Deriving mathematical models from neural networks: A method for deducing individual effects of factors on a response variable

Abstract

This research outlines a novel approach to obtaining mathematical models from neural networks. The target scenario is one where a response variable depends on a number of factors, each factor has an effect which is a function of the factor and the response variable is the sum of the effects of the factors. A neural network was trained such that response values were generated from factor values. It was assumed that each effect was zero when the underlying factor was set to zero. The effect of a factor could be isolated by setting all other factors to zero, so that the response value became equal to the effect of the factor being isolated. In that way each effect was isolated and then modelled as a function of the factor. Thus, the technique was developed, for modelling a response variable as a function of its input factors.