Cluster Number Selection Using Finite Mixture Model and Penalized Fisher Class Separability Measure

It is well known that classical clustering algorithms have a problem of determining an appropriate number of clusters. In this paper, the method of selecting cluster number is proposed using finite mixture modeling algorithm and penalized Fisher class separability measure criterion, at the same time the data set is partitioned into several appropriate clusters. The cluster number searching process can be partitioned so that the hierarchical clustering analysis is used to construct a ´tentative´ clustering for initial examination, followed by the iterative optimization, which continues to improve the clustering. The iterative optimization strategy is designed using Finite mixture model and penalized Fisher class separability measure. The parameterized architecture of the data set is described by Finite mixture model, which parameters are iteratively learned using expectation-maximization, or EM algorithm. The mixture model complexity, i.e. the number of clusters is selected to maximize the penalized Fisher class separability measure criterion.