Summary:
In this paper a method for the analysis of the influence input variables have on the output of a fitted model is presented. It is based on the statistical study of the derivatives of the output of the model with regards to its standardized inputs (also called I/O sensitivities). If the model is not using a particular regressor (input variable) to estimate its output, then the corresponding sensitivities will have a low degree of significance. Moreover, if the model is able to estimate the derivatives of the underlying process from a set of samples of the input/output mapping (training set), then the sensitivities of the model will be an estimation of the sensitivities of the underlying process. We have tested the technique on different problems, with a special focus on time series prediction problems, although it is not the only possible application domain. We apply the method to connectionist models (Artificial Neural Networks), but the method can be applied to any non-linear model, as far as its I/O sensitivities can be computed or estimated. This procedure has also been applied to recurrent models. The aim of this analysis is the selection of an appropriate subset of input variables. It is also a useful tool for the explanation of the underlying process.
Keywords: Variable selection,feedforward networks, recurrent neural networks, time series prediction
Published on paper: June 1998.
Citation:
A. Muñoz, T. Czernichow, Variable selection using feedforward and recurrent neural networks. Engineering Intelligent Systems for Electrical Engineering and Communications. Vol. 6, nº. 2, pp. 91 - 102, June 1998.
