An ensemble extended belief rule base decision model for imbalanced classification problems

Abstract

Class imbalance is a common problem in real-world applications and usually poses a major challenge to artificial intelligent (AI)-based decision models. The present work introduces a novel ensemble decision model which utilizes an explainable and fast-growing rule-based system, called extended belief rule base (EBRB) decision model, to alleviate the impact of class imbalance, where the proposed ensemble EBRB model includes two core components: a diversity-based base EBRB construction scheme and a consistency-based ensemble EBRB inference scheme. Specifically, for the purpose of enhancing diversity in the construction scheme, various kinds of oversampling techniques are applied to construct diverse base EBRBs firstly, followed by the calculation of attribute weights based on information gain. As for the inference scheme, the proposed ensemble EBRB model aims to produce inferential outputs not only integrating the rules activated from all base EBRBs, but also taking into consideration the consistency of the activated rules. In experimental study, twenty-six imbalanced classification datasets are used to demonstrate the effectiveness of the proposed ensemble EBRB decision model. Results demonstrate that the proposed model outperforms conventional EBRB systems and other typical imbalanced classifiers.