A structure optimization method for extended belief-rule-based classification system

Abstract

The widely applied belief-rule-based(BRB) system has demonstrated its advantages in handling both qualitative and quantitative information. As an extension of BRB system, the extended belief-rule-based(EBRB) system bridges the rule-based methods and data-driven methods by efficiently transforming data into extended belief rules(EBRs). Many works have been done to apply EBRB system in addressing classification problems. However, the problems of making use of all attributes indiscriminately and activating almost all EBRs still affect the accuracy and computational efficiency of EBRB system. In this paper, a structure optimization method for EBRB(SO-EBRB) system, including attribute optimization and rule activation, is proposed to address aforementioned problems. In the attribute optimization, a weighted minimum redundancy maximum relevance(MRMR) method is proposed, where the relevance between attributes and label as well as the redundancy among attributes are used to evaluate attributes. Afterwards, the proposed attribute weight calculation method is utilized to assign attribute weights for the EBRB system. In rule activation, an improved minimum centre distance rule activation(MCDRA) method, which considering the weights of attributes in distance calculation, is used to activate customized EBRs for input query data. 15 benchmark classification data sets are utilized to verify the effectiveness of the proposed SO-EBRB method. The results show that, compared with conventional EBRB system, the SO-EBRB system achieves higher classification accuracy, lower rule activation ratio and less response time. Additionally, comparison between the proposed method and some state-of-art machine learning algorithms demonstrates that the SO-EBRB system achieves prominent performance in addressing classification problems.