Data attributes decomposition-based hierarchical neural network

Abstract

The “black box” problem in neural network is being much concerned, which contributes to more and more researches on the structures of the neural network. Hierarchical neural network (HNN) is one kind of the neural networks that pays attention to the inner structure of network with the presentation of modular parts. In order to reducing the dependence of expert system in HNN, in the paper, a data attributes decomposition-based hierarchical neural network (DADHNN) is proposed through analyzing the information of data attributes based on two kinds of hierarchical structure. Also, two datasets from UCI repository and the production datasets of purified terephthalic acid (PTA) solvent system of a chemical plant are both used for the practical application. The application results show that the DADHNN method can establish the subnets automatically and have explainable ability to users, which provides a new way to the industry product-processing.