Data Analysis Using the Nonlinear Dimension Reduction Algorithms

Abstract

The problem of the preliminary structure analysis of the datasets using the t-SNE and the UMAP algorithms, which are the algorithms for the nonlinear dimensionality reduction, has been considered. The possibilities of the t-SNE and the UMAP algorithms are investigated in the context of obtaining the adequate visualization results for the datasets in the two-dimensional space, associated with the datasets in the original multidimensional space, when deciding on the choice of the clustering algorithm which best (in the sense of a certain criterion) determines the number and the structure clusters. The technique for the preliminary structure analysis of the datasets using the nonlinear dimensionality reduction algorithms, which makes it possible to perform adequate visualization in two-dimensional space for the initial multidimensional data and form reasonable recommendations for choosing the number of clusters and the clustering algorithms has been proposed. The examples of the preliminary analysis of the time series groups, acting as the datasets, confirming the effectiveness of the proposed method have been given.