next up previous contents
Next: Classification Up: Clustering the SOM Previous: Hierarchical SOMs

Other clustering methods

Another way to cluster the map units of the SOM is based on the assumption that there are always some neurons which lie between the clusters. These will get fewer hits from an input data set than neurons in the clusters, so the borders between clusters are indicated by low-hit areas [53]. There are however several defects with this method, for example the fact that because the SOM approaches the probability density function of the input data, in the ideal case there will be an equal number of hits from the training data in each unit of the map.

Having labeled samples naturally opens new possibilities. For example all units having a certain amount of hits from a certain class can be joined to form one cluster. However, since the SOM is based on unsupervised learning this topic will not be discussed here.

An interesting way to postprocess the SOM was proposed by Pedrycz et al [33]. They used fuzzy sets and linguistic variables to analyze interrelationships between different vector components. Each component was expressed with fuzzy variables and associated linguistic labels, e.g. "Large", "Medium" and "Small". A linguistic description for a sample vector was achieved by aggregating these labels from all components. Using the fuzzy variables the membership value of each map unit for each description could be calculated, and by selecting the descriptions with largest spread and highest membership values, the map could be interpreted linguistically.

Juha Vesanto
Tue May 27 12:40:37 EET DST 1997