FEATURES SELECTION FOR USE IN PREDICTIVE ANALYTICS MODELS OF REGIONS FOREIGN ECONOMIC ACTIVITY

Abstract

The work is devoted to the actual problem of designing and selecting features in the tasks of constructing predictive models of indicators of foreign economic activity of regions. The aim of the work is to develop a methodology for the use of graph models and dimensionality reduction methods for the selection of features in the construction of predictive analytics models of foreign economic activity of regions. As an approach to describe the structure of foreign economic relations, graph models were used that implement the possibility of building on the basis of the UMAP dimension reduction algorithm. To build an optimal graph based on the UMAP algorithm, it is necessary to vary the number of nearest neighbors for each vertex and the minimum metric distance to establish a connection between the vertices. It is shown that the clique symmetry coefficient of the graph makes it possible to estimate the local connectivity of points in the constructed graph, forming a generalized idea of the network structure from the position of the presence of clusters in it. The Gini index of the graph allows us to assess the correspondence of the global structure of the graph to real networks. The selection of features is based on the analysis of the graph clicks, which provides maximum mutual information MI with a minimum of features, which maximally reduces distortions in describing the structure of regional foreign economic relations. The application of the described approach made it possible to eliminate the multicollinearity of features, to select indicators, to expand the possibilities of using the existing data set by including new indicators that introduce new useful information about the subject area into the model. The method of feature selection considered in the paper can be rationally used to construct interpreted predictive models of foreign economic activity indicators and as one of the ways to reduce the dimension of the model feature space. The results obtained allow us to conclude about the advantages of the considered approach to the implementation of the selection of features in the construction of predictive models of indicators of foreign economic activity of regions.