Ensuring Accuracy and Stability of Multidimensional Data Clustering Using Kohonen Self-Organizing Maps Based on Automatic Data Reduction

The article investigates a method for improving the accuracy and stability of multidimensional data clustering by combining Kohonen self-organizing maps with an automatic feature reduction procedure. A method for removing insignificant features using an interquartile range algorithm integrated into a software application for building SOMs is proposed. A comparison of the clustering results before and after reduction is performed with unchanged training parameters and a fixed seed value, a parameter that determines the initial state of the random number generator and ensures the reproducibility of the initialization of the neuron weight coefficients. It has been proven that the removal of features with low variability does not violate the topological structure of SOM, preserves the stability of cluster separation, and reduces computational costs. A high degree of correspondence between the data before and after reduction has been confirmed based on statistical analysis. It has been established that the proposed method increases the efficiency of SOM and is suitable for analyzing large multidimensional samples