Algoritmos para geração da frente de Pareto da regressão Lasso

Abstract

Modeling problems can involve a very large number of input variables, especially when we are interested in studying experimental data and obtaining an explanatory model for a certain phenomenon or event from them. In general, it is desired that the model be interpretable and that a clear conclusion can be obtained about the relationship of each input variable to the response, where a large number of variables can make such interpretation difficult. The use of Lasso regression is a viable option for obtaining models with a smaller number of input variables, while maintaining the accuracy obtained by them. However, generating models from Lasso requires greater computational effort compared to other methods, and for this reason, it is important that the process of generating these models is efficient. In this study, we conducted an analysis of different algorithms for generating models from Lasso, as well as ways to reduce computational effort when we want to obtain multiple models for different regularization parameter values for a given problem, by approximating the Pareto front of Lasso.