Clustering of socio-economic data with Kohonen maps
Data sets of socio-economic feature vectors often involve a large number of high dimensional samples. In order to exhibit manageable representations and to extract useful information from these data, some preliminary reduction of the complexity may be required. For this purpose, the Kohonen maps connectionist model is an emerging approach, providing a valid alternative to traditional techniques for unsupervised data analysis. It maps the feature space onto a two-dimensional lattice, while attempting to preserve some topological properties of the original samples. On a data set supplied by EUROSTAT, this paper shows how problems stemming from the non-deterministic behaviour of the algorithm can be overcome to produce a suitable Kohonen map. Tools for gaining more insight into the data representation are proposed. Kohonen's algorithm is interpreted in terms of gradient descent along a potential function. The connectionist approach is compared with three traditional techniques, namely vector quantisation, principal component analysis and hierarchical clustering.
Bibliographic Reference: Paper presented: PASE '92, Prague (CS), December 6-9, 1992
Availability: Available from (1) as Paper EN 37338 ORA
Record Number: 199310359 / Last updated on: 1994-11-29
Original language: en
Available languages: en