Skip to Main Content
The aim of the paper is to propose and evaluate a hybrid approach to generate a representative training dataset of the required size. Prototype selection is understood as a selection of the representative prototypes from the original training dataset. The basic assumptions underlying the proposed method is that the prototype selection is carried out after the training dataset has been grouped into clusters, and that prototypes are selected from each of thus obtained clusters. Under these assumptions the number of clusters produced has a direct influence on the size of the reduced dataset. When the number of clusters exceeds the required final size of the training dataset, clusters need to be merged. Clusters merging may not be an easy task in case clusters have a heterogeneous structure. The paper considers the problem of cluster merging and proposes to eliminate the problem of the cluster heterogeneity through reaching a consensus-based solution.