EMCLUST

Expectation-Maximisation clustering

   [LABELS,W_EM] = EMCLUST (A,W_CLUST,K,LABTYPE)

Input
A	Dataset, possibly labeled
W_CLUST	Cluster model mapping, untrained (default: nmc)
K	Number of clusters (default: number of classes in A)
LABTYPE	Label type: 'crisp' or 'soft' (default: label type of A)

Output
LABELS	Integer labels for the objects in A pointing to their cluster
W_EM	EM clustering mapping

Description

The untrained classifier mapping W_CLUST is used to update an initially  labeled dataset A by iterating the following two steps
1. Train W : W_EM = A*W_CLUST 2. Relabel A : A = prdataset(A,labeld(A*W_EM*classc))  This is repeated until the labeling does not change anymore. The final  classification matrix is returned in B. The final crisp labeling is  returned in LABELS. W_EM may be used for assigning new objects.

If K is given, a random initialisation for K clusters is made and labels  of A are neglected. If K is omitted the given labeling is used as  initailisation.

An early stopping of the EM algorithm is controlled by PRTIME.

LABTYPE determines the type of labeling: 'crisp' or 'soft'. Default: label  type of A. It is assumed W_CLUST can handle the LABTYPE requested.  Only in case LABTYPE is 'soft' the traditional EM algorithm is followed.  In case LABTYPE is 'crisp' EMCLUST follows a generalised k-means  algorithm.

See also

mappings, datasets, prkmeans, prprogress,