An Asymmetric Criterion for Cluster Validation - Developing Concepts in Applied Intelligence - page 4

which is based on cluster stability evaluates the primary clustering results instead of
final clustering.
3 Proposed Clustering Ensemble
In this section, first the proposed clustering ensemble method is briefly outlined, and
then its phases are described in the subsequent subsections in more detail.
Fig. 1. The proposed clustering ensemble method.
The main idea of the proposed clustering ensemble method is to utilize a subset of
the best performing primary clusters in the ensemble instead of all of them. It seems
that every cluster does not have a good quality. So, in this method just those clusters
which satisfy enough stability to participate in the combination are chosen. The
cluster selection is done based on cluster stability which is defined according to
Normalized Mutual Information, NMI.
Fig. 1 depicts the proposed clustering ensemble procedure. First, a set of B primary
partitions is provided using K -means and Linkage methods to create the necessary
diversity for an ensemble. Then, the Improved Stability, IStability, is computed for all
clusters of each obtained partitions. The manner of computing IStability is described
in next sections in detail. After that, a subset of the most stable clusters is selected to
participate in the final decision committee. This is done simply by applying a
threshold over the IStability value of any cluster. In the next step, the selected clusters
construct the co-association matrix. Several methods have been proposed how to