On the use of cluster validity for evaluation of MIGSOM clustering

Organizing data into homogeneous groups (clusters) is one of the most fundamental tasks of recognizing interesting repartitions in the data set. However, the challenge of clustering is how to evaluate the obtained results. In addition, clustering directly a data set, can be computationally heavy according to the data dimension. For this reason, it is convenient to cluster quantization prototypes to reduce computational cost. Self-Organizing Map (SOM) technique is appropriate for vector quantization and clustering high dimensional data. Nevertheless, SOM suffers from its static structure. Therefore, we have proposed a dynamic method called MIGSOM (Multilevel Interior Growing Self-Organizing Maps) which presents a different architecture of those which are proposed in the literature. The irregular and multilevel structure of MIGSOM is developed during the training process according to the data direction. Thus, the main objective of this paper is to evaluate clustering of MIGSOM prototypes with k-means. Therefore, in this paper we review our learning method MIGSOM. Then, we compare the use of three cluster evaluation measurements (DB, F_measure, Conn_Index) applied to the trained neural map by MIGSOM, GG and SOM.