Extending the SOM algorithm to visualize word relationships

Abstract

Self Organizing Maps (SOM) are useful tools to discover the underlying structure of high dimensional data. However the algorithms proposed in the literature rely on the use of symmetric measures such as the Euclidean. Therefore when asymmetry arises they fail to reflect accurately the object proximities and the resulting maps become often meaningless. This is a serious drawback for several applications such as text mining in which the object relations are strongly asymmetric. In this paper, we propose two variants of the original SOM algorithm that are able to deal successfully with asymmetric relations. The algorithms are tested using real document collections, and the performance is reported using appropriate measures. The asymmetric algorithms improve significantly the maps generated by their symmetric counterpart. © Springer-Verlag Berlin Heidelberg 2005.