A new variable selection algorithm for LSTM neural network

Abstract

This paper proposes an accurate and reliable input variable selection algorithm by embedding a nonnegative garrote (NNG) algorithm into long short term memory (LSTM) neural network to perform data-driven modeling on a highly nonlinear and dynamic time-delay dataset. Firstly, an LSTM deep neural network is trained, and a well-trained LSTM network is obtained by optimizing the parameters of LSTM through a grid search algorithm. Secondly, the initial input weights of LSTM are compressed accurately by the NNG algorithm, and block cross-validation is applied to the optimization calculation process to achieve input variable selection. Finally, the performance of the algorithm is verified by the improved Friedman time-delay artificial datasets. Simulation results show that the algorithm could construct a more simplified and better predictive model than other traditional algorithms.