A minimum centre distance rule activation method for extended belief rule-based classification systems

Abstract

Originating from the belief-rule-based (BRB) system, the extended belief rule-based (EBRB) system combined the advantages of the rule-based method and those of data-driven methods. By transforming the data set into extended belief rules and using evidential reasoning (ER), the EBRB system has expanded the application of BRB systems and demonstrated their capability in addressing classification problems. Nevertheless, the problem of activating nearly the entire rule base in every classification process is embedded in the EBRB scheme. There have been advances in rule activation for the EBRB system; however, the introduction of subjective information into the classification, high computational costs and long response times are common problems facing existing rule activation methods. To solve the problems facing rule activation for EBRB systems, a minimum centre distance rule activation (MCDRA) method for EBRB systems is proposed. In MCDRA, no subjective information is required, and no time-consuming iteration procedure is necessary. Two components of the proposed MCDRA, i.e., the filtering procedure and the selection procedure, are designed to eliminate unrelated samples of input query data and to select and activate the highly related samples to the input query data. A total of 12 benchmark data sets are used to test the performance of EBRB with MCDRA (M-EBRB). The experimental results show that compared with other rule activation methods, the proposed method obtains satisfactory rule activation ratios, accuracies and response times. Additionally, M-EBRB performs well on noisy data and comparatively with both the fuzzy-rule-based classification system (FRBCS) and several machine learning classification algorithms. In addition, MCDRA can be utilized as a generic rule activation method and can be used to optimize other rule-based classification systems.