An integrated approach for modelling of multivariate data based on statistical principles and neural networks

Abstract

For multivariate systems, empirical modelling using statistical methods is an issue in the presence of multicollinearity. Recently, neural networks provide a wide class of general-purpose, flexible non-linear non-parametric mapping, ignoring the fact of underlying data property. In this work, a composite system is designed to integrate the statistical principles of multivariate data, particularly multicollinearity, into the neural network architecture. The designed system is trained and validated with a simulated database based on the extent of correlation structure among the input variables. A real-life case with a large multi-dimensional system is illustrated under the proposed system.