Using Curves of Permanence to Study the Contribution of Input Variables in Artificial Neural Network Models: A New Proposed Methodology

Abstract

Understanding the influence of some factors on a particular phenomenon can be very relevant in many cases of decision-making. An example would be the identification of the level of influence that factors such as smoking, stress and lack of exercise have on the predisposition to heart disease. Knowing which of these inputs are relevant for a person to become a cardiac patient, it is possible to take some preventive measures. This article presents a new method to assist the not so simple task of feature selection, using the statistical function called curve of permanence. In this work we show parmanence curves applied on result data from the executions of some existing algorithms of feature selection, all of them based on Artificial Neural Networks (ANN). The objective of this study is to propose a technique that provides robustness to the process of determine the values of contributions of the inputs of an ANNs.