An Asymmetric Criterion for Cluster Validation - Developing Concepts in Applied Intelligence - page 6

Fig. 2. Computing the Stability of Cluster C i .
Fig. 3. Two primary partitions with k=3. (a) True clustering. (b) Spurious clustering.
This approach to evaluate the stability has some drawbacks which can appear
under some circumstances. Here, a drawback of this procedure is introduced. Then an
approach which is based on a novel asymmetric measure named Alizadeh-Parvin-
Moshki-Minaei, APMM criteria, is proposed.
Fig. 3 shows two primary partitions with the evaluated stability of each cluster and
the process of its computation. In this example K-means is applied as the base
clustering algorithm and the true number of clusters, k=3, is forced. For this example
the number of all partitions in reference set is 40. In 36 partitions the result is
relatively similar to Fig 3.a and in 4 times the top left cluster is divided into two