DSELPROTO

Select prototypes from given dissimilarity matrix

    N = DSELPROTO(D,K,CRIT,NINIT)
    N = D*DSELPROTO(K,CRIT,NINIT)

Input
D	Dissimilarity matrix  between objects (rows) and protoypes (columns)
K	Number of prototypes to be selected
CRIT	'kmeans', 'kmedoids' or 'kcentres' (default: 'kmeans')
NINIT	Indices of preselected prototypes (optional)

Output
N	Indices of selected prototypes

Description

Select and rank a subset of K prototypes (columns) of the matrix D using  a greedy forward selection approach such that the maximum NN distance  from all objects to these prototypes minimizes

• for CRIT is 'kmeans' or 'kmedoids': mean(min(D(:,N(1:K),[],2)).

-	for CRIT is 'kcentres' : max(min(D(:,N(1:K),[],2)).

This routines tries to sample the given prototypes (possibly the original  objects) such that they are evenly spaced judged from the dissimilarities.  This may be used as a systematic initialisation in kmeans, kmedoids or  kcentres procedures.

The dissimilarity matrix D might be rectangular (prototypes different  from objects). It might also be asymmetric.

See also

datasets, mappings, selproto,