The experimental analysis of the method of clustering and ranking of multidimensional data using the Kohonen neural network

The paper proposes a methodology of clustering and ranking data using the Kohonen neural network based on space-correlation properties of a training sample regardless of the network learning algorithm. The possibility of applying the promising method of linear transformation of training samples coordinates for clustering weakly correlated spatially inseparable data is shown experimentally. The paper demonstrates the usage of ranking to highlight the border instances and define the level of closeness to neighborhood cluster, which makes it possible to solve the problem of finding cluster boundaries in spatially inseparable data. The necessity of the multilayer clustering is justified in the case of uneven spatial data distribution. The method of clustering and ranking is illustrated by the example of analysis of financial statements empirical data. The technique is applicable to samples of small and medium size.