Evaluation of belief entropies: from the perspective of evidential neural network

Abstract

In Dempster-Shafer’s theory, the belief entropy for total uncertainty measure of mass function has attracted the interest of many researchers in recent years. Although various belief entropies can meet some basic requirements, how to judge the performance of belief entropies is still an open issue. This paper proposes a novel evidential neural network (ENN) classifier to evaluate different belief entropies in practical application. Driven by the least commitment principle (LCP), the maximum entropy is integrated into the traditional divergence-based loss function. The proposed loss function consists of divergence and maximum entropy parts, which considers not only the distribution difference but also the degree of approaching the maximum entropy. Some classification experiments are conducted in 7 real-world datasets to validate the effectiveness of the proposed evaluation method.