Neural network ensembles with missing data processing and data fusion capacities: applications in medicine and in the environment

Abstract

An important way to reach a qualitative improvement of Artificial Neural Networks (ANNs) is to incorporate biological features in the networks. Our proposal introduces modularity at two different levels, first, at the network level and second, at the intrinsic level of the networks, generating neural network ensembles (NNEs). We designed three NNEs which incorporated new capacities with regard to the processing of missing data, introduced hybrid modularity, and also used modular ANNs for building the NNEs. We have investigated a suitable NNE design where selection and fusion are recurrently applied to a population of best combinations of classifiers. In this paper we explore the ability of the proposed NNE in different automated decision making applications, especially for those with inherent complexity in their information environment. We present some results on dementia diagnosis and on automatic pollutants detection.