Development of a novel variable selection algorithm for LSTM with LASSO

Abstract

In the paper, a nonlinear regression with long short-term memory (LSTM) and least absolute shrinkage and selection operator (LASSO) is developed. The LSTM is used to handle strong nonlinearity, dynamic and time-series, and the LASSO is applied to perform input variable selection for LSTM. Firstly, a deep neural network of LSTM is constructed and trained from the initial data set. After that, the LASSO is introduced to shrink the input weights of well-trained LSTM, in which the Monte Carlo method (MCM) and moving block cross-validation (MBCV) are applied to perform the optimization. The performance of the proposed algorithm is verified via two artificial data sets, and it shows that the proposed results have a better superiority than other algorithms.