On the Procedures of Generation of Numerical Features over Partitions of Sets of Objects in the Problem of Predicting Numerical Target Variables

Abstract

Analysis of criteria for the solvability/regularity of problems and of the correctness of algorithms is applied here to the problem of prediction of the values of numerical variables. It is shown that partial regularity is a necessary and sufficient condition for the solvability of the corresponding system of the classification problems. Cross-validation experiments conducted on several datasets from the field of biomedicine (non-invasive diagnostics of magnesium concentration in blood plasma), bioinformatics (prediction of the protein secondary structure), and solid-state physics (prediction of the properties of high-temperature superconductors) have demonstrated the effectiveness of the developed methods for generating “synthetic” informative numerical features and for increasing the accuracy of prediction of the numerical target variables.