The direct-kernel self-organizing map (DK-SOM)

This paper introduces the Direct-Kernel Self-Organizing Map (DK-SOM) as a novel learning/visualization technique. DK-SOM is a self-organizing map just as the original SOM, but operates on a kernel transformation of the data, where the kernel transform is applied in a pre-processing phase. DK-SOM has the advantage over SOM that it operates in a latent feature space just as support vector machines, but otherwise keeps the SOM algorithm intact. For problems with more features than data, DK-SOM can offer a distinct speed-up over SVMs. DK-SOMs are capable of incorporating distance measures that are more natural to the problem domain via the definition of an appropriate kernel. The prime application areas for DK-SOM are data visualization and multiclass data mining.