Automatic Estimation of Cluster Number in Fuzzy Co-clustering Based on Competition and Elimination of Clusters

Fuzzy co-clustering induced by multinomial mixture model (FCCMM) is one of the effective methods for analyzing cooccurrence information data. In FCCMM, we have to determine the number of clusters in advance. Furthermore, we have to select the best solution front a lot of trials with various patients of the number of clusters and random initial values. In Gaussian mixture models, a robust EM clustering algorithm, which is robust to initial values and can automatically estimate the optimal her of clusters, have been proposed in order to resolve such problems. It introduces an entropy-based penalty term with respect to cluster volumes to the objective function of the standard EM algorithm and obtains the optimal number of clusters by continually eliminating clusters with low competitiveness in volumes. In this paper, we propose a method which automatically estimates the optimal number of clusters without being influenced by the initial values by introducing the entropy-based penalty term with respect to cluster volumes to the objective function of FCCMM in a similar manner to the robust EM algorithm and de strate its estimation performance through numerical experiments.