Clustering of the self-organizing map using particle swarm optimization and validity indices

In this paper, an automatic clustering algorithm applied to self-organizing map (SOM) neurons is presented. The connections of the SOM grid are pruned according to a weighted sum of a set of measures of connection strength between adjacent neurons. The coefficients of the weighted sum are obtained through particle swarm optimization (PSO) search in the multidimensional problem space, where the fitness function is the composed density between and within clusters (CDbw) validity index of strongly connected groups of neurons, while scanning through different values of the minimum cluster size so as to find stable regions with a reasonable trade-off between their length and their mean CDbw value. Simulation results are further presented to show the performance of the proposed method applied to synthetic and real world datasets.