Novel Approach for Improving ML Model Performance by Means of Subgroup Discovery Methods

Neganova E.A. Novel Approach for Improving ML Model Performance by Means of Subgroup Discovery Methods Master's thesis. Perm: National Research University Higher School of Economics, Department of Information Technologies in Business, 2020. The master’s thesis contains 109 pages including 12 Appendix pages, 26 tables, 38 references, 65 figures, 26 formulae. The research is aimed at developing a novel approach for improving machine learning models' performance by automating the process of hypothesis generation on adding new features using Subgroup Discovery class methods. In the first chapter, the iterative process of machine learning models creation is described, existing approaches for improving models' performance as well as tools that automate the process of feature engineering are introduced. The choice of Subgroup Discovery class methods is justified, requirements to the novel approach and its limitations are defined. The second chapter contains a description of the developed approach. In the third chapter three cases of the approach application within various subject areas are presented: the anti-fraud system development, Titanic passengers' survival prediction and energy consumption forecasting. Keywords: machine learning, hypothesis generation, model performance improvement, feature engineering, subgroup discovery.