Algorithmic approaches for studies of variable influence, contribution andselection in neural networks

Two methods of studying variable influence and contribution in neural netwo rk (NN) models are examined in this work. The first approach, a variable se nsitivity analysis method, is based on sequential zeroing of weights (SZW) of the connection between the input variables and the first hidden layer of an established NN model. The second approach is based on systematic variat ion of variables (SVV) while the other variables are either kept constant o r systematically varied synchronously. It is shown that there is a close re semblance between the results obtained by the proposed method for studies o n variable influence and contribution in artificial NN models and the natur e of the functions used to generate these synthetic data sets. The standard NN models are thus suitable not only for function approximation and nonlin ear relationships, but also to a high degree able to represent the nature o f the input variables. We are thus able to demonstrate that highly intercon nected NN models, which are sometimes considered to be black boxes, can be highly transparent. The information generated about the variables, using th e methods proposed in this work, can thus serve as a guide to the interpret ation of influence, contribution, and selection. The methods proposed in th is study are further compared to other sensitivity analysis methods as stat istical sensitivity analysis (SSA) and beta-tests. Furthermore, the methods applied to the synthetic data sets were used on three real data sets, givi ng, for instance, additional information on the effect of principal compone nt (PC) regularization of input variables. (C) 2000 Elsevier Science B.V. A ll rights reserved.