Topology preservation in nonvectorial SOM

Although the SOM algorithm has been widely used with vectorial data, its principle is not restricted to metric vector spaces. Indeed, any set of items for which a similarity or pseudo-distance measure is available could be mapped onto the SOM grid in an ordered fashion. As Kohonen and Somervuo (2002) pointed out, the optimal speed of shrinking of the neighbourhood range function on nonvectorial SOM algorithm should be experimentally determined. This paper presents the use of the UDL monitoring algorithm for the nonvectorial approach to SOM learning rule.