BSD-FE: a fast and effective framework for automated feature engineering

Abstract This study addresses the problems of high computational cost, limited transformation strategies, and dimensional expansion encountered in automated feature engineering (AutoFE) and proposes a new approach based on the Bernstein-search differential evolution (BSD) algorithm (BSD-FE). BSD-FE is designed as a flexible differential evolution approach that reduces sensitivity to parameter settings and adapts to a wide range of data types through an adaptive crossover structure. In the study, 15 classification and 10 regression datasets were used; new features were generated using unary/binary operators, spline and quantile-based discretization, and filter-based feature selection strategies. BSD-FE was tested on random forest, XGBoost, and LightGBM models and compared with seven different AutoFE methods according to macro- F 1 and inverse relative absolute error metric. The results show that BSD-FE provides balanced and superior performance in terms of both accuracy and computation time. It produced an average of 24 new features and significantly reduced convergence time by limiting dimensional growth. The findings reveal that BSD-FE offers a reliable and generalizable AutoFE solution across different tasks and data types.