A density-based clustering of the Self-Organizing Map using graph cut

In this paper, an algorithm to automatically cluster the Self-Organizing Map (SOM) is presented. The proposed approach consists of creating a graph based on the SOM grid, whose connection strengths are measured in terms of pattern density. The connection of this graph are filtered in order to remove the mutually weakest connections between two adjacent neurons. The remaining graph is then pruned after transposing its connections to a second slightly larger graph by using a blind search algorithm that aims to grow the seed of the cluster's boundaries until they reach the outermost nodes of the latter graph. Values for the threshold regarding the minimum size of the seeds are scanned and possible solutions are determined. Finally, a figure of merit that evaluates both the connectedness and separation selects the optimal partition. Experimental results are depicted using synthetic and real world datasets.