A fast belief rule base generation and reduction method for classification problems

The belief rule-based (BRB) system has been one of the most significant rule-based systems for its ability to deal with various kinds of information under uncertainty, and it has shown great potential for classification problems. However, the combinatorial explosion problem hugely limits the application of the BRB system as excessive rules could not only increase the computation cost, but also impact the performance of the BRB system. Therefore, motivated by this problem, this paper proposed a fast and accurate belief rule base generation and reduction method. Firstly, a fast belief rule base generation method is introduced, where similar rules are grouped and combined to ensure all the possible situations are covered without generating excessive rules. Then, a redundancy-based belief rule base reduction method is proposed, where the redundancy degree of the belief rule that represents the degree to which a belief rule is affected by other rules is introduced, and it is calculated to identify redundant rules. Furthermore, the conventional evidential reasoning (ER)-based inference process is retained from conventional BRB systems. Thirty classification benchmarks from the well-known UCI machine learning repository are tested to validate the effectiveness of the proposed method, and the results are compared with other rule-based systems, improved BRB systems, and other machine learning methods. Comparison results show that the proposed method could effectively reduce the size of the BRB without costing its accuracy. Furthermore, sensitivity analysis and robustness analysis are conducted, which further show the effectiveness and robustness of the proposed method.