case TYPE is omitted the default measure for multilevel  clustering is 'actl'. For a single clustering all measures are returned  in a structure or printed on the screen.

It is assumed that the cluster labels LABC are indices to cluster  prototypes with true labels as given by the correponding entries in LABT.  The true labels LABT can be derived as doubles from a PRTools dataset A by LABT = GETNLAB(A);

The following performance measures are available;

ROC	Relative operating characteristics, see CLUSTROC.
AUC	Scalar, area under the ROC curve [e1,e2].
MINE	Scalar, MIN(E1+E2) with E1 and E2 the two errors of the ROC.
ACTL	Classification error of assigning all objects to the true class  of the cluster prototypes (active learning).
COMB	Classification error based on the true class of the cluster  prototypes and combining multilevel cluster confidences.
NEST	Classification error based on the true class of the cluster  prototypes after nested combining cluster levels, see RECLUSTN
ADRI	The adjusted Rand index, see Wikipedia.  It is between 0 and 1 and 1 for consistent clusterings.
NMI	Normalised mutual information between 0 and 1.

See also

datasets, mappings, knnc, cluste, clusth, clustk, clustkh, clustm, clustf, clustr, dcluste, dclustf, dclusth, dclustk, dclustm, dclustr, reclustn, clustcerr, clustc, clustnum, clustroc, plote,