A Novel Robust Variable Selection Algorithm for Multilayer Perceptron

Abstract

In the study, a robust embedded variable selection algorithm for nonlinear modeling of datasets with outliers is proposed. The proposed algorithm, named LAD-LASSO-MLP, is an iterative two-step approach. Firstly, a multilayer perceptron (MLP) is trained with the initial dataset. Secondly, the least absolute shrinkage and selection operator (LASSO) is introduced to shrink input weights of the trained MLP and select the significant variables, in which the residual of LASSO is robustified by the least absolute deviations (LAD). Determination of the shrinkage parameter is carrying out via cross-validation. Finally, a couple of compared artificial dataset simulation are provided to verify the effectiveness of the proposed algorithm. Simulation results demonstrate that such estimation has strong resistance to outliers of response variable.