Análise comparativa de métodos de seleção de variáveis em problemas de classificação

Abstract

In this study, a comprehensive comparison between the SHAP method and the Lasso method for variable selection is presented. The methodologies of both techniques are explored and juxtaposed, utilizing key selection metrics such as precision, recall, F1 score, and accuracy in both simulated and real database scenarios. The tests applied show SHAP as a good competitor for variable selection methods, with results even slightly superior to Lasso in the cases presented here, both on a simulated and real basis, maintaining competitive accuracy levels in relation to the complete model. Despite being somewhat close in accuracy, SHAP notably reduces the variable space, demonstrating its selection prowess. Additionally, a robustness study involving perturbation values in training, testing, and combined datasets confirms the resilience of the variables selected by SHAP, particularly in terms of accuracy. These analyses underscore the efficacy of the SHAP method as a versatile and potent tool for variable selection, promising improved interpretability and performance in machine learning applications.