Clustering the Self-Organizing Map Based on the Neurons' Associated Pattern Sets

Abstract

This paper presents an automatic clustering system, built as a committee machine, which is used to cohesively partition the self-organizing map. In the proposed method, each expert from the committee machine analyzes the connections of the neuron grid based on a particular similarity matrix, and thus decides which ones should be pruned by gradually removing them and observing the intervals of stability. Those intervals are regarded as the ones in which the number of clusters found through connected components remain constant. The output of each expert is a connectivity matrix that effectively expresses which connections should remain as a binary true or false value. The final stage of the committee machine consists of combining the outputs of the experts, and through majority voting establish which connections should remain in the grid, and hence performing the segmentation of the map. The system was evaluated through its application to synthetic and real world data sets.