Мінімаксні підходи для відбору ознак

Abstract

У роботі розглядаються мінімаксні підходи до відбору ознак у задачах інтелектуального аналізу даних.

The paper considers minimax approaches to feature selection in data mining tasks. The main existing feature selection methods, including filter, wrapper, and embedded approaches, are analyzed, and their limitations under noise, structural heterogeneity, and distribution shift are identified. Special attention is paid to the minimax formulation of the feature selection problem, where not the average model risk but the maximum error over a set of scenarios is optimized. It is shown that such an approach can improve model robustness, reduce the impact of spurious features, and enhance performance in worst-case situations.