The methods of extracting the contribution of variables in artificial neural network models - Comparison of inherent instability

摘要

The methods for quantifying variable importance in neural network based model are used in many scientific fields. These methods open the "black box" model and give the information about relative importance of explicative variables. However, the contribution of independent input variables is usually calculated based on the single, best neural model. It was proved in several scientific reports that the use of a single neural network architecture can lead to misleading results. In this work, the novel strategy, based on a group of neural models containing models of the same architecture trained, starting from different random values of connection weights as well as models of different architectures is proposed. The results are based on the model of relationship between chemical honey parameters as well as the temperature and dielectric loss coefficient of honey. This new approach promises to produce relatively precise and reliable results and is valuable in real-world applications. (C) 2016 Elsevier B.V. All rights reserved.