A novel sensitivity-based method for feature selection

Abstract Sensitivity analysis is a popular feature selection approach employed to identify the important features in dataset. In sensitivity analysis, each input perturbed one-at-a-time and response of machine learning model examined determine feature's rank. Note that existing perturbation techniques may lead inaccurate ranking due their parameters. This study proposes novel involves using complex-step. The implementation complex-step framework deep neural networks as method provided this paper, its efficacy determining for real-world datasets demonstrated. Furthermore, filter-based methods are employed, results obtained from proposed compared. While classification task indicated outperformed other methods, case regression task, it was found perform more or less similar methods.