An Adaptive Variable Selection Algorithm for Gated Recurrent Unit Based on Sensitivity Analysis and Nonnegative Garrote

Abstract

In this study, sensitivity analysis and nonnegative garrote (NNG) are combined to realize adaptive variable selection for gated recurrent unit (GRU). Firstly, the sensitivity analysis based on variance decomposition is used to quantify the correlation between each input variable and the output variable, and the total sensitivity index of each input variable is calculated. Secondly, the total sensitivity index is added to the NNG as an adaptive weight vector to achieve adaptive variable selection. Finally, an artificial dataset with the characteristic of time series is used to verify the effectiveness of the proposed algorithm. Simulation results show that the proposed algorithm can identify the relevant variables effectively and improve the predictive performance of the model.