Hybrid mean impact value with nonnegative garrote for input variable selection of multi-layer perceptron
Hongxun Wang, Fengying Ma, Kai Sun · Proceedings of the 2020 4th International Conference on Electronic Information Technology and Computer Engineering · 2020
In the paper, a new adaptive variable selection algorithm for nonlinear regression multi-layer perceptron (MLP) is proposed. The proposed algorithm applies the nonnegative garrote (NNG) to shrink the input weights of the trained MLP. An adaptive indicator is designed by introducing the mean impact value (MIV) algorithm, which can improve the coefficient shrinkage efficiency of the NNG algorithm. The proposed algorithm uses cross-validation and Bayesian information criterion to calculate the optimal shrinkage parameter. The performance of adaptive variable selection algorithm is demonstrated by an example of artificial dataset. Simulation results show that the algorithm has better model accuracy and simplicity than other state-of-art algorithms.