Classifier ensemble with evolutionary optimisation enforced random projections

Abstract

An effective multi-classifier fusion (MCF) system is demanding in the clinical context in terms of integrating various diagnosis/prognosis predictive models to arrive at a stable and consentaneous medical decision. In this study, we introduced a novel MCF framework for a classifier ensemble with the evolutionary optimisation of random-projections (termed CLEER). The proposed CLEER generated a number of diverse base classifiers via training on the mapped data from the Bernoulli random projection. It innovatively framed the classifier fusion into an evolutionary computation architecture wherein the required diversity and accuracy were enforced by optimising the random projection components using a genetic algorithm. The efficacy of CLEER has been demonstrated via extensive evaluations using twenty public datasets from various research fields, as well as four clinical datasets. A comparative analysis showed that the ensemble diversity was effectively enhanced on using CLEER, and more accurate classifications were achieved as compared to the state-of-the-art benchmark ensemble methods. The proposed CLEER could serve as a potential tool for the fusion of diagnostic or prognostic models for assisting in medical decision making.