Evidential reasoning based ensemble classifier for uncertain imbalanced data

Various studies have focused on the classification of uncertain or imbalanced data. However, previous studies rarely consider the classification for uncertain imbalanced data. To address this research gap, this study proposes an evidential reasoning (ER) based ensemble classifier (EREC). In the proposed method, an affinity propagation based over-sampling method is developed to obtain the balanced class distributions of the training datasets for individual classifiers. Using the balanced training datasets, ER-based classifiers are constructed as individual classifiers to handle data uncertainty, in which attribute weights are learned from the similarity between the values of attributes and labels. With trained individual classifiers, final results are generated by combining the results of individual classifiers using the ER algorithm, in which the weights of individual classifiers are determined according to the classification performance on out-of-bag data. The pro-posed EREC is applied to the diagnosis of thyroid nodules using the datasets of five radiol-ogists, obtained from a tertiary hospital located in Hefei, Anhui, China. Using real datasets and UCI datasets, the EREC is compared with 12 representative ensemble classifiers and other oversampling methods based ensemble classifiers to highlight its high performance. (c) 2021 Elsevier Inc. All rights reserved.

Automated summary

This study introduces an evidential reasoning-based ensemble classifier to handle uncertain and imbalanced data, filling a research gap in the field. By developing an oversampling technique and constructing ER-based classifiers, data uncertainty is effectively managed. Experimental comparisons with real and UCI datasets demonstrate the high performance of the proposed method.