Selection of useful features in Kohonen map based clusterization of multidimensional data

摘要

A modification of the rules of Hebbian learning for selection of useful features in clusterization of multidimensional data is considered. The modified algorithm not only groups objects by similarity of descriptions, but also automatically generates a selection of useful features as a way of improving the performance of the grouping process. A series of comparative experiments is presented in which the modified algorithm proves to be more efficient than a Kohonen map designed for a preliminary selection of useful features.