Bi-Directional Chains of Neural Nets for Multi-Target Regression

TLDR: In this paper, the significance of directionality problem has been discussed and is addressed by proposing an ensemble based methodology, which can be used in both classification and regression using chain models, which although mostly competent, possess the issue of a uni-directional dependency.

Abstract: Multi-Target Regression refers to the problem where a set of n independent variables are used to predict the values of k target variables where both k and n are greater than 1. Most methods provide the provision for a regression problem with multiple targets including decision tree regressors and artificial neural networks. However, these methods end up making an assumption that their is no inter-dependency among the target variables. In numerous problems, this assumption turns out to be false which can be notably seen with the variance inflation factor and co-relation of these variables. This consideration was addressed in both classification and regression using chain models, which although mostly competent, possess the issue of a uni-directional dependency. In this work, the significance of directionality problem has been discussed and is addressed by proposing an ensemble based methodology. The comparative analysis of the proposed model is studied against the pre-existing models to explore the improvements in the performance of the model.