Assessing the contribution of variables in feed forward neural network

Neural networks are being used as tools for data analysis in a variety of applications. Neural network technique is cited in the literature as a ‘Black Box’ approach and criticized most for the lack of interpretability of the network weights obtained during the model building process. Some attempts have been made in the past in this direction to interpret the contributions of explanatory variables in prediction problem using the weights of neural network. In the present study, a new approach is proposed to interpret the relative importance of independent variables in neural networks and a comparison with the connection weight approach is presented. The performance of this approach is studied for various data characteristics and is found to be a better method in comparison to a well known method existing in the literature. An example from behavioral science is also considered to illustrate how the performance of the proposed approach translates to a real life situation.