Improving the performance of self-organizing maps via growing representations

Self-organizing maps (SOMs) are a type of artificial neural network that through training can produce simplified representations of large, high dimensional data sets. These representations are typically used for visualization, classification, and clustering and have been successfully applied to a variety of problems in the pharmaceutical and bioinformatics domains. SOMs in these domains have generally been restricted to static sets of nodes connected in either a grid or hexagonal connectivity and planar or toroidal topologies. We investigate the impact of connectivity and topology on SOM performance, and experiments were performed on fixed and growing SOMs. Three synthetic and two relevant data sets from the chemistry domain were used for evaluation, and performance was assessed on the basis of topological and quantization errors after equivalent training periods. Although we found that all SOMs were roughly comparable at quantizing a data space, there was wide variation in the ability to capture its underlying structure, and growing SOMs consistently outperformed their static counterparts in regards to topological errors. Additionally, one growing SOM, the Neural Gas, was found to be far more capable of capturing details of a target data space, finding lower dimensional relationships hidden within higher dimensional representations.