A Set-Medoids Vector Batch SOM Algorithm Based on Multiple Dissimilarity Matrices

Abstract

This paper gives a batch SOM algorithm that is able to training a Kohonen map taking into account simultaneously several dissimilarity matrices, that are obtained using different sets of variables and dissimilarity functions. This algorithm is designed to provide a partition and a set-medoids vector representative for each cluster, and learn a relevance weight on the training for each dissimilarity matrix by optimizing an objective function. These relevance weights change at each algorithm's iteration and are different from one cluster to another. The proposed algorithm provides a collaborative role of the different dissimilarity matrices, aiming to cluster and visualizing the data while preserving their topology. Several examples illustrate the usefulness of the proposed algorithm.