Optimal cluster selection based on Fisher class separability measure

In this paper, a novel hierarchical clustering algorithm is proposed, where the number of clusters is optimally determined according to the Fisher class separability measure. The clustering algorithm consists of two phases: (1) Generation of sub-clusters based on the similarity metric; (2) Merging of sub-clusters based on the Fisher class separability measure. The proximity matrices are constructed. Each subcluster comprises patterns close to each other in proximity metric. The trellis diagram is used for searching of subclusters. Connections between consecutive layers in the trellis diagram are weighted by the similarity metric. The threshold for the merge of sub-clusters is numerically designed according to Fisher class separability measure. The proposed algorithm can pre-process the data for the supervised learning. It also can be applied for the optimal determination of basis functions for radial basis function (RBF) networks.