On Cluster Analysis Via Neuron Proximity in Monitored Self-Organizing Maps

A potential application of self-organizing or topographic maps is clustering and visualization of high-dimensional data. It is well-known that an appropriate choice of the degree of smoothness in topographic maps is crucial for obtaining sensible results. Indeed, experimental evidence suggests that suitably monitored topographic maps should be preferred as they lead to more accurate performance. This paper reconsiders the basic toolkit for cluster analysis —based on the relative distance from each pointer to its immediate neighbours on the network— from this monitoring perspective. It is shown that the idea works nicely, that is, much useful information can be encoded and recovered via the trained map alone (ignoring any possible density estimate available). Moreover, the fact that a topographic map is not restricted to metric vector spaces makes this learning structure a perfect tool to deal with biological data, such as DNA or protein sequences of living organisms, for which only a similarity measure is readily available.